JP2765672B2 - Control method and computer system for data transfer and data removal - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION In a data processing system comprising a plurality of operating systems supporting attachment to a shared DASD and a shared electronic storage (SES), a data processing system on disk, in a cache included in the shared electronic storage. , And in the local processor storage. DASD
The management of data movement between and the shared electronic storage mechanism ensures continuous access to data as it moves between storage locations, while ensuring that the most frequently referenced data remains resident in the storage mechanism. In order to optimize the use of shared electronic storage to avoid congestion in storage or I / O subsystems including DASD,
One must address the problem of providing a means to efficiently move modified data to the ASD.

Specifically, the present invention increases system efficiency in ensuring that cached versions of data blocks are moved to secondary storage after being updated in a multi-system data sharing complex. About things.

The present invention relates to the following application. Japanese Patent Application No. 3-299273 entitled "Configurable, Recoverable Parallel Bus"
G. Bartow et al., U.S. Patent Application No. 839657, filed February 20, 1992. "High Performance Channels For Data Processi
U.S. Patent Application No. 839 to NGBarrow et al.
No. 652 (filed on February 20, 1992). "Frame-Group Transmiss
U.S. Patent Application No. 839986 entitled "Ion And Reception For Parallel / Serial Buses" (February 2, 1992
0 day application). "Communication Messages Between Process
US Patent Application No. 860380 to DAElko et al. entitled "Ors And A Coupling Facility." Sysplex Shared Data Coheren
U.S. Patent Application No. 860805 to DAElko et al., entitled "cy Method and Means". "Method and Apparatus For Distribute
d Locking of Shared Data, Employing A Central Coup
U.S. Patent Application No. 86 to DAElko et al.
No. 0808. DAElko entitled "Command Quiesce Function"
Other U.S. Patent Application No. 860330. "Storage Element For A
DAElko entitled "Shared Electronic Storage Cache"
Other U.S. Patent Application No. 860807. "Command Retry System"
U.S. Patent Application No. 860378 to DAElko et al. "Integ
US Patent Application No. 860800 to DAElko et al. entitled "Rity Of Data Shared Between Caches Over A Link."
ent Of Data Objects used To Maintain State Informa
J. entitled "Action For Shared Data At A Local Complex"
U.S. Patent Application No. 860797 to A. Frey et al. "Recovery Of Dat
a Objects Used To Maintain State Information For S
United States Patent Application No. 860647 to JA Frey et al. entitled "hared Data At A Local Complex."
For Managing Connections Between Processors And A
U.S. Patent Application No. 860646 to DAElko et al., Entitled "Coupling Facility." Method And Apparatus For Notification
Of State Transitions For SharedLists of Data Entr
JAFrey et al., U.S. Patent Application No. 860809, entitled "ies."
ethod And Apparatus For Performing Conditional Ope
DAElko entitled "rations on Externally Shared Data"
Other U.S. Patent Application No. 860655. "Apparatus And Method
For List Management In A Coupled Data ProcessingSy
US Patent Application No. 860633 to JAFrey et al., entitled "Stem."
Interdicting I / O And Messaging Operations In A Mul
U.S. Patent Application No. 860489 to DAElko et al., entitled "ti-System Complex." Method And Apparatus For Coupling Dat
No. 860803 to DAElko entitled "A Processing Systems".

[0004]

BACKGROUND OF THE INVENTION Typically, conventional hardware caches in processors include storage arrays having fixed-size blocks of data. Such storage appears to the hardware as a linear block array. Data blocks in the cache are managed by a cache directory. The cache directory invalidates data blocks based on LRU storage references, or casts in or out of cache. A well-known type of hardware cache is a central processing unit (C
PU) and DASD (Direct Access Storage).

[0005] In a database system in which a plurality of independently operating computer systems share data, coherency of the data in different systems may occur.
Global locking is required to maintain rency). AJ van de Goor says "COMPU
TER ARCHITECTURE AND DESIGN ", Addison Wesley (1989
Years), the problem of data coherency is that sharing data between a very large number of processors can lead to multiple paths to the data and the opportunity to modify the data locally, resulting in inconsistent copies of the data. Argued that the possibility exists.

Several solutions to the problem of data coherency have been proposed. They are all based essentially on the existence of a global lock on data retrieved from a central location. Assuming that the data is paged, one computer system of a multi-computer system sharing data stored on a disk acquires a global lock on a page of data and obtains that page. Update. The lock informs other computer systems that the page has been acquired for an update. Before releasing the lock on the page, the computer system holding the lock writes the page to disk and then requests that the copy of the page held in its local cache be invalidated. Generate a message and send it to another computer system. This lock on the page will not be released until an acknowledgment is received from any other computer system that has access to the page. Similar solutions are disclosed in U.S. Pat. Nos. 4,399,504 and 49657.
No. 19 describes this in detail. A commercial product from the applicant of the present invention that incorporates this solution is an IMS / VS system with data sharing capabilities.

[0007] Prior art global locking systems offer significant benefits in maintaining data coherency. However, its inherent overhead penalty is that I / O procedures must be performed when updating pages, and that messages must be exchanged after I / O procedures to notify other systems and release locks. It is mentioned.

Prior art global locking, when used in a single system with non-data sharing, requires that each transaction that updates the data write modified data to storage along with the log record in storage before the transaction is updated. Enforcing a commit policy that is not fully committed creates extra overhead in maintaining data coherency (consistency) between transactions. This requires one I / O procedure per page for each modified transaction, increasing overhead costs.

In contrast, in the case of single system non-data sharing, IBM DB2 follows a policy that does not require an I / O process to write updated pages back to storage to commit the transaction. I have. When using the above protocol in a data sharing situation where multiple computer systems access one or more data storage sites, write back to storage is required and message delays may cause performance problems. May be significantly reduced.

In a multi-computer data sharing system containing multiple storage stages, the second storage stage consists of one or more direct access storage devices (DASD), which are shared by independently operating computer systems. It is contemplated. Typical nomenclature for hierarchically arranged storage systems classifies DASD and other such storage mechanisms as "secondary" storage. It should be noted that secondary storage includes all features that require the data to be moved to the "primary" storage before the CPU can directly reference it. Further, it is contemplated that caching techniques will help provide fast, frequently accessed storage for shared data. For various reasons, data is placed into a shared cache by the database system after it is obtained from DASD. Note that the shared cache is included in the primary storage stage of the multi-computer data sharing system.

In such a structure, one computer system obtains a data block to be written to DASD from a shared cache, while at the same time another computer system obtains the same data block by its shared cache. There is a potential danger in such cases as obtaining from a shared cache and modifying it back to the shared cache. In this situation, retrieving the modified data block from the shared cache for storage on DASD is a "cast out" of that data block.
out) ". Castout reads the page to be cast out from shared memory,
Write to DASD, then "unchanged" in shared memory
(That is, the contents of the corresponding pages are the same in the DASD and the shared memory. Conversely, when the contents of the two pages are different, “changed” is marked).

[0012]

Efficient cache management of a shared cache requires that the shared data blocks be cast out periodically or based on a threshold of changed data blocks in the cache. is there. When a block of data is cast out,
It is marked "unmodified" and is a candidate for removal from the cache. When one computer system writes a new version of the page to shared memory during the time interval between a read operation and a delete operation, a significant danger arises if the castout is performed by another computer system. The danger is that the deletion will erase the new version of the data block. To prevent this from happening, a high level of locking or serialization and queuing is typically used in the shared cache. The problem with high levels of locking is that doubling the cost because removing a page from the cache requires two additional multi-system interactions, locking and unlocking. Also, the writing of the modified version by the second system is delayed, which can have undesirable performance consequences.

Heretofore, the term "class" has been defined as a VM / 37 for controlling printing, punching, and network operation.
It has been used in record processing controls, such as the 0 spool file class. However, it was not known to be used for SES devices for controlling the castout process under the control of CPC.

[0014]

[Means for Solving the Problems]

Sysplex hardware structure: FIG. 1 shows a sysplex (SYStem comPLEX: system complex) system. The system includes a plurality of computer processing complexes (CPCs) CPC-1 through CPC-M. These represent an arbitrary number of CPCs from one to many, all of which are connected to one or more SES (shared electronic storage) devices (FIG. 1 shows one SES device 101). .

Each CPC is of the type shown in FIG.
It may be a multiprocessor such as the currently available IBM ES / 9000 model 900 designed according to the BM ESA / 390 scheme. The IBM ESA / 390 system is described in “Enterprise System Architecture (ESA / 390) Manual (POP) [Enterprise Systems Architecture (ESA)
/ 390) Principles of Operations (POP)], IBM
It is specified in document number SA22-7201-00.
Each CPC has one or more operating systems. If there is a CPC with multiple operating systems, its resources are transferred to IBM PR / SM
Features must be used to logically partition between multiple operating systems. An inter-system channel (ISC) is connected between SES 101 and CPC1-M. The ISC connected to the CPC exchanges signals with microcode or hardware in the CPC.

Each CPC in the sysplex operates on a storage hierarchy. This storage hierarchy is, for example, each CP of CPC.
Dedicated high-speed hardware caches 201-1 through 201-N in U, shared hardware cache 20 accessible to all dedicated processor caches
2. It may include a main storage (MS) 204 shared by all processors in the CPC, a hardware storage (HSA) 205 associated with the MS but without MS addressability. However, DASDs are grouped by DASD controls (controls) so that any CPC in the sysplex can access any DASD in the group. This is referred to herein as "Sysplex DASD" 207.

The CPC / SES physical connection 208 is provided by a corresponding channel connected at one end to the MS controller in the CPC and at the other end to the SES device. This channel bus can be composed of serial optical fibers. The bus may be a single fiber, but it consists of multiple fibers operating in parallel, and "striping" data between them (interleaving)
May be.

In the hardware sense, the SES can be thought of as a large random access memory that can be used in common by all CPCs connected to it.
Attached CPCs can use SES to store shared data records and files temporarily or semi-permanently. Therefore, the SES uses the extended storage (E) in the CPC that is common to all CPCs attached to the SES.
S) It can be considered as a component of the storage hierarchy in the system, with hierarchy levels approximately corresponding to the stages.

In a sysplex using the present invention, 1
One or more SES entities can physically connect to the MS / ES of every CPC in the sysplex. SEs all CPCs in the sysplex
There is no need to connect to S. For example, SES may be connected to only a group of CPCs operating the same programming subsystem. And different groups of CPCs connect to different SESs in the sysplex,
Different programming subsystems can be run.

A fundamental feature underlying the present invention is the use of SES as a fast cache for data that is typically stored on a common DASD in a sysplex. However, the physical connection of CPC / SES / DASD need not be directly in the hierarchical path. Which C in the sysplex
PCs can also access records from SES much faster than from common DASD. That is, S
During the ES, data elements or records can be quickly accessed without the electromechanical delays seen with DASD, such as waiting for head movement between tracks or waiting for track rotation to reach the required DASD record.

A special command for allocating the SES cache is provided. Also, a plurality of caches can be allocated in the same SES, for example, by using different programs to cause each cache to process data shared by the added plurality of CPC subsets.

