JPH09185462A - Data look-ahead method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accelerate filed data access, etc., by enabling the data look-ahead of random access case. SOLUTION: A CPU 10, main storage device 20 and secondary storage device 30 are connected through a system bus 40. An application program(AP) 21, operating system(OS) 22, AP buffer 23, system buffer 24 managed by the OS and specified storage area 25 exist in the main storage device 20. The plural pairs of the numbers of blocks with which the secondary storage device 30 is accessed at random in the past by the AP 21, and the numbers of look-ahead data are held in the specified storage area 25. Corresponding to the contents in the specified storage area 25, an access pattern analytic part 220 of OS 22 analyzes the access pattern of AP 21, and the look-ahead of data block to be accessed in near future from the secondary storage device 30 to the system buffer 24 is performed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pre-reading method of file data or the like stored in a secondary storage device in an electronic computer system having a randomly accessible secondary storage device.

[0002]

2. Description of the Related Art In the current computer system,
There is a large difference in access speed between the main storage device and a secondary storage device such as a hard disk (HD). Generally, file data is stored in a secondary storage device such as an HD, and when accessing this, it is necessary to read it once into the main storage device. Most of the file access overhead is
This is due to the data transfer from the secondary storage device to the main storage device. In order to reduce this overhead, conventionally, only when the file data on the secondary storage device is sequentially accessed by the application program (AP), only the data for one block next to the last read block, The operating system (OS) uses a pre-reading method of transferring from the secondary storage device to the main storage device in advance. According to this, at the time when the AP issues a read request for the next block, the data transfer of the block is already started from the secondary storage device to the main storage device, and the file access can be speeded up.

[0003]

The above-mentioned prior art is not considered in terms of prefetching other than sequential access, and performance is remarkably deteriorated in random access other than sequential access in which so-called file data or the like is discontinuously accessed. There was a problem of doing.
An object of the present invention is to enable prefetching of data even in an access pattern which was conventionally regarded as random access and to speed up file data access and the like.

Further, in the above-mentioned conventional technique, only one block is prefetched, and when a hard disk is used as the secondary storage device, there is a problem that seek or rotation waiting of the disk head frequently occurs. An object of the present invention is to reduce the seek and rotation waiting of the disk head and improve the data transfer efficiency.

Still another object of the present invention is to reduce the damage caused when the prefetching prediction is wrong and the prefetched data is not accessed by the application program. Another object of the present invention is to enable prefetching of file data and the like from the AP by the same procedure as the conventional access method.

[0006]

[Means for Solving the Problems] What is random access?
By accessing data in non-contiguous offsets such as files, everything is by no means random. An application program (AP) that handles a database or the like is often random access, but often has some regularity.

Therefore, in the present invention, the OS analyzes the data access pattern of the AP that accesses the data stored in the secondary storage device continuously / discontinuously to predict the data to be accessed in the near future, The data is prefetched from the secondary storage device to the main storage device. Further, in the present invention, the amount of data to be read in advance is dynamically changed according to the prediction result. Specifically, a plurality of sets of addresses of previously accessed data of the secondary storage device and the number of prefetched data are held in a specific storage area, and when the access request to the secondary storage device is made, the specific storage area is stored. Based on the contents of, the data to be accessed in the future is predicted and the amount of data to be read ahead is calculated.

By performing pre-reading for a data access pattern that can be predicted to be accessed in the future with a high probability, the data has already been transferred to the main memory or is being transferred when the pre-read data is accessed by the AP. Become. Thereby, the waiting time for data transfer from the secondary storage device of the AP to the main storage device is shortened,
It is possible to speed up file access and the like.

Further, as the probability of predictive accuracy increases, the amount of data to be read ahead at one time is gradually increased. Since the one-time read-ahead data is collectively transferred from the secondary storage device, the disk rotation waiting time and seek time are reduced, and file access can be speeded up.
Further, when the probability is low, the amount of data to be prefetched at one time is relatively small, so that it is possible to reduce the damage when the AP does not access the prefetched data due to a wrong prediction.

Further, the AP reads a file, data, etc. by issuing a READ system call to the OS. However, the prefetching is realized by the OS without changing the AP interface of the READ system call.
Does not need to change the conventional access method.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a schematic configuration diagram of the entire system according to an embodiment of the present invention. In the figure, the central processing unit (CP
U) 10, main storage device 20 and hard disk (HD)
The secondary storage device 30 is connected to the system bus 40. On the main memory 20, various application programs (AP) 21, operating system (OS) 22, buffer for AP (AP buffer) 2
3. OS managed buffer (system buffer) 24
And so on. Further, the secondary storage device 30 has a disk cache mechanism 31. These configurations are the same as the conventional one. The configuration related to the present invention includes an access pattern analysis unit 220 for prefetch control as one of the functions of the OS 22, and an access pattern analysis unit 22 of the OS 22.
This is a specific storage area 25 newly provided on the main storage device 20 for use with 0. As will be described later, the data (block) prefetched from the external storage device 30 is
Held in the system buffer 24 managed by
When the READ system call for the data is issued from 1, the system buffer 24 is copied to the AP buffer 23.

FIG. 2 is a conceptual diagram of the system buffer.
The system buffer has a configuration in which a plurality of system buffers are chained with the cue head at the head. The system buffer is used when issuing a READ / WRITE request from the AP or a read-ahead READ request by the OS described later.
It is assigned by the OS. One system buffer holds one block of data. The size of one block is usually 8 Kbytes. Each system buffer includes a queue chain pointer of the system buffer, a flag indicating completion of data transfer and an access type (READ / WRITE), a block number of a READ / WRITE request,
It holds information such as a memory address indicating an actual data storage location on the main storage device (memory). In the unallocated system buffer, the flag, block number, memory address, etc. are empty. As described above, the system buffer is a general term for an area that actually stores data and an area that stores its management information.

FIG. 3 is a diagram showing an example of the structure of the specific storage area 25. The specific storage area 25 has N sets of storage areas, with a READ block number storage area 251 and a prefetch block number storage area 252 as one set. The READ block number storage area 251 stores the block number requested by the AP for READ. The block here is a unit of data read from the secondary storage device, and here is 8K.
B. The block number is the block serial number (block address) from the beginning of the file stored in the secondary storage device. When the AP makes a READ request for a further continuous block, the block number in the READ block number storage area 251 is updated to the continuous block number. R is stored in the prefetch block number storage area 252.
The number of blocks read ahead when the block having the block number stored in the EAD block number storage area 251 is read is stored. Therefore, the READ block number storage area and the prefetch block number storage area in a certain storage area have a one-to-one correspondence. Since the specific storage area 25 has N sets of storage areas, a maximum of N non-consecutive block numbers in the file can be stored, and the number of prefetched blocks is stored corresponding to each. be