KR19990047097A - How to Manage Disk Cache on High-Speed Parallel Computers - Google Patents

Abstract

The present invention relates to a disk cache management method of a high-speed parallel computer running in an input / output processing node to improve the input / output performance.

In the present invention, a microkernel-based operating system (MISIX) matches various types of administrators functionally serverized by mutual message communication. The file manager managing user files issues messages for input / output requests to the disk cache manager. The disk cache manager has a two-step cache management method consisting of lines and blocks as caching units. The line is composed of the most effective units when performing I / O from the physical device, and the block serves as the basic unit of input / output of the operating system buffer cache.

Accordingly, the present invention provides a disk cache management method of a high-speed parallel computer that improves overall performance by inserting adjacent blocks into a disk cache to increase the cache hit ratio and minimize the cost of physical I / O due to cache failure.

Description

How to Manage Disk Cache on High-Speed Parallel Computers

The present invention relates to a disk cache management method of a high-speed parallel computer running in an input / output processing node to improve the input / output performance.

In a typical computer system, data stored on a disk is loaded into a main memory to quickly respond to a user's input / output request. The operating system's buffer cache was used for this purpose, but the failure of the buffer cache due to frequent I / O requests involved physical I / O. Due to the physical I / O, applications that expect I / O performance such as OLTP show a serious I / O bottleneck, which is a fundamental cause of performance degradation. Therefore, disk cache was used as the secondary cache.

However, such a conventional disk cache has a limited capacity and a management method similar to the operating system buffer cache, causing frequent physical I / O at the time of cache failure and fetching, and thus, it is recognized as a role of expansion of the buffer cache.

In addition, the preloading of the conventional block-based cache management drives a separate I / O command, and since the physical I / O unit is a sector unit of the disk, the cost due to the cache failure increases.

Accordingly, the present invention provides a microkernel-based disk cache manager at the input / output node to increase the cache hit ratio through a two-step cache management method for lines and blocks, and to reduce the disk access time per unit input / output using an effective physical input / output unit. The purpose is.

The present invention for achieving the above object is a first step of determining whether the cache line requested from all cached lines exist, and if the cache line does not exist immediately to determine whether it is a write and write request, and as a result of the check immediately write and write request In this case, the second step of setting the write unit to the block and calling the block I / O manager (hereinafter referred to as BIOM) and ending, and if the read result is not to write immediately, checks whether the read unit is performed. A third step of setting a line, calling a block input / output manager (BIOM), and then setting a line and a block state flag to be valid (hereinafter referred to as VALID) and terminating; The fourth step of proceeding to the delayed write processing step and ending; and the cache line zone requested in all the cached lines. Whether the cache line exists and proceeds to the invalidated line state processing step (hereinafter referred to as INVALID); and if the failure is confirmed from the step, the process proceeds to the third step. Otherwise it comprises the sixth step of terminating.

1 is a block diagram of a high-speed parallel computer system to which the present invention is applied.

2 is a configuration diagram of an input / output node to which the present invention is applied.

3 is a disk cache matching diagram to which the present invention is applied.

4 is a main flow diagram of a disk cache to which the present invention is applied.

5 (a) to 5 (f) are an overall flowchart of a disk cache management method to which the present invention is applied.

<Description of Signs of Major Parts of Drawings>

101: X-bar interconnection network (Xcent Net) 102 and 201: processor node (PN)

103: input / output node (ION) 104: communication control node (CCN)

105: Xcent Network Matcher (XNIF)

106 and 203: Peripheral Interconnect Bus (PCI Bus)

107: block device controller (SCSI) 108: peripheral device (DEV)

202: shared memory 301: micro kernel (MK)

302: Processor Manager (PM) 303: File Manager (FM)

305: Disk Cache Manager (DCM) 306: BIOM

307: User process

A high-speed parallel computer (hereinafter referred to as SPAX) is a symmetric multiprocessor (Symeritric Multi) with one node connected to four Intel P6 processors 201, shared memory 202, and local resources 204, and a P6 bus 203. Processor (hereinafter referred to as SMP) structure, which is interconnected to an Xcent Net 101 to form a cluster, and is a parallel processing system that constitutes an entire system as 16 clusters. Each node is equipped with various functionally modular managers 302 and 303 based on the microkernel 301 to operate and communicate with each other by way of message transmission.

The nodes are functionally divided into the processor node 102, the input / output node 103 and the communication control node 104, which are logically divided by the type of devices connected to the PCI bus 106 inside the node. The I / O node 103 mounts SCSI controllers on four PCI buses, and the communication control node 104 mounts communication controllers on four PCI buses to transmit messages as dedicated servers managing block I / O devices and character I / O devices, respectively. A server node that controls the processing and ordering of matching and input / output request units.