Each SES cache has a directory 10
2, including a data area 103, a local cache register 104, and a cache control (mechanism) 105. If the data area portion of the cache is not used,
The size may be set to 0. Each valid directory entry in the SES cache stores the name of a data element registered in SES by any attached CPC. The SES may or may not include a copy of the data in the registered data element. The name registered with SES is 1 in the sysplex.
It is also the name of one or more copies of that data element in one or more CPCs. In addition, the directory name also identifies a copy of the data element stored (or about to be stored) on one of the DASDs 1-K in the DASD bank connected to the directory 207.

The present invention (claims 3 to 5 and claim 9)
One of the objects of the present invention is that in a shared data multi-computer system, even if there is a page of data being cast out in the shared memory (SES) prior to deletion, the shared memory (SES) may be The present invention provides a method and structure for ensuring that a later (ie, newer) version of a page written to a.

An important advantage of the present invention is that in shared memory,
No high-level locking or serialization and queuing mechanisms are required to ensure consistency between page versions when removing certain pages from shared memory for entry into secondary storage (DASD) That is.

An important object of the present invention (claims 6 to 8) is to provide a method for operating a multi-system, data sharing complex, wherein data is cached in a shared electronic storage device (SES). It is in. In a multi-system, data sharing complex, a database system running on a first computer system reads a modified page in a shared cache as a first step to write the page to secondary storage. While another database system can attempt to cache a more recently updated version of the same page in the shared cache. The present invention
Detecting such a condition and removing the updated copy of the page from the shared cache after the first computer system stores the previous version of the page in secondary storage without blocking mechanisms such as locking. Can be prevented.

The present invention is directed to a sysplex that uses multiple operating systems (OS). Each CPC in a sysplex logically partitions its resources among different OSs, and multiple independent CPCs within a single CPC, such as through an IBM Processor Resource / System Management (PR / SM) system. An OS can be provided. Thus, a sysplex can have any combination of OSs running on its different CPCs. Some CPCs each have one OS,
Another CPC is a plurality of OPCs running independently of each other.
Has S. One or more subsystems can run under any OS in any CPC, including subsystems such as IBM's DB2, DFP, IMS, VSAM, and the like.

Different copies of the same database subsystem program can run independently and simultaneously in different CPCs. According to the present invention, such different programs can simultaneously access the same or different data elements or records in the database in the MS / ES local cache (LC) of different CPCs.

It is a primary object of the present invention (claims 1 and 2) to ensure continuous access to data during its movement between storage locations, with the most frequently referenced data resident in the storage facility. Optimized use of shared electronic storage to remain in place, and efficiently modified data from shared electronic storage to DASD to avoid congestion on shared storage or I / O subsystems including DASD To manage data movement between DASD and shared electronic storage while addressing the various problems of providing a means to move data to and from the DASD.

The most significant innovation in the technique of operating a shared cache of the present invention (claims 3-9) is that a more recent version of a page is not deleted from the cache while a previous version is written to secondary storage. No additional serialization mechanism, such as high-level locking, is required to ensure this. This cache is accessed by a directory that stores a directory entry for each data page stored in the cache. Essential to the invention is the provision of a castout field in the directory entry for each page in the cache that contains the identity of the computer system currently performing the castout operation. This field is
Works in conjunction with the change field used to indicate whether the page has changed. If this change (bit) field indicates that the page has been modified during an ongoing castout operation, it prevents the page from being deleted, thereby causing the latest version of the page in the shared cache to be Save for out operation. The present invention assumes that the castout field is zero.
The function is performed by permitting the castout only when there is no ongoing castout, and deleting the page only when both the castout ID and the change bit fields are 0. The present invention requires the addition of certain operations to the normal set of read / write operations, a "read" operation for castout.
This read operation for castout involves placing the identity of the requestor in the castout ID field and setting the change field to zero.

SES structure: The SES cache is a structure in SES which is composed of an aggregate of a data area element, a directory, and a local cache control (part). S
The ES cache structure is created at the request of programs that access data shared between CPCs, and these programs require coherency and integrity for copies of locally cached data items.

SES directory: SES cache
The directory is a basic component of the SES device for obtaining SES coherency control. Forming the SES data area with the SES cache improves sysplex performance, but is optional. Without the SES data area, data records in the sysplex would only be accessible from the sysplex DASD. In database software operation, S
Performance provided by rapid access to shared records in the ES data area will be lost. The SES local cache register (associated with the SES directory entry) still identifies which CPC LCB in the sysplex has a copy of the shared data element.

SES local cache register:
Entries in this local cache register identify an additional local cache that contains a copy of the data element identified by the associated directory entry. Each entry in the local cache register contains a cache coherency vector associated with the local cache and the coherency vector used to represent the validity of the locally cached copy of the data element. Provide enough information to locate the local cache entry.

Overview of the process flow: SES cache storage is typically smaller than DASD storage. Therefore, the changed data is transferred from the SES cache to the backup DAS
D must be transferred periodically. This process is called castout and is controlled by the program,
The following operations are required. A SES read re-operation is issued to set castout serialization and copy the data block to main storage. An I / O operation is performed to DAS the data block
Copy to D. An SES lock release operation is issued to release the castout serialization.

The data items associated with the castout are:
Maintained in the castout class by the SES cache. The castout class is used to improve the efficiency of the castout process by allowing a program to batch data transfers to DASD in a single I / O operation. DASD and S
The process of performing the movement of data between ES caches is the subject of the present invention.

Objects in the SES directory that enable efficient castout processing include the following (see also FIGS. 3 and 8):

A castout class (CCL). A two-byte value that identifies the castout class assigned to the name.

Castout lock (CO). A 2-byte value indicating the status of data castout. When the castout lock is zero, the data has not been castout. When the castout lock is non-zero, the value of the first byte of the castout lock is
Move the data block from SES cache to DASD
Identify the local cache that is casting out to The value of the second byte identifies the castout process on the local system. If the castout lock is not zero, the data bit must be one.

A change bit (C). A one-bit value that, together with the castout lock, indicates the changed state of the data. When the change bit is 1, the data is cached in a changed state. Change bit is 0
If the data is not locked for castout, then the data is not cached or cached but not modified. If the change bit is 0 and the data is locked for castout, the data is cached in a modified state. When data is in the modified state, the latest version of the data resides in the cache.
When the change bit is one, the data bit must also be one.

The castout process can be started in one of several ways:

Event Driven Castout: After each write operation, the count of changed data elements is compared. This triggers a castout process for the specified castout class when a write command occurs and exceeds a defined threshold. This comparison is part of the transaction processing.

Timer-driven castout: a set of counters is required at timed intervals. This comparison is part of the background program and not part of any transaction processing. A balance must be struck between responsiveness to the castout process and transaction overhead in obtaining the count.

Commands that support the castout process in SES include:

Read castout class command: The read castout class command returns a list of name blocks to main storage. The directory entry is processed when the name field matches the input name in the masked state and the directory entry is enqueued to the specified castout class queue.

Read castout class information command: The read castout class information command is:
Returns statistics for a range of castout classes.

Read Directory Command: The Read Directory command scans a directory and returns a list of directory entry information blocks.

Read for Castout command: The Read for Castout command returns the contents of the named data area to the local cache and locks its name for castout.

Castout / Lock release command:
The unlock castout lock command resets the castout lock to zero and stores its user data field in the directory entry for the namelist.

Write and Register Command: The write and register command stores the contents of the data block in the data area and registers a local cache entry when the change control matches the change state of the data.

Write-on-registration command: A write-on-registration command is a command whose name is assigned in the directory,
If the local cache entry is registered and the change control matches the change state of the data, store the contents of the data block in the data area.

During mainline operation, the database manager is responsible for managing the storage of the SES cache. The database manager is also responsible for managing the movement of changed data elements from the SES cache to DASD. The timeout interval established at initialization causes the operating system to
The processing routine is periodically given control for managing the S-cache area and managing the transfer of SES cache data to DASD. The above mechanism is the basis for performing management of data movement between DASD and SES cache in system software. Techniques are used to implement a non-blocking serialization protocol between writing and casting out data from SES using this primitive service. Furthermore, the above primitive services are the basis for determining a set of modified data elements to be moved from SES to DASD in such a way that the overall performance of the combined system environment is optimized.

An example of one protocol for managing data movement from SES cache to DASD is as follows. A read castout class information command is issued for each castout class and a candidate castout class is selected. For each selected castout class, a read castout class command is issued. A list of the most recently updated directory entries is returned. A read command for castout is issued for each selected directory entry. An I / O request is created to cause the data read from SES to be written to DASD. When the I / O operation is completed, a castout lock release command is issued.

[0052]

【Example】

SUMMARY: The present invention maintains the structure of a multi-system data sharing complex that includes a shared cache (SES) in the form of non-volatile electronic memory, and coherency among several potentially different data versions. While operating the cache to share data resources.

If the data coherency strategy in a multi-computer data sharing system requires the writing of updated data pages, an architecture with several important features can be used. Such an architecture is shown in FIG. 1, which comprises a plurality of independently operating computer systems CPC-1 through CPC.
−M, these computer systems share data stored in direct access storage devices (DASD) 109A, 109B-109N. DAS
D109A, 109B-109N may include, for example, a multi-disk disk drive. Characteristically, this is called "secondary storage area". The architecture includes M computer systems CPC-1 through CPC-M, each of which includes a database management system (DBMS), which creates, organizes, and modifies a database containing data on DASD. To control access to data in the database. A high speed non-volatile electronic memory (SES101) is also provided in the system and serves as a cache shared by the computer system. The memory 101 is connected to the computer systems CPC-1 to CPC- by high-speed links 106-1 to 106-M.
M. Hereinafter, this memory 101 will be referred to as “memory”
Or, it is called “NV storage device” or “SES”.

When comparing the access to the memory 101 with the access to the secondary storage area, it can be said that the memory 101 is a relatively high-speed semiconductor memory. Further, the memory 101
Is connected, for example, by a fiber optic communication channel that allows very high-speed data transfer. Relatively speaking, input / output operations are performed in a relatively high-speed memory 101, while, as is well known, a relatively low-speed DA
Input and output by SD can take tens of milliseconds.

The memory 101 includes management logic 110, preferably in the form of a processor, which manages the storage operations of any memory. The management logic 110 comprises a high performance processor that includes local program storage and dedicated memory, the processor having a management logic based on messages to and from the computer systems CPC-1 through CPC-M. Be able to engage in access transactions.

With respect to the computer systems CPC-1 through CPC-M, these entities each have a multiprocessor architecture with a dedicated cache and can each support an IMS / VS or DB2 type database management system. For example, a system such as IBM / 3090 can be provided.

In essence, the invention is implemented in the data sharing complex shown in FIG. 1 and described above. In the memory 101, there is a semiconductor memory called a shared cache 103. Shared cache 103 may include conventional multi-port high-speed random access memory, which is preferably non-volatile. Shared cache 10
3 is used to store a data block. For example, data pages can be stored using shared cache 103. One page is shown as 111.

The management logic 110 in the SES comprises one or more processors or microprocessors and programs or microprograms executable by those processors. These processors receive the messages sent by the CPC commands and execute those commands in SES. The local cache control 105 and the directory 102
Accessed by management logic 110. Using the well-known hashing lookup algorithm, SE
You can access the S-cache directory. The local cache control (part) is a data structure including a plurality of entries.
1 identifies a computer system operably connected to the computer system. For example, the computer system CPC
-1,... CPC-M are connected, they will be listed in the local cache control 105.

The shared cache 103 operates as a “store-in” type cache instead of a “store-through” type cache. A “store-in” cache is a cache that can be written without having to write updated pages to secondary storage at the same time (“store through”).

The SES cache is a component of a three-level storage hierarchy in a network of additional processors. The lowest level of this hierarchy is DASD, the middle level is the SES cache, and the highest level is the local cache in processor storage. DASD and SES cache are shared by the processor and accessed by I / O operations and message operations, respectively. The local cache can be defined in each processor and is accessed using CPU instructions.

Data that moves through the storage hierarchy is given a name. The data size is variable and ranges from 1 to 16 times the size of the data area element. The data area element size is fixed for each SES cache, and the minimum size is a power of 2 and the minimum size is 256 bytes. The name is a 16-byte value assigned by programming control. The data resides permanently in DASD storage.

A copy or a new version of the data can reside in any combination of the SES cache area or local cache area. Specifically, the data item resides in the SES cache area, but may not reside in any of the local caches, may reside in the SES cache storage and part of the local cache, -Resides in a part of the cache, but may not reside in the SES cache area.

As an example, assume that the data element referenced in the storage hierarchy is a page. Then, the pages cached in shared cache 103 are identified by software-assigned names.
Therefore, a request for a read or a write in the shared cache 103 is made on the page (P) for which
Must be specified. The directory 102 is
It is indexed by the name of the page that is the subject of a read or write command as before. FIG. 3 shows a representative entry of the directory 102. For the present invention, the important fields of the directory are shown in FIG. It includes a name field 301, a data bit (D) field 303, a data area pointer field 306A-306N, a change bit (C) field 304, a valid bit (V) field 302, a data area size field 307, and a castout. Includes a lock identification (CO) field 308, a castout class field 309, and a local cache register index 305 that provides access to the local cache register entry shown in FIG. The local cache register entry contains the identification (402A-402C) of the local cache added to the SES cache and a valid bit (401A-401C). local·
Two examples are shown that result in different types of entries in the cache register 305.

Continuing with the above example, page name field 301 is the field that management logic 110 uses to index into directory 102. Assuming that the management logic 110 receives a read or write command, the command is accompanied by the value of a parameter P that identifies the page. Management logic 110 presents the value of P to the hashing process, which uses the value used by management logic 110 to quickly access the directory using the page name if it already exists. Generate After the page name field is located, the page address field 306 is used to indicate the address of the identified page in the shared cache.

The management logic 110 creates, manages, and deletes directory entries as needed. These activities are performed using known mechanisms, but directory 1
The exact structure of the 02 entry is unique to the present invention. The management logic 110 has a conventional structure for obtaining and putting data from the shared cache 103, but the read and write operations that it follows are unique to the present invention. The management logic 110 also has normal cache management capabilities to generate "cache miss" and "cache hit" responses. These responses are generated in response to commands issued from a computer system connected to the shared cache 111. A “cache miss” indicates that the identified data element (eg, page) is not present in shared cache 111, and a “cache hit” indicates that the identified data element (eg, page) is not in shared cache 111. Indicates that there is.

The commands are generated by a computer system (CPC) in the multi-system complex of FIG. These commands elicit responses from management logic 110. The inventors contemplate that commands and responses are exchanged between the computer system and the management logic by a message protocol. Further, the inventor contemplates that access to shared cache 101 is synchronous in that the computer system issuing the command can maintain a delay until a response is received from management logic 110. Shared cache 1
The speed of the semiconductor memory forming 01 reduces the delay inherent in the synchronous message passing structure.

The inventor has also determined that the computer system of the multi-system data sharing complex of FIG. 1 can obtain access to the DASD using conventional means, for example, the shared disk function of the IBM IMS specification system. Is intended. As is known, such accesses are asynchronous in that the computer system is not delayed while a read or write command is dispatched to DASD.

As shown in FIG. 1, the NV storage device (SE
S) 101 is not directly added to any secondary storage device. The computer systems CPC-1 through CPC-1
Each DBMS of the PC-M is stored in the NV storage device 101.
Knows its existence and is responsible for the contents cached therein. In the system of FIG. 1, updated data elements (eg, pages) are explicitly written to NV storage by the DBMS.

The responsibility for maintaining stealable pages in NV storage device 101 is left to one of the DBMSs running on the computer system,
Or shared by those DBMS. Management of page space in NV storage to ensure availability for new pages is performed by periodically writing updated pages from NV storage to secondary storage via a castout operation. This castout operation is preferably asynchronous with the transaction commit operation. Data element (for example, page)
Is preferably updated a plurality of times before being written to the secondary storage device.

As shown in FIG. 1, each computer system includes an identified buffer. This buffer is
Used to stage data exchanged between the computer system and the NV storage device 101. For example, the computer system CPC-1 includes the buffer 107A and the computer system CPC-1.
−M is provided with a buffer 107C. All computer systems in the data sharing complex use such dedicated buffers. In addition, when one of the computer systems provides a read or write command to the management logic 110, that computer system sends the address of the location where the requested data should be entered or obtained into its dedicated buffer.

The present invention allows data to be updated from the shared cache while allowing the data to be cast out to be updated.
Regarding casting out to secondary storage. Once updated, a data element (eg, page) is considered "dirty" until it becomes consistent with its updated version in secondary storage. When the change bit in the directory entry for a data element (eg, a page) is 0, the page is called a "clean page." Note that the version of the page in DASD in this case is the same as the version of the corresponding page in the shared cache 101. Conversely, if the change bit (C) for the page is 1, the page is "dirty" and the version of the cached page is different from the version of the page in secondary storage (from Also new).

The present invention is based on registering a set of unique commands and the changes resulting from their execution in the affected directory entries.

Local Cache: A local cache (LC) is defined for the SES support mechanism on a CPC that includes a local cache as described in US Patent Application No. 860797. A CPC instruction initializes control in the SES support facility and the local cache
Assign a token. The size of the local cache varies from system to system. Each local cache is 1
One or more local cache buffers (LC
B) 107.

SES cache: The SES cache is
This is a structure in SES that is composed of a collection of data area elements and directories. This is specified by a structure identifier. SES
The cache is created by a cache structure allocation command. This command is issued by the initialization means in the processor and determines the attributes of the SES cache, namely the size and number of data area elements, the number of directory entries, the number of storage classes, and the number of castout classes.

The local cache is added to the SES cache by a local cache addition command. This command initializes the controls in the SES facility and associates the local cache with a set of paths that the SES cache uses to issue generated commands to the SES support facility. The local cache is added to the SES cache, allowing it to participate in the storage hierarchy. Coherency of copies of data in the local cache and in the SES cache is maintained by controls in the SES cache and is enforced by cross-invalidate (XI) commands issued as generated commands to various SES support mechanisms. cache·
The coherency process is described in US patent application Ser. No. 860,805.
No.

SES cache directory: SES
A cache directory is a collection of directory entries arranged as an associative array. Directory entries are partitioned into storage classes. The group of changed directory entries is divided into castout classes. When a named data object is placed in the upper two levels of the storage hierarchy, its state and location are registered by the SES cache directory. The status information indicates whether the data has changed, has not changed, or has been locked for castout. In the position information, the data is S
Whether it resides in the ES cache area and which local cache contains the copy. A record (located in the SES cache data area by a pointer at position 406A-N) is referenced therein by any of several different items such as a record, data element, data item, block, page, and the like. .
Certain SES read and write commands are stored locally in the SES cache directory.
Register a cache copy. The SES write command and the SES invalidation command deregister the local copy.

Local cache entry valid state: When data is in the local cache, the status of the data is valid or invalid. The valid state of the local cache entry is maintained by control in the SES support facility. Data is validated by CPU instructions and invalidated by an invalidation process associated with SES write and SES invalidation operations. The valid state of the data is tested by a CPU instruction. A valid named data object must be registered in the SES cache directory. Local cache coherency is maintained by the invalidation process. The process of maintaining data coherency using these controls is described in U.S. Patent Application No. 860805.

Castout process: SES cache storage is usually smaller than DASD storage. Therefore, the changed data must be periodically transferred from the SES cache to the backup DASD. This process is called castout and is controlled by the program and requires the following actions: An SES-read operation is issued to set castout serialization and copy the data block to main storage. An I / O operation is performed to DAS the data block
Copy to D. -An SES-unlock operation is issued,
Release castout serialization.

The data objects associated with this process are maintained in a castout class by the SES cache. The castout class is used to improve the efficiency of the castout process by allowing a program to batch data transfers to DASD in a single I / O operation. DA
The process of performing the movement of data between SD and SES cache is the subject of the present invention.

Reclaim Process: The least recently used (LRU) unmodified data and directory resources are reclaimed by the SES cache as needed to satisfy new requests. Data objects are mapped by the program into one of several storage classes. Each class has a reuse vector for controlling the reuse process. This allows S between storage classes to compensate for changes in workload characteristics.
It is possible to dynamically adjust the ES storage allocation. The reuse vector is initialized by the program. The process of utilizing such SES cache area management objects is the subject of U.S. patent application Ser. No. 806,807.

Instrumentation: Instrumentation information is provided to the program to assist in the data allocation and storage process between storage classes.

Cache Structure: A set of cache structure objects is established for each SES cache structure.
The cache structure object consists of: -Cache control (part)-Directory-Castout class control (part)

Cache control (part): Maximum castout class (MCC): A 2-byte unsigned binary integer that specifies the number of castout classes. Valid castout class values range from 1 to the maximum castout class value.

LCID Vector (LCIDV): A bit string whose initial value is 0. The bit positions start from 0 and increase sequentially. The bit at position (i) in the string is set to 1 when the local cache with LCID value (i) is added. This bit is 1
, A local cache identifier is assigned. The bit at position (i) indicates when the local cache has been purged and an LCID deallocation was requested.
Alternatively, it is reset to 0 when the cache structure is deallocated. When this bit is 0, no local cache identifier is assigned.

Local cache control (part): The local cache control is performed when the local cache
Initialized when added to the cache and deleted when the local cache identifier is deallocated. Local cache control is valid when a local cache identifier is assigned.

Directory: The directory is the SES
This is a storage location of state information and location information for the cache and additional local cache. Each named data block appearing in the SES has an associated directory entry. Directory entries are arranged as an associative array and are accessed using the value of the name field.

The fields of the directory entry (FIG. 3) are summarized as follows. -Castout class-castout lock-change bit-data bit-data area size-local cache register-name-valid bit-user data field

The castout class (CCL) (30
9): A two-byte value that identifies the castout class assigned to the name.

Cast Out Lock (CO) (30)
8): A 2-byte value indicating the castout state of data. When the castout lock is zero, the data has not been castout. When the castout lock is non-zero, the value of the first byte of the castout lock identifies the local cache that is casting out the data block from the SES cache to DASD. The value of the second byte identifies the castout process on the local system. If the castout lock is not zero, the data bit is one.
Must.

Change Bit (C) (304): A one-bit value indicating the changed state of the data in combination with the castout lock. When the change bit is 1, the data is cached in a changed state. If the change bit is 0 and the data is not locked for castout, the data has not been cached or has been cached but not changed. If the change bit is 0 and the data is locked for castout, the data is cached in a modified state. When data is in the modified state, the latest version of the data resides in the cache. When the change bit is 1, the data bit is also 1
Must.