Therefore, in order to improve the input / output performance of the storage device 108 mounted on the various input / output nodes 103 in a parallel processing system such as SPAX, a two-stage cache management method of caching units in lines and blocks is provided. In addition, since the preparation time, seek, and latency time of the disk access determine the total access time by using the line unit as the physical input / output unit, the time according to the size of the data transfer amount is almost ignored until a certain amount. By devising the I / O, we devise a method to increase the hit ratio of the disk cache as well as the prefetching effect.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a high-speed parallel computer system to which the present invention is applied, a processing node (PN) 102 having four central processing units and internal resources, an input / output node (ION) ( 103 and a Communication Control Node (CCN) 104 are connected to an Xcent Net 101, which is a high-speed interconnection network, to communicate in a mutual message exchange. Each node is connected to a Block Computer Controller (SCSI) 107 and a Peripheral Devices (DEV) 108 mounted on a Peripheral Component Interconnect Bus (PCI Bus) 106. As a kind, they are classified into the input / output node 103 and the communication processing node 104.

2 is a structural diagram of an input / output node to which the present invention is applied, in which one node is connected to four Intel P6 processors 201, shared memory 202, and local resources 204, and a P6 bus 203. A processor (SMP) structure is shown.

The processor 201 uses Intel's P6, and four PCI buses 203 may be connected to a block device (SCSI) controller 107 to connect various types of block devices. Memory 202 has a capacity of 1 gigabyte and is used for disk caches and kernel regions containing a microkernel and several managers.

3 shows a control flow between nodes when disk input / output is requested in the system.

When a file input / output request is made to the kernel by the user process 307, the processor manager (PM) 302 of the PN determines a given device number and requests the file manager (File Manager: FM) 303. The file manager 303 manages file I / O, and if it fails to find a desired file in its buffer cache, it sends an I / O request of a desired disk block to a disk cache manager (DCM) 305 which it manages. Notify by message. The disk cache manager (DCM) 305 of the I / O node converts the request block into a line and finds it in the disk cache, and if successful, terminates the request, regresses, and if it fails, interfaces with the BIOM 306 to physically control and data messages. Request I / O At this time, the unit of physical input / output is provided, and the intermediate memory copy is eliminated by setting the disk data exchange area to the disk cache and the device.

4 shows the main processing flow diagram of the disk cache management method.

This is the main flow diagram that is processed when an input / output request arrives as a message in step 401, calling the search function in step 402 and ending in step 403.

5 (a) to 5 (f) show an overall flowchart of the cache management method.

Fig. 5A shows the main flow of the disk cache search.

In step 501, it is determined whether there is a cache line for the requested block in all cached lines. If it is not present, in step 502, it is determined whether the cache policy is write-through and write request. If it is an immediate write and a write request, the write unit is set to a block in step 503, and the BIOM for physical input / output is called in step 504 to end.

If it is not an immediate write in step 502, it is determined whether a read request in step 505. If it is a read request in step 505, the read unit is set to a line in step 506, the BIOM is called in step 504, and the line and block status flags are set to VALID in step 507 to end.

In the case of the write request of the delayed write in step 505, the process terminates after the process of receiving a new cache line or replacing the existing cache line in step 508.

If there is a line of the block requested in the cache line in step 501, the process is performed according to the line status flag in step 509, and if the result fails in step 510, the process proceeds to step 505. If it fails, it exits.

5 (b) shows an INVALID line state processing flowchart.

When the cache line requested in step 501 exists, it is determined in step 511 whether the line status flag is INVALID, and if the current line status flag is INVALID, in step 510 a failure is recorded and terminated. In step 512 the next possible line status flag is examined, processed and terminated.

Fig. 5C shows the flow of write processing for delayed write in the absence of a cache.

If it is a delayed write in step 505, it is determined whether a line is allocated from the free list in step 513, if a line is allocated, a write is performed in the corresponding block in step 514, and a line status flag in step 515. Is set to PVALID, the corresponding block status flag is set to DIRTY, and the remaining block status flags are set to INVALID. If the line is not allocated, a replacement target is selected and replaced in the current cache line in step 516, and the flow proceeds to step 514.

Fig. 5 (d) shows a flow chart of dirty (hereinafter referred to as DIRTY) line state processing.

In step 517, it is determined whether the current line status flag is DIRTY. If not, it proceeds to step 523. If the step is DIRTY, in step 518, it is checked whether the current block status flag is DIRTY. If the step is DIRTY, it is determined whether it is a write in step 524. If the read is not a write, the read is performed in the cache and ends. If it is a write in step 524, update the block in step 525, set the write unit to a line in step 526, call BIOM in step 528 to complete the physical write, and then step 528 Exit by setting the line and block status flags to VALID.

If it is not DIRTY in step 518, it is determined in step 519 whether the current block status flag is INVALID, and if the step is not INVALID, processing for it is terminated in step 529. If it is INVALID, the read unit is set to block in step 520, the BIOM is called in step 521, and the block status flag is set to VALID in step 522 to end.