Data bit (D) (303): A one-bit value indicating whether data is in the SES cache. When the data bit is 1, the data has been cached. When the data bit is 0, the data is not cached.

Data area size (DAS) (307): 5-bit unsigned binary integer that specifies the size of the data area as an integer multiple of the data area element size. directory·
When an entry is assigned, the initial value is 0 and is 0 until the data bit is set.

The local cache register (LC
R) (FIG. 4): What is a local cash register?
5 is a table that contains information about the storage locations of locally cached copies of data blocks. Each row of the table corresponds to one directory entry.
In the column, the local cache identifier (402A, 40
2B, 402C) and valid bits (401A, 401B, 401) relating to the local cache entry number.
C) is included.

Example B of FIG. 4 illustrates one embodiment of a local cache register (LCR) structure. This entry contains j LCID fields, where j is the maximum number of local caches that can be added to the SES cache in the sysplex. Each entry in the LCR contains a valid bit (V). This V bit is set to 1 if the field represents a certain LCID. If V is 0, the field is whatever LCI
It does not represent the D value.

The LCID value can be found in the local cache by S
Allocated by operating system software when a request is made to connect to the ES cache. The entries in the LCR are LCI
They are arranged in the order of the D values. The valid field for LCID1 is at the first LCR entry position, the valid field for LCID2 is at the second LCR entry position, and so on until the valid LEN field for LCIDj.

Name (N) (301): The name contains a 16-byte value specified by the program when the named data object is registered in the cache.

Valid bit (V) (302): 1-bit field indicating the valid state of the directory entry. There are two possible values: 0 for invalid and 1 for valid. Directory entries are initialized to the invalid state. When in the invalid state, the directory entry is available for a name assignment. The valid state is
Indicates that the name has been assigned, and the remaining fields in the directory entry are valid.

User data field (UDF)
(310): The user data field contains an 8-byte value associated with the data when it was first changed in the SES cache, and is maintained until the data table entry is reused. . The user data field is valid when the data is cached in a modified state.

Castout class control: The castout class control consists of: -Castout class count (CCC)-castout class queue (CCQ)

Castout class count (CC
C): 4-byte unsigned binary integer associated with a castout class. The value indicates the number of data area elements assigned to the castout class directory entry. This count is returned on a write-when-registered or write-and-register command when the change signal is processed for that castout class, and when the castout class is within the specified range, Returned on read class information command.

The castout class queue (CC
Q): A subset of directories that specify directory entries in the castout class, arranged in order.
The queue is ordered by update, with the most recently updated directory entries at the bottom of the queue and the most recently updated entries at the top of the queue. There is a castout class queue for each castout class in the cache.

[0102] The castout class count and total changed count are periodically compared to a threshold to initiate the castout process. This is done in one of the following ways.

Event-Driven Castout: After each write operation, the count of changed data elements is compared. This triggers a castout process for the specified castout class when a write command occurs and exceeds a defined threshold. This comparison is part of the transaction processing.

Timer-driven castout: a set of counters is required at timed intervals. This comparison is part of the background program and not part of any transaction processing. A balance must be struck between responsiveness to the castout process and transaction overhead in obtaining the count.

Cache structure operand: SES and DA
The cache structure operands involved in moving data between SDs are summarized below.

Change control (CHGC): A one-bit value that selects which process is used to write the data. The two possible values are: 0 Write data unchanged. 1 Write data changed.

The local cache identifier (LCI
D): One byte unsigned binary integer identifying the local cache.

Name (N): A 16-byte value that identifies a data block in the storage hierarchy.

Name Mask (NM): A two-byte value that determines the bytes used for name comparison.

Name Replacement Control (NRC): A one-bit value that controls the name replacement process. 0 Suppress replacement 1 Replace name

Replacement Name (RN): A 16 byte value that identifies the data block to be replaced in the local cache.

Resume token (RT): controls processing of an operation for a directory that spans a plurality of commands.
8-byte value.

Cache Structure Processes: Castout Class Removal Directory Designated Directory
This is done for the entry. The processing is as follows. -The directory entry is removed from the castout class queue specified in the directory entry. -The castout class count of the castout class specified in the directory entry is decremented. -The total changed count of the storage class containing the directory entry is decremented.

Change signal processing: Change signals are processed for a specified data area and castout class. The processing is as follows.

The change bit is set in the directory entry to indicate that the data is cached in a modified state.

If the change signal is the first change signal processed for the directory entry, the user data field is stored in the directory entry. The first change signal occurs when the data is not initially cached, or is cached unmodified. When the data is cached with the first change, the change signal is the second and subsequent change signals, and the user data field is not stored.

The castout class is stored in the directory entry, and this directory entry is placed at the bottom of the castout queue. This may be the initial placement in a castout class, may represent a change in class specification, or may be an update of the data in that castout class.

The castout count is updated. If this is the first placement in a castout class, the castout class count is incremented by one. If this is a class change,
The castout class count of the source class is decremented and the castout class count of the target castout class is decremented.
The class count is incremented by one. If this is an update in a castout class, the castout class count is not updated.

The setting of the change bit in the directory entry must be done before the completion of the command.

Name Assignment: A name assignment obtains an invalid directory entry, marks it invalid, initializes its name, and appends the directory entry to the directory to provide a name entry. It is processed.

If no invalid directory entry is available, a valid directory entry is
Is reused, its current contents are cleared, and it is added to the directory. The oldest (oldest used) directory entry that has not changed is reused.

Name Comparison: Name comparison is a byte-level comparison and is controlled by a mask field.
The 16 bits from left to right in the mask field correspond one-to-one with 16 bytes from left to right in the name field. The comparison is advanced byte by byte from left to right. When the corresponding bit in the mask is 1, the byte groups are compared.
Otherwise, no comparison is made and operation continues with the next byte.

Write Modified Data: When a data element is stored in a modified state, the castout class count of the specified castout class and the total number of modified data elements in the SES Is returned.

Cache Structure Commands: The following commands describe actions taken on SES objects in SES, as shown in the process flow diagram. -Read castout class information (Fig. 5)-Read castout class information (Fig. 6)-Read directory (Fig. 7)-Read for castout (Fig. 9)-Unlock castout lock (Fig. 10)-Read and Registration (Fig. 11)-Writing and registration (Figs. 12 and 13)-Writing when registered (Fig. 14)

Read Castout Class: The processes invoked by the read castout class command are summarized in FIG. The read castout class command returns a list of name blocks (NB) to main storage. When the name field matches the masked input name, the directory entry is processed and enqueued to the specified castout class queue.

When all unmasked bytes are equal (501), the name comparison is successful. If the mask is 0, all names are returned.

The castout class is scanned. A name block is added to the returned list, and the processed count is incremented by one for each name in the castout class that matches the masked input name (50
2).

Scanning of the castout class is controlled by the resume token request operand. Processing starts when the token value is 0, and when the token value is non-zero,
Processing resumes from the location specified by the token. Processing is completed when the data block is full, when processing of the entire castout class has finished, or when a model-dependent timeout has been exceeded.
When the data block is full (503), a resume token is generated (504) and returned with a list of directory elements (505) and a processed count. When the end of the castout class is reached (506),
The list of directory elements and the processed count are returned to the program (507). When a model-dependent timeout occurs before reaching the end of the table (508)
Is returned with a resume token generated (509) and a list of directory elements (510) and processed count.

The castout class scan begins with the directory entry specified by the resume token and proceeds in the order specified by the castout class queue array. The ordering of the scans is necessary for each execution of the read castout class command, but not throughout the individual execution of that command. Movement on the queue between successive read castout class commands can change the position of the scan start point on the next run. As a result, some elements of the castout class queue may be lost or returned many times. When the value of the restart token is 0, the scan starts at the top of the castout class queue.

Returning the entire castout class may require issuing multiple read castout class commands. If the program requires a complete, consistent queue, the program must prevent write activity to the castout class until the scan for that castout class is complete.

The processes invoked by the read castout class information command are summarized in FIG. The read castout class information command returns statistics on a range of castout classes.
A counter for the largest subset in the specified range of the castout class and a counter for the range of the largest subset are placed in the data block. These counters are used to store data in castout class order.
Put on the block. The first counter is the one for the castout class designated as start of range (SR). The last stored counter depends on the size of the requested range and the value of the end of range (ER).
If the value at the end of the range is less than the value of the largest castout class and the entire range fits in the data block, the largest subset is equal to the entire range (601).
In this case, the start value of the requested range is placed in the SR data operand and the end value of the range is placed in the ER data operand (602).

If the data block is full and the castout class value of the last stored counter is less than the maximum castout class value (603),
The start of range value is placed in the SR data operand, the final castout class value is placed in the ER data operand (604), and a response code is returned indicating the presence of additional entries (605).

If the value at the end of the specified range is greater than the maximum castout class value and the data block is not full (606), then the start value of the range is SR data.
The largest castout class value is placed in the ER data operand (607) and an out-of-range response code is returned (608).

If the beginning of the requested range is greater than the maximum castout class value (609), a value of 0 is stored in the SR and ER operands (61).
0), an error response is returned (611).

Read Directory: The process invoked by the Read Directory Under Mask command is summarized in FIG. This read directory command scans the directory and returns a list of directory entry information blocks or a list of name blocks. The directory entry is processed (701) when the name field matches the input name in the mask state and the changed state matches the requested state.
If all the unmasked bytes are equal, the name comparison is successful (702). A 0 mask causes all names to be processed.

When the change state selection control is 0, all directory entries that match the name of the mask state are processed. When the change state selection control is 1, only items that have been changed or locked for castout and that match the name of the mask state are processed.

The directory is scanned. Directory list in the returned directory block (NB)
An entry information block is added (703 and 70).
4) The processed count is incremented by one when the request is for changed data, the name in the directory entry matches the input name in the masked state, and the changed state of the data matches the changed state selection criteria. Is done.

The scanning of the directory is controlled by the resume token request operand. When the token value is 0, the process is started. When the token value is non-zero, the process is restarted from the location specified by the token. Processing is complete when a data block is full, when the entire directory has been processed, or when a model-dependent timeout has been exceeded. When the data block is full (705), a resume token is generated (706) and returned (707) with the list of directory blocks returned and the processed count. When the end of the directory is reached (708), the list of directory blocks to return (709) and the processed count are returned to the program. If a model-dependent timeout occurs before reaching the end of the directory (710), a resume token is generated and returned (711) with the list of directory blocks to return (712) and the processed count.

The directory entry information block (D
FIG. 8 shows the format of EIB).

To scan a directory, the directory read command must be executed many times. The directory may change during the scanning process. The set of directory entries that must be processed before the scan is complete consists of valid directory entries that exist and remain valid at the start of the directory scan process. Directory entries that are added or deleted after the scan starts do not need to be processed. Directory entries need only be processed once. Directory entries updated after processing do not need to be processed again.

Read for Castout: FIG. 9 illustrates the process invoked by the Read for Castout command.
Are summarized below. The read command for castout is
Returns the contents of the named data area to the local cache and locks the name for castout.

Data has been cached and modified but not locked for castout (9
01) when the value of the LCID operand is stored in the first byte of the castout lock, the value of the CPID operand is stored in the second byte of the castout lock,
By resetting the change bit to 0 (902),
The data is marked locked for castout. A data response code indicating completion is returned (90
3). The data is returned to the main location specified by the data address in the message command block and, if any ancillary data is present, it is returned in the message response block. The storage class is incremented.

If the data is locked for castout (904), the value of the castout-in-progress lock is returned (905).

If the data is not cached or cached but not modified (906), the command completes with an exception response code (907).

If a name has not been assigned to the directory (908), an error response code is returned.

Unlock Castout Lock: The process invoked by the Unlock Castout Lock List Unlock command is requested and shown in FIG. The unlock castout lock command resets the castout lock to 0 and stores the user data field in the directory entry for the namelist.

Each list item is processed sequentially from the item specified by the SL request operand. Processing continues until the end of the list is reached, an error occurs in processing the list item, or a model-dependent timeout is exceeded. The processing of each list item is as follows.

The directory entry containing the name is found. If the data is locked for castout and the value of the castout lock matches the value of the local cache identifier, the castout lock is set to 0 (1001) and the user
A data field is stored in the directory entry (1002). If both the change bit and the change bit excess indication (CBO) are 0, the directory
The entry is removed from the castout class queue (1003). If the change bit is 0 and the change bit excess indication is 1, the change bit is set to 1 (1
004). The processing for the list item is completed.
The next list item is processed. If the end of the list has been reached (1005), the command is completed and a completion response code is returned (1006).

If a name is assigned to the directory and the castout lock does not match the value of the local cache identifier (1007), the list item is not processed. The value of the list item is placed in the current list item response operand (1008) and the value of the castout lock is placed in the castout lock value response operand (1009). No further processing is performed and a response is returned.

The model-dependent timeout was exceeded (10
10) At that time, the value of the last list item that was successfully processed is placed in the current list item response operand (101).
1), a response code indicating a timeout is returned (10
12).

If no name is assigned to the directory (1013), the list item is not processed. The value of the list item is placed in the current list item response operand (1014). No further processing is performed and a response is returned (1015).

When a non-zero response code is returned, the list has not been completely processed. The CLI response operand identifies the point at which processing was completed. All list items before the specified list item have been processed, and all list items after the specified list item are unprocessed.

When an exception response code is returned, an error occurred during the processing of the list entry. The program is
The list processing can be continued by setting the value of the list start operand (SL) to one greater than the value of the current list item (CLI) and reissuing this command.

Read and Register: FIG. 11 is a process flow diagram of the read and register command. The read and register commands return the contents of the named data area to the local cache and register a local cache entry. When the data is not cached, only the registration operation is performed. The read and register commands also assign a name to the directory when the name is not currently assigned.

When the name replacement control (NRC) request operand is 1 (1101), the local cache entry specified by the request operands of the replacement name and the local cache identifier is deregistered (110).
2). When the name replacement control is 0, the registration is not canceled.

When data is cached (11
03), a local cache entry is registered (1104), a reference signal is started for that storage class (1105), and the data is returned with a change bit (1106).

When a name is assigned to a directory but the data is not cached (110
7) The local cache entry is registered (1
108), a reference signal is started for that storage class (1109).

When a name is not assigned to a directory and directory assignment is not suppressed (1
112), a directory entry allocation operation is performed (1113). If the cache is not full, a directory entry is allocated, a local cache entry is registered (1114), and a reference signal is initiated for that storage class (1115).

If the name is not listed in the directory and the assignment is controlled (1111), the command is completed and an exception response code is returned.

Write and Register: FIGS. 12 and 13 are process flow diagrams of the write and register commands. The write and register commands store the contents of the data block in the data area and register the local cache entry when the change control matches the change state of the data.

The write and register commands also assign a directory entry when no current name is assigned to the directory.

When the name replacement control (NRC) request operand is 1 (1201 and 1301), the local cache entry specified by the request operand of the replacement name and the local cache identifier is deregistered (1202 and 1302). ). When the name replacement control is 0, the registration is not canceled.

When data is not cached, a data table entry is assigned (1
203, 1204, 1303, and 1304). Directories may or may not be assigned names. If not, a name is also assigned (1205 and 1305). If the allocation is successful, if the change control is 0, the data remains unchanged and SE
Written to the S-cache (1206 and 1210),
When the change control is 1, the data is written in the changed state (1
306 and 1310), local cache entries are registered (1207, 1211, 1307, 13)
11).

When the data is already cached without any change (1214), if the change control is 0, the data is written without change (1215).
If the change control is 1, the data is written in a changed state (1315).

When the data is already cached in the changed state and the change control is 0 (1216), no data is written and no local cache entry is registered. Change control is inconsistent with the change state of the data. The command completes and returns an exception response code.

Data is cached in a changed state, and when the change control is 1 (1315), the data is written in the changed state and a success response code is returned.

Write at Registration: FIG. 14 is a process flow diagram of the write at registration command. The write-on-registration command stores the contents of a data block in the data area if its name is assigned in the directory, a local cache entry is registered, and change control is consistent with the change state of the data. I do.

When a local cache entry is registered and data is not cached (1401 and 1402), a data table entry is allocated (1403 and 1404). If the assignment is successful and the change control is 0 (1405), the data is written unchanged. If the assignment is successful and the change control is 1 (1406), the data is written in a changed state.

When the local cache entry is registered and the data is already cached without any change (1409), when the change control is 0, the data is written without change (1410). When the change control is 1, the data is written in a changed state (1411).

If no local cache entry has been registered (1412 and 1413), the command is completed and an exception response code is returned.

When the local cache entry is registered, the data is already cached in a changed state, and when the change control is 0 (1414), no data is written. Change control is inconsistent with the change state of the data.

A local cache entry is registered, cached in a changed state of data, and when the change control is 1 (1415), the data is written in a changed state (1411).

Management of Data Movement Between DASD and SES Cache: FIGS. 15 to 19 show an outline of a system structure including an SES cache, a local data manager using the SES cache, and operating system support. . FIG.
3 summarizes the processing performed by the base manager's buffer manager. When a database that is shared between systems in a sysplex is first accessed, a local buffer pool can be built (1501) and operating system services that provide support for the SES cache structure are provided. A call can be made to allow the buffer manager access to the SES cache structure (1502) and a timeout interval can be set using operating system services (1506).

An operating system service that supports the SES cache structure is called to allow the buffer manager access to the SES cache structure (1502). These services first determine if the requested SES cache structure has already been allocated (1503). If the SES structure has not been allocated, the allocate cache structure command is called to create the SES cache structure (1504). The operating system service selects a local cache identifier (LCID) and defines a DEFINE VECTOR (vector definition).
Issuing a CPU instruction causes a cache coherency bit vector to be created. local·
Issue a cache attach command to associate the local cache with the SES cache (1505).

During mainline operation, access to the database is supported by the buffer manager. Many different protocols can be followed. Examples of using the main line are shown in FIGS. 16 and 17 (1507).
details of). -Retrieve data from DASD-Store data in SES cache-Retrieve data from SES cache-Store data in DASD

Access to data is dictated by a program running on the CPC and making requests to the database manager. For example, such a program can be executed as a result of a request from an end user to execute a transaction program.

When a request is made to the database buffer manager, the buffer manager determines whether the requested data is already in the local buffer (1601). If the data is in the local buffer, the data is checked for validity (160
2). The data in the local buffer may be invalid because another instance of the database manager has performed an update (1703 and 170 in FIG. 17).
5). If the data is valid in the local buffer, an entry used to update the SES data element reference by the reference list processing command is created in the name list (1603).

If the data is not currently valid in the local buffer, a buffer is allocated and a validity indicator is allocated (1604). A read and register command is issued to register the local buffer in the SES cache and retrieve the data if the data is stored in the SES cache (1605). Data is not retrieved from SES cache (test is 160
6), the data is read from the DASD (1608) and the read request is completed by returning the requested data to the caller.

In the case of an update request, the buffer manager updates the contents of the local buffer (1701).
A determination is then made whether to store the data in the SES cache (1702). If the data is to be stored in the SES cache, a write-in-register command is issued (1703). As part of the processing performed by the write-on-register command, all other copies of the data element in the local buffer of the other instance of the database manager are invalidated. If the data should not be stored in the SES cache, the data can be written to DASD (1704). If the data was being written to DASD, a Complementary Invalidate Copy command is issued to invalidate all other copies of the data element in the local buffer of the other instance of the database manager (1705). .

During mainline operation, the buffer manager is also responsible for managing the storage of the SES cache (1508, also detailed in FIGS. 18 and 19). The buffer manager is also responsible for managing the movement of changed data elements from the SES cache to DASD. The timeout interval established during initialization allows the operating system to manage the SES cache area and
In order to manage the transfer of the ES cache data to the DASD, control is periodically passed to the processing routine.

FIG. 18 shows an outline of one protocol for managing the SES cache area. This process is given control at periodic intervals established by the database manager during initialization of use of the SES cache structure. A storage class information read command is issued for each storage class (1801). The hit ratio of each storage class is calculated (1802). The obtained hit ratio is compared with a performance target set for each storage class, and an adjustment value of the storage class size is calculated (1803). For each storage class in which the allocated SES resources need to be changed, a set reuse vector command is issued to activate the reuse vector (1804). The local buffer reference was stored during mainline operation (160
3) A reference list processing command is issued for each storage class (1805).

FIG. 19 shows an outline of one protocol for managing data migration from the SES cache to the DASD.
This process is given control at periodic intervals established by the database manager during initialization of use of the SES cache structure. A read castout class information command is issued for each castout class, and a candidate castout class is selected (1901). A read castout class command is issued for each selected castout class (1902). A list of the most recently updated directory entries is returned. A read command for castout is issued for each selected directory entry (1903). An I / O request is created to cause the data read from SES to be written to DASD (1904). When the input / output operation is completed, a castout lock release command is issued (190).
5).

When all data using the SES cache is no longer accessible, the database manager can disconnect from the SES cache (1509). The database manager transfers all modified data elements from the SES cache to the DAS
D or move the modified data element to SES
You can choose to keep it in the cache. S
If no modified data elements remain in the ES cache, the SES cache structure can be deallocated and can remain allocated for future use. The database manager is SE
An operating system service is called to disconnect from the S-cache structure (1509). A local cache disconnect command is issued to disassociate the local cache from the SES cache. If necessary, the operating system makes the local cache identifier (LCID) available for reassignment. A DEFINE VECTOR CPU instruction is issued to release the cache coherency vector. The operating system service determines if the SES cache structure should be deallocated (1511). If the SES cache structure is to be deallocated, a deallocate cache structure command is invoked (1512). The local buffer pool is then freed by the database manager (1513).

Castout Class Assignment: A castout class is assigned to a data element by a write-when-registered command and a write-and-register command when the data is stored in a modified state. The following is performed as part of the change signal processing.

The castout class is stored in the directory entry, and that directory entry is placed at the bottom of the castout class queue. This may be the initial placement in the castout class, may represent a change in class specification, or may be an update of data in the castout class.

The castout class count is updated. If this is the first placement in the castout class, the castout class count is incremented by one. If this is a class change,
The castout count of the source castout class is decremented and the castout class count of the target castout class is incremented by one. If this is an update in a castout class,
The castout class count is not updated.

When the change signal is the first change signal processed for that directory entry, the user data field is stored in the directory entry. The first change signal occurs when the data is not initially cached, or is cached unchanged. When the data is cached with the first change, the change signal is the second and subsequent change signals, and the user data field is not stored.