5 (e) shows a processing flowchart of the PVALID line status flag.

In step 530, it is determined whether the line status flag is PVALID. If it is not PVALID, it is an undefined status flag value. In step 533, an error is recorded and ends. If the line status flag is PVALID in step 530, it is determined whether the block status flag is INVALID in step 531. If the current block status flag is INVALID, the process returns to set failure in step 532 and terminates. If the current block status flag is VALID in step 534, the process ends according to the cache policy and the request.

5 (f) shows a flowchart for performing a processing procedure according to a cache policy and a request type.

If in step 531 the block status flag is not INVALID, then in step 535 it is determined whether it is an immediate write policy and a write request; if it is an immediate write policy and a write request, in step 536 the cache block is updated, and In step 537, the BIOM is called in block 538 as the write unit, and the writing is completed immediately.

If it is an immediate write policy in step 535 and not a write request, it is a read request or a write request of delayed write, so it is determined whether to write in step 539. If it is a read instead of a write, the block is taken and ends. If it is a write, the block is updated at step 540, and at step 541, it is determined whether the current line status flag is DIRTY. If the line status flag is not DIRTY, it is set to DIRTY in step 542, and in step 543, the block status flag is set to DIRTY, and if the line status flag is DIRTY, step 543 is reached.

As described above, the present invention is a two-step cache management method in which a disk and a block are separated into lines and blocks in a large disk cache that guarantees the integrity of data during input / output of a block device. Minimize the cost of physical I / O due to cache failure. Therefore, it is possible to selectively use delayed writes and immediate writes that adapt to application areas while improving overall performance, and thus has an excellent effect on widening user selection.

Claims (6)

A first step of determining whether the cache line requested from all the cached lines exists, and if the cache line does not exist, to immediately read and write the request; A second step of setting the write unit to a block and calling the block I / O manager when the check result is an immediate write and a write request; A third step of checking whether it is a read or not if the read result is a read request and setting a read unit as a line, calling a block I / O manager, and setting a line and a block status flag as valid if the read request is completed; A fourth step of, if the check result is a delayed write request instead of a read request, proceeds to the delayed write processing step and ends; A fifth step of determining whether there is a cache line requested from all the cached lines, and if the cache line exists, proceeding to an invalidated line state processing step; And a sixth step of checking whether there is a failure from the step, and if the failure occurs, proceeding to the third step, and terminating the failure. 2. The process of claim 1 wherein the invalidated line state processing step Determining whether the line state flag is invalidated, and if the line state flag is invalidated, recording the failure and ending the first step; And a second step in which the line state flag proceeds to an in-validated or dirty line state processing step. The method of claim 1, wherein the delayed write processing step is After checking whether the line is allocated to the free list, if the line is allocated, the block is written to the corresponding block, and the line state flag, the corresponding block state flag, and the remaining block state flags are set as provisional, dirty, and invalidated, respectively. The first step of terminating, And a second step of selecting and replacing a replacement target in the current cache line and repeating writing to the corresponding block if the line is not allocated as a result of the checking. 3. The dirty line state processing step of claim 2, wherein A first step of checking whether the current line status flag is dirty and proceeding to the provisional line status processing step if it is not dirty, and if dirty, a first step of checking whether the current block status flag is dirty; A second step of checking whether it is dirty if the result of the checking is read, and if the read is not a write, the process terminates after reading from the cache; A third step of updating the block if the check result is written, setting a write unit to a line, calling a block input / output manager (BIOM) to complete a physical write, and then setting the line and block status flags to valid Wow, A fourth step of checking whether the current block status flag is invalidated if the result of the checking in the first step is dirty, and processing and ending in a processing step according to a cache policy and a request if not invalidated; And if the check result is invalidated, calling the block input / output manager (BIOM) with the read unit set to the block, and then setting the block status flag to valid and ending the fifth step. How to manage disk cache. 5. The process of claim 4, wherein the provisional line state processing step A first step of checking whether the line state flag is provisional, if not, recording an error and ending, and if the line state flag is provisional, checking whether the block status flag is invalidated; And a second step of setting and returning failure if the block status flag is invalidated, and ending if the block status flag is invalidated, and performing processing according to a cache policy and a request if not invalidated. How to manage disk cache on your computer. 6. The method of claim 4 or 5, wherein the cache policy and the processing according to the request A first step of checking whether it is an immediate write policy and a write request, and if the immediate write policy is a write request, updating the cache block and invoking the block I / O manager with the write unit as a block to immediately complete and terminate the write; A second step of taking a block if the read policy is an immediate write policy and the write request is not a write request, and if the read is not a write, the terminal is checked; A third step of updating the block if the check result is written, checking whether the line status flag is dirty, setting the line status flag to dirty if it is not dirty, and ending the process by setting the block status flag to dirty Wow, And if the line state flag is dirty, proceed to a fourth step of proceeding to the line state flag.