Allocation of castout classes is based on the DASD volume on which the data element resides permanently. The data manager using the SES cache assigns the castout class to each DAS
Associate with D volume.

Castout Processing: To begin the castout processing, the castout class count and the total changed count are periodically compared to a threshold. This is done in one of the following ways:

Event-Driven Castout: After each write operation, the counts returned on the write-on-registration command and on the write and register commands are compared. This triggers a castout process for the specified castout class when a write command occurs and exceeds a defined threshold. This comparison is part of the transaction processing.

Timer-driven castout: A set of counters is requested at a timed interval using the read castout class information command. This comparison is part of the background program and not part of any transaction processing. A balance must be maintained between responsiveness counts for castout processing and transaction overhead during acquisition.

In the castout process, the read castout class information command can be used to determine the data element count for a range of castout classes. Based on the count returned and the management protocol used, the castout process for the DASD volume can be initiated by programming. Issue a read castout class information command by programming for each DASD volume to which the modified data element is to be written. Selecting a data element that matches a particular name mask allows programming to select only those data elements that are associated with a particular data set. The data element names and user data fields associated with the selected castout class and name mask are included in the name block returned by the read castout class information command. Select a data element from the data returned by programming, and
Transfer of data to SD can be started.

The castout process retrieves a data element from the SES cache using a read command for castout. The read for castout command sets the castout lock (CO) to ensure that a single process is writing changed data to DASD. When the castout lock is set, the change bit is reset. Ensuring that a single process is writing changed data to DASD prevents race conditions that can cause modifications to the data to be lost. For example: Process 1 retrieves a data element for writing to DASD. -The data element is then modified in the SES cache. Process 2 retrieves the data element for writing to DASD. Process 2 completes writing the data element to DASD. Process 1 completes writing that data element to DASD.

A write operation to a data element with a castout lock set in SES is completed. The change bit can be turned on by this write operation.

The castout / lock release command is
Emitted when a data element has been successfully recorded on DASD. The castout lock release command resets the castout lock. If the change bit was set by a write operation, the change bit remains set at the end of the castout unlock command. If the process performing the castout fails, the unlock castout command can force the change bit to be set. Forcing the change bit to set protects the changed nature of the data element when it is not known whether the data element was successfully written to DASD by programming.

The list cast unlock command provides programming an efficient means of completing the castout process when multiple data elements are written to DASD. This is a cast-out scheme that allows programming to effectively retrieve changed data elements for a single DASD volume and build an efficient channel program to transfer the changed data to DASD. It is a complementary process to having a class queue.

Change Rate Management: The general castout process by the data manager can manage the SES cache so that the changed data is at most about 20%. The program monitors the percentage of changed data in the SES cache and identifies changed data elements as D
The castout process for moving to the ASD is started. The choice of which data elements to move to DASD to maintain less than 20% changed data is a program decision. For efficiency, it is best to select a castout class that contains enough data elements to write a channel program that transfers a large number of physical blocks.

Management of oldest changed data: When data is written and change control indicates changed data, processing of the change signal is started. The change signal is
When it is the first change signal processed for an entry, the user data field is
Stored during entry. The user data field indicates that the data element has changed but the change is still
Used by the program to contain the time value representing the oldest point in time that has not been moved to D. Of all data elements in the SES cache used by the program, the data element with the oldest time value is the one that the program must process log data to recover the database in case of a SES cache failure. Represents the earliest time point. The longer the time interval that the log data requires processing to recover from a SES cache failure, the longer it will take to recover. Thus, in addition to managing the amount of changed data in the SES cache, the program will
It has the advantage of being moved to ASD. To recognize the oldest modified data in the SES cache, process the name blocks returned by the read directory and read castout class commands.

Management of DASD Volumes: Regardless of the percentage of changed data elements in the SES cache or the age of the oldest changed data, programming causes a particular event in a data set to occur when a particular event occurs. You will choose to move the data elements associated with the volume to DASD. For example, when the last user stops accessing a database, the program moves all changed data elements associated with that database to DASD. data·
When a backup of a set or volume is to be performed, the program moves all changed data elements associated with that data set or volume.

Scenario for Managing Data Movement Between DASD and SES Cache: The above mechanism provides the basis for performing management of data movement between DASD and SES cache by system software. Techniques for implementing a non-blocking serialization protocol between writing and casting out data from SES using these primitive services are described below. Using these primitive services, data elements can be
Techniques for implementing a non-blocking serialization protocol for caching during S are described in US patent application Ser. No. 860,805.
No. In addition to these two protocols,
The primitive services described above are converted from SES to DASD in such a way that the overall performance of the
Provides a basis for determining the set of modified data elements that are moved to

Non-blocking serialization for castout: The following usage scenario assumes that the data element referenced in the storage hierarchy is a page. In that case, the shared cache (11 in FIG. 1)
Pages cached during 1) are identified by software-assigned names. Therefore, any read or write request in the shared cache
You need to specify the name of the page for which you are requesting. The data element need not be a page for this protocol to work properly.

A non-blocking serialization protocol for caching data in a shared cache is illustrated using FIGS. In FIG. 20, a page is read using the read and register (RAR) command described in FIG. FIG. 21 is a diagram showing how to use the write and register (WAR) command described in FIG. 12 and FIG. FIG. 22 shows how the write-when-registered (WWR) command described in FIG. 14 is used for non-blocking serialization for casting out data from a shared cache.
Shown in FIG. 23 is a diagram showing how to use the cast-out read command described in FIG. FIG. 2 illustrates the use of the castout / lock release command described in FIG.
It is shown in FIG. FIG. 25 is a process showing the reuse function of the SES cache. FIG. 26 shows the flow of this protocol.

In FIGS. 20, 21, and 22, the local
The collection of valid bits in the cache register (401A, 401B, 401C in FIG. 4) is called a system valid bit vector (SVBV). In this example, SVB
V can be considered to consist of one bit per system added to the cache. System valid bit Bit per system added to the cache. If vector (SVBV) 1, the page cached in the identified system memory is valid. If 0, the page cached in system memory is not valid.

Describing the collection of valid bits in the local cache register as a system valid bit vector is only a simplification of the concept in describing the following protocol. In a preferred embodiment, each bit entry in the system significant bit vector is
There is a one-to-one correspondence with the valid bit field of the local cache register.

In FIGS. 20 to 24 and 26, the local cache identifier (LCID) is called the system identifier (SI). The function of SI is the same as the function of LCID. In the following protocol, SI is local
One field is provided in the SVBV of any entry stored during cache control and subsequently created as a result of processing any entry and join command in the directory.

In the following protocol description, CONNECT
The function of the T (connect) command is satisfied by the add local cache command. READ PAGE
The function of the (page read) command is satisfied by the read and register command (RAR) shown in FIG. W
The RITE PAGE (page write) command is shown in FIG.
2 and the write and register command (WAR) shown in FIG.
Filled by CONDITIONAL WRI
The function of the TE (conditional write) command is satisfied by the write-on-registration (WWR) command shown in FIG.

These commands are based on the possibility that while one system is trying to cache an updated version of a page, another system may be casting out the page, and the data integrity of that page is While supporting non-blocking serialization for casting out pages in the multi-system data sharing complex of FIG. In the practice of the present invention, commands intended for a given data element (eg, a page) are serialized with other commands for the same page by management logic (110, FIG. 1) in memory (101, FIG. 1). Is done.

In this protocol, the memory system 101 in FIG. 1 supports the following commands.

A CONNECT command executed by a software system connected to NV storage device 101, such as the example of a database system in the multi-system data sharing complex of FIG. This CONNECT
In response to the command, the management logic 110 places the identity of the connected system in the local cache control 105 and places it in the SV bit vector of every entry currently in the directory 102 and any subsequently created entries. Provide a field for the connected system.

READ PAGE (S, P, Buffe
r Address) command. Where S identifies the system issuing the command, P identifies the requested page, and Buffer Address indicates the address in the system's buffer where the page should be distributed.

WRITE PAGE (S, P, CB =
1, Buffer Address) command. This command is also referred to as "unconditional" WRITE. WRIT
When the E PAGE command is issued, the parameters entered with this command include the C parameters corresponding to the C fields for the identified page. The updated nature of the page being written is indicated by setting the change bit to one.

[0212] CONDITIONAL WRITE
(S, P, CB = 0, Buffer Address)
command. The CONDITIONAL WRITE command is used by a conditional writing computer system to "cast out" or put pages acquired from secondary storage into NV storage.
C = 0 indicates that the page is unmodified.

READ FOR CASTOUT (S,
P, Buffer Address) command. This command starts a castout process for page P. System S to cast out page P
The DBMS in the middle tells the management logic that P from the NV storage
To the buffer address in system S. The system then writes the page to disk. After disk input / output, DBMS is UNLOCK-C
ASTOUT ID (castout ID lock release)
Issue a command.

UNLOCK-CASTOUT ID
(S, P) command. This command tells NV storage that page P has been successfully written (cast out) to secondary storage.

Page read: FIG. 20 shows the READ PA
5 shows an operation flow of processing of a management logic mechanism performed in response to a GE command. Initially, the management logic 110
(Standby) state 2001, from which UNCOND
ITIONAL WRITE, CONDITIONAL
It can go into any of at least three processes: WRITE or READ PAGE. FIG.
At 0, assume that a READ command of the above form has been received. This is the READ process step 2002
Indicated by When providing a READ PAGE command, the computer system issuing it will include itself (SI, i.e., the i-th system), the requested page (P), and the read page in the computer system to which the read page should be distributed Identify the buffer address. The read process performed by the management logic 110 can have the following three cases.

If an entry for the identified page P is present in the directory 102 and the D bit for that page is set to one, indicating that the requested page is in the shared cache 103, the first A case occurs. These two conditions are tested in decisions 2003 and 2004, respectively. If both determinations are yes, the reading process proceeds to step 2005.
Sets the bit corresponding to the identified i-th computing system (SI) in the SV bit vector of the entry, and returns the data page to the specified buffer address in step 2006, step 2007 Returns the cache hit instruction.

What this case means is that whenever a read request is issued for a page already in the shared cache 103, that page is unconditionally returned to the requestor with a cache hit indication. . The S bit for the requestor is conditioned to a first state (1) to indicate that the copy of page P processed by the system is up to date.

In the second case, assume that there is an entry for page P in directory 102, but that page has not yet been placed in shared cache 103. In this case, following the page reading step 2002, a decision Yes is taken in decision 2003 and the decision 200 is taken.
Take a no route at 4. At this time, the S bit for the requesting system is conditioned to the first state (ie, "1") and a cache miss is issued. In this case, a previous READ PAGE command is received by management logic 110, resulting in a directory entry being created, but the page has not yet been retrieved from DASD and placed in the shared cache.

In the last case, there is no directory entry for page P and management logic 110
Takes a no route in decision 2003 and proceeds to step 201
0, 2011, and 2012 are sequentially executed. In this regard, at step 2010, the management logic creates a directory entry for page P (assuming that there is available storage), and initially populates the entire SV bit vector for the created entry. To the second state (preferably "0"). The D bit is then set to 0, CB is set to 0, and S for the requesting system
The bit is conditioned on the first state. Finally, a cache miss is issued in step 2009 and the logic enters a wait state.

In this last case, the system SI has stored the first READ request for page P, and before that page is placed in the shared cache 103, any subsequent READ requests will follow the procedure steps 2002, 20.
03, 2004, 2008, 2009. Requested page from DASD to shared cache 1
03, the READ request is sent to step 200
2, 2003, 2004, 2005, 2006, 200
7 will be followed.

Page Write: The process used by the management logic to write updated pages to the shared cache 103 is shown in FIG. Page is READ PAGE
It may be updated after being retrieved from the shared cache 103 by a command. Also, after a page is obtained from DASD, it may be changed before being placed in the shared cache 103. Assume that system SI has acquired page P, updated the page, and now has to write the page to shared cache 103. The system SI writes the updated page P,
Issues a WRITE command of CB = 1. This WRI
FIG. 2 shows the processing performed by the management logic in response to the TE command.
It is shown in FIG.

In FIG. 21, there are the following three cases. First, there is an entry for page P in directory 102 and D = 1 (the page is in shared cache 103). Second, there is an entry for the page P, but the page has not yet been entered into the shared cache 103 in any way. Third
Does not have any entry for page P in the directory.

In the first case, the management logic 110
The write process performed by FIG.
2, 2103, 2104, then step 21
05, 2106, 2107 and 2108 are executed. In step 2105, the CB field in the directory entry for page P is set to one. Next, in step 2106, the existing version of page P is
Overwritten with the data at the buffer address supplied with the ITE command. Then, all S bits in the SV bit vector for the data entry for the page are set to 0, except for the S bit for the system issuing the WRITE command. Upon exiting step 2107, the management logic exits step 2107.
At 108, the command acceptance indicator is returned to the requesting system, and then the wait state 2101 is entered.

There is a directory entry for page P, but the page is not in the shared cache.
In the case of, the management logic process 110 goes through the steps 2102 and 2103 from the waiting state 2101 and
In decision 2104, a no route is taken, and step 2109,
2110, 2111, 2107 and 2108 are executed.
In step 2109, the change bit in the directory entry for the requested page is set to one.
At step 2110, space for page P is allocated in shared cache 103, data is moved to the allocated space from the buffer address of system SI,
The page's cache address is placed in the page address field of the page's directory entry. Next, at step 2111, the data bit is set to 1 in the directory entry of page P, and step 2107 is performed, followed by the management logic returning the command accepted indicator and the waiting state 2101
to go into.

In the last case where there is no directory entry for page P, the NO path is taken at decision 2103 and steps 2112 through 2116 and 2108 are performed, and then the management logic enters the wait state 2101. At step 2112, a directory entry for page P is created, and at step 2113 space for the entry for that page is allocated in the shared cache. At steps 2114 and 2115, the relevant bits in the directory entry are conditioned.
In step 2113, all bits in the SV bit vector for the page are initialized to zero and the D bit is set to one. Next, in step 2115, the S bit (bit SI) for the requesting system is conditioned to one, the change bit is also set to one, the page has been modified, and the address of the page is placed in the entry. To indicate that At step 2116, the directory entry for page P is placed in the directory and the page is placed in the cache at an allocated location. The management logic then returns the acceptance indicator and enters a wait state.

The write process of FIG. 21 is unconditional in that the request is never rejected. In addition, no matter which WRITE case is executed, the change bit for the affected page is set and all bits of the SV bit vector for the affected page are zeroed, except for the writing computer system. . As can be seen from the description below for CONDITIONAL WRITE, if the S bit for the non-writing computer system is zeroed during WRITE, then the down-level page will not enter the shared cache.

Conditional write: Next, management logic 110
Reference is made to FIG. 22 to understand the CONDITIONAL WRITE process implemented therein. In this process,
The system SI has received a cache miss from NV storage and then has acquired a page from secondary storage but has not modified it and is preparing to cache it in the shared cache 103. Assume that In the parameter set sent to the management logic 110, the system SI sets the change bit to 0 to indicate that the page owned by the computer system is equivalent to a secondary storage version of the page.
The system SI is a CONDITIONAL of CB = 0.
Issue a WRITE command. Again, there are three cases.

In the first case, the directory entry for the page and the page itself are in the cache. The directory 102 is checked for an entry corresponding to the page P. Assuming the page is in the directory, the yes path is taken from step 2203, and the data bits of the item are examined in step 2204 to determine if the page is in the cache. If the path of yes is taken in decision 2204, the page is in the cache. At this point, the S bit for system SI is examined in the SV bit vector of the page entry to determine if another system has modified the page.
If bit SI is set to 1, no change is made to that page and the decision 2205 takes the yes path. Here, management logic 110 determines that the page is in the cache, the page submitted by the SI is equivalent to the page in the cache, and that overwriting of the page is not required. Accordingly, the management logic 110 proceeds to step 2206 where the CONDITIO
Returns the appropriate code indicating NAL WRITE acceptance.
However, the shared cache 10
Note that no data is transferred to 3. The management logic then enters a wait state. If the NO route is taken in decision 2205, the command is rejected in step 2211 and the management logic enters a wait state.

In the second case, there is a directory entry for page P and a yes path is taken in step 2203. However, at step 2204, the management logic 110 determines that the page is not in the cache, takes a no path at decision 2204, and then determines 22
Assume to go to 07. At decision 2207, page P
Check the SI bit of the SV bit vector for If the bit is set to 1, the system SI owns a valid page. At this point, decision 220
Take the path of Jesus at 7. Space is allocated in the shared cache, and page data is moved to the allocated space from the buffer address of the system SI. In step 2209, the D bit is set to 1 and the CB bit is set to 0. At step 2210, the page address is set in the directory entry and the entry is placed in the directory. Finally, the management logic writes the CONDITIONAL W to the system SI.
Return the RITE acceptance indicator and enter the standby state.

In this case, assume that the result of decision 2207 is no. At this time, the system SI does not own a valid page, the conditional write process ends in step 2211, and the management logic 110 returns to the CONDITI
The rejection of the ONAL WRITE command is returned to the system SI, and then the standby state 2201 is entered.

Finally, the management logic checks the received CO
In response to the NDITIONAL WRITE command,
If it is determined that no entry relating to the page P has been created in the directory 102, a negative route is taken in decision 2203, the command processing ends, and step 2211 is executed.
The command is rejected, then the management logic enters a wait state 2201.

When using these commands, the SV bit vector and the change bit field effectively serialize the cast-in of the page to the shared cache 103 and, after the page is cast-in and changed, the secondary storage. Obviously, a later down-level version of the page obtained from the area guarantees not to overwrite the updated version of the page in the shared cache. This guarantee is driven by a non-blocking serialization protocol for moving pages from secondary storage into a shared cache. This protocol works because if the command causes a cache miss because of a missing directory entry or data, the NV storage device will read the first READ PA
This is because the GE command starts tracking cache operations for that page (in the SV bit vector). An unconditional WRITE command by another system is
Set the change bit for the page,
Reset the S bits of all systems except the writer system in the vector. Subsequent CONDITION
The processing of the AL WRITE command is as follows: a) SV bit
Writing the system S-bit in the vector, or b) no directory entry exists for the page. NV storage rejects the CONDITIONAL WRITE command in both cases.

After an updated page has been written into cache 103, what should be done to ensure that the castout version of the page does not cause the deletion of a newer version of the page in the shared cache? The problem remains. The present invention is directed to a method in which a castout operation in the multi-system data sharing complex of FIG. 1 is performed when a previous version of a page has not been cast out to secondary storage.
READ FOR CA in a protocol that ensures that the updated version of a page is not deleted from
STOUT (read for castout) command and U
Use the NLOCK CASTOUT LOCK command. This guarantee is valid even before the possibility of the page being updated during a castout.

Read for castout: READ FO
FIG. 2 shows the command processing in the R CASTOUT command.
3 is shown. First, the management logic 110 is in the standby state 23
01 and RE from system SI for page P
Transition is made in response to receipt of the AD FOR CASTOUT command. In response to receiving the command,
The management logic checks the CO field of the directory entry for page P. If the CO field is empty (preferably indicated by the value 0), the management logic takes the yes path at decision 2303 and returns to step 23
04 C in directory entry for page P
The B field is set to 0, and in step 2305 the identity of the requesting system is placed in the CO field of that entry. Next, at step 2306, the management logic returns the copy of page P in shared cache 103 to the buffer address specified in the command parameter set and enters a wait state at 2301. At this point, the DBMS in system SI begins the process of writing that version of page P in its buffer to secondary storage.

If the determination 2303 takes the negative path, there is a possibility that another system is performing a castout operation, in which case the other fields of the CO field become non-zero, and the determination 2
At 303, a no route is taken. If the CO field is non-zero, the system is performing a castout operation and the management logic 110 rejects the command in step 2307 and returns to the wait state 2301.

Release of castout lock: FIG.
The processing of the management logic for the UNLOCK CASTOUT LOCK command is shown. First, management logic 1
10 is in a standby state 2401. U in step 2402
NLOCK CASTOUT LOCK command (UN
Upon receipt of (LOCK CO), the management logic compares the entry in the CO field of page P with SI, the identity of the system issuing the command. If the values are equal, the requesting system returns a READ FOR
Indicates that the system has started the castout process by the CASTOUT command. If they are equal, the decision 2403 takes a yes path, the CO field is set to 0 in step 2404, and the management logic enters the wait state 2401. If it is determined that the CO value and the system ID value are not equal and there is a possibility of a software error, a NO route is taken in decision 2403, the command is rejected in step 2405, and the standby state 2401 is entered again.

FIG. 25 shows a process for deleting a page cast out using the commands shown in FIGS. 23 and 24. Relatively speaking, "delete"
Means the removal of a directory entry and the dedication of the cache space pointed to by the directory. We have determined that any suitable D inside the management logic 110
It is intended to use the ELITE process to dedicate the cache space holding page P and remove its associated directory entry from the directory. Inevitably, the deletion process is a READ FOR
CASTOUT / UNLOCKCASTOUT LOC
It will depend on the completion of the K sequence, followed by the removal of the castout page P.
FIG. 25 shows how the castout process prevents pages updated by a WRITE command during that process from being deleted. FIG.
5 does not assume that the deletion always occurs immediately after the castout.

In FIG. 25, when a directory entry is to be created but there are no free entries available, a DELETE process is called within management logic 110 to delete page P from shared cache 103. It is. This is step 2502. The delete process first checks the CB field of the directory entry for page P, and then
Check the O field. This is judgment 2503 and 25
04. In decision 2503, if the change bit is set to one, an indication is drawn that the version of the page P in the shared cache does not match the version in secondary storage. In this case, a castout process must be started to write the shared cache version to secondary storage. In this case, decision 250
Take a no path at 3. In decision 2504, READ F
The execution of the OR CASTOUT command initiated the castout process, but the subsequent UNLOCK
The execution of CASTOUT LOCK may not have completed the process. In this case, the CO field is non-zero, and taking a No path in decision 2504 will end the delete process. The no path from decision 2503 and 2504 is step 25
At 05, the entry will not be deleted. If the determinations 2503 and 2504 take the path of Yes, respectively, the conditions that the shared cache version and the secondary storage version match, and the completion of the castout process are satisfied.
Directory entry for directory 102
Will be deleted. At this point, the appropriate cache management policy could be invoked to steal the cache space allocated to page P.

Non-blocking castout process flow: FIG. 26 shows a READ FOR CASTOU
T (RFC) command and UNLOCK CASTOUT
9 illustrates the operation of the castout process using the LOCK (UNLOCK) command. Guaranteeing that the version of the shared cache of the page P matches the version of the secondary storage is indicated by a WRITE command executed between the RFC command and the UNLOCK command during this operation. This operation is referred to as the directory
An example will be given regarding the expression of an entry. In this expression, only the page identification P, change bit CB, and CO field of the entry are shown. Directory for page P
The entry is indicated by reference numeral 2601. In addition, the time evolves vertically along the arrow marked TIME.

In FIG. 26, the system S1 performs the READ
Assume that the castout process is initiated by issuing a FOR CASTOUT command 2602. Thus, the RFC process shown in FIG. 23 starts. In processing the command, the NV storage management logic sets the change bit to 0, and sets the identification (S1) of system 1 to the CO entry of the page P directory entry.
Put in the field. The management logic returns a copy of page P from the shared cache to a buffer in system S1. The system S1 then proceeds to 2605 to write a copy of page P to secondary storage.
Start the SK I / O process. Meanwhile, after the completion of the RFC process and before the completion of the WRITE TO DISK process, the system S2 executes the WRITE command 2606.
Execute The system S2 indicates that the CB value of the directory entry of the page P is 1, and the CB field of the entry 2601 is changed to this value during the WRITE process 2607 performed by the management logic according to FIG. At this point, an updated version of page P that is being written to secondary storage is present in the shared cache, and a mismatch occurs between that version and the version cast out by system S1.

Shortly after the execution of the WRITE command 2606, the WRITE DISK 26 in the system S1 is executed.
05 is completed (2608). At this time, the system S1
Is UNLOCK CASTOUT LOCK command 2
Issue 609 and start the corresponding processing by the management logic at 2610. During operation 2610, the management logic zeros the CO field of directory entry 2601, thereby unlocking page P for another castout process. The important point to note here is UNLOCK CASTOU
T LOCK command returns directory entry 260
That is, the CB field in 1 is not reset.
This causes the management logic 110 as shown in FIG.
Does not delete page P. The newer version of P will be cast out by a cast out process that is later started by S1 or S2. Thus, if the management logic attempts to delete page P in step 2611, the deletion of page P is prevented and another page is considered for deletion.

[Brief description of the drawings]

FIG. 1 illustrates a sysplex system including multiple CPCs and one SES device.

FIG. 2 illustrates a single CPC (Central Processor Complex) as found in the prior art.

FIG. 3 shows a preferred form of an entry in the SES directory.

FIG. 4 shows a SES local cache register (LC)
It is a figure showing preferred form of an item in R).

FIG. 5 is a flowchart illustrating a read castout class command.

FIG. 6 is a flowchart illustrating a castout class information read command.

FIG. 7 is a flowchart illustrating a directory read command.

FIG. 8 shows a directory entry information block (DEI).
It is a figure showing the format of B).

FIG. 9 is a flowchart illustrating a castout read command.

FIG. 10 is a flowchart showing a castout / lock release command.

FIG. 11 is a flow chart illustrating a read and register command.

FIG. 12 is a flowchart showing a read and register command when SES change control indicates an unchanged record state.

FIG. 13 is a flowchart showing a write and register command when the SES change control indicates a changed record state.

FIG. 14 is a flowchart showing a write command at the time of registration.

FIG. 15 provides an overview of the buffer manager of the database manager and the associated flow of use of the SES cache.

FIG. 16 is a flowchart of a buffer manager process of a read operation.

FIG. 17 is a flowchart of a buffer manager process of a write operation.

FIG. 18 is a flowchart of SES cache area management of the buffer manager.

FIG. 19 is a flow diagram of the transfer of SES cache data to DASD by the buffer manager.

FIG. 20 is a flowchart showing a processing flow of a read command supporting a non-blocking castout protocol.

FIG. 21 is a flowchart showing a processing flow of an unconditional write command supporting a non-blocking castout protocol.

FIG. 22 is a flowchart showing a processing flow of a conditional write command supporting a non-blocking castout protocol.

FIG. 23 is a flowchart showing a processing flow of a read command for castout supporting a non-blocking castout protocol.

FIG. 24 is a flowchart showing a processing flow of a castout lock release command supporting a non-blocking castout protocol.

FIG. 25 is a flowchart showing a processing flow of a storage reuse process that supports a non-blocking castout protocol.

FIG. 26 is a flowchart illustrating a method of a non-blocking castout protocol for caching data in a shared cache.

[Explanation of symbols]

 101 SES Cache (High Speed Non-Volatile Memory) 102 Directory 103 Shared Cache 104 Local Cache Register 105 Local Cache Control 106 High Speed Link 107 Local Cache Buffer 109 Direct Access Storage (DASD) 110 Management Logic 111 Shared Cache

Continued on the front page (72) Inventor Jeffrey Alan Fry United States 12524, Fishkill, NY, Greenhill Drive 24 Ai (72) Inventor Chandrasekalan Mohan United States 95120, San Jose, CA, Portswood Drive 727 (72) Inventor Indell Pal Sing Nalang, USA 95070, Serra Oaks Court, Saratoga, California 13778 (72) Inventor Jeffrey Mark Nick United States 12524, Fishkill, NY, Plymouth Road 43 (72) Inventor Inventor Jimmy Paul Strickland United States 95070, Allcott Way, Saratoga, California 18929 (72) Inventor Michael Dustin Swanson 12603, United States of America Pokiepsi, New York , College Avenue 95 (58) investigated the field (Int.Cl. ^6, DB name) G06F 12/08 G06F 12/08 310 G06F 12/12

Claims (9)

(57) [Claims] A method for controlling data transfer from a shared storage area to an input / output device in a central processing unit complex (CPC), comprising: a plurality of central processing unit complexes (CPC);
Shared electronic storage (SE) having shared storage shared by C
S) configuring a sysplex to include a cache and a direct access storage device (DASD) connected to the CPC; and controlling the direct access storage device (DASD) to store data items in the sysplex. When a data item indicated by a directory entry in the SES cache is changed, a castout class for the directory entry in the SES cache is associated with the changed data item, and the changed data item is selected by the castout class. Enabling a plurality of data items to be transferred to DASD with a single input / output instruction. 2. A method for controlling data transfer from a shared storage area to an input / output device in a central processing unit complex (CPC), comprising: a plurality of central processing unit complexes (CPC);
Shared electronic storage (SE) having shared storage shared by C
S) configuring a sysplex to include a cache and a direct access storage device (DASD) connected to the CPC; and controlling the direct access storage device (DASD) to store data items in the sysplex. When a data item indicated by a directory entry in the SES cache is changed, a castout class for the directory entry in the SES cache is associated with the changed data item, and the changed data item is selected by the castout class. Enabling a plurality of data items to be transferred to DASD with a single I / O instruction; and a user data field for indicating a time at which the associated data item is first written into the SES cache. Directory Associating an associated data item with a copy of the same data item in the DASD to indicate when a change flag field is set in the directory entry. Writing to a field; and a data transfer control method comprising: 3. Management memory means for acquiring data from the memory in response to a READ command, generating data corresponding to the data blocks in the memory, and deleting the data blocks from the memory. At least one slower than said memory for storing data
A computer system comprising: a storage mechanism; and a plurality of processing systems connected to the memory and the storage mechanism, wherein the computer system serializes removal of data stored in the storage mechanism from the memory; Data in the memory generated by the management logic means;
When the corresponding data of the block includes a change bit field and a castout identification field, and when a READ command is supplied from the first processing system to obtain a data block to be put into the storage, the READ command is issued. In response to the command, a change bit field is set to indicate a match between the version in the memory of the data block and the version in the storage, and the value identifying the first processing system is set to a value. A computer system for setting a castout identification field of corresponding data and responsive to a value in the castout identification field to prevent deletion of a block of data from the memory. 4. A WRI from a second processing system for changing a version of the data block in memory.
When the TE command is issued, the modified bit field is changed in response to the WRITE command to indicate a mismatch between the version in memory and the version in the storage of the data block; 4. The computer system of claim 3, wherein in response to the inconsistency indicated therein, the data block is prevented from being deleted from memory. 5. A data definitive system of claim 4
A transfer control method, wherein the first processing system has a data
Deleting a value from the castout identification field when entering a block into the storage; and (a) updating the data block in memory after preventing the deletion of the data block from memory. READ from the processing system to get the version
Providing a command; and (b) in response to the READ command of step (a),
A change bit to indicate a mismatch between the version in memory of the data block and the version in storage.
Setting a field and setting a castout identification field of the corresponding data of the data block to a value identifying the processing system; and (c) storing an updated version of the data block in the processing system. (D) removing a value identifying the processing system from the castout identification field in response to writing the updated version of the data block to storage in step (c); e) if the change bit field indicates a match between the version in memory of the data block and the version in storage, delete the data block from memory; otherwise, delete the data block. Steps (a)-(e) until the block is deleted from memory
Performing a data transfer. 6. A cache, management means for controlling access to a data block in the cache and deleting a data block from the cache, a storage mechanism slower than the cache, A method for controlling removal of data from a cache in a multi-computer data sharing system including a plurality of processing systems connected to the cache, wherein the first storage system includes the storage mechanism. Reading a data block from a cache; and modifying the data block in the cache by a second processing system while the first processing system is placing the data block into storage. The deletion of the data block from the cache. And, when there is no change, the data recovery control method comprising the step of deleting the data block from the cache. 7. A cache, management means for controlling access to data blocks in the cache and deleting data blocks from the cache; a storage mechanism slower than the cache; In a multi-computer data sharing system including a plurality of processing systems connected to a cache, a method for controlling removal of data from the cache, comprising: a processing system having a data block to be input to the storage mechanism; The data in the cache, including a castout lock field having an identification (ID).
Creating corresponding data for the block; (a) preventing the deletion of a data block possessed by the processing system when the castout lock field includes an identification of the processing system; (b) C) preventing removal of the data block from the cache if the cached version of the data block does not match the version in the storage of the data block; A data removal control method including a step of deleting a block. 8. The data block until the data block is deleted.
The method according to claim 7, wherein steps (a), (b) and (c) are repeated in order. 9. A memory, management logic means for obtaining data from the memory in response to a READ command and maintaining a directory having corresponding data of data blocks in the memory, and slower than the memory for storing data. And a plurality of processing systems connected to the memory and a plurality of processing systems connected to the memory, wherein a data transfer control method including a READ process for serially moving data blocks from the memory to the storage mechanism. Processing in the memory having a change field indicating whether a data block in the memory matches a version of the data block in a storage, and a data block to be input to the storage. A data block in the memory including an identification (ID) field for identifying the system. Creating corresponding data for the data block; issuing a request from the first processing system for inputting a data block from the memory to the storage mechanism; responding to the request; Setting a change field to indicate a match between the version in memory and the version in the storage mechanism, and entering the identification (ID) of the first processing system into the identification field. Entering the data block into the storage; removing the first processing system identification (ID) from the identification field after the data block is entered into the storage; A data block from the memory if it indicates a discrepancy between the data block and the data block in storage. Preventing removal of the block, otherwise, the data transfer control method comprising the step of deleting the data block from the memory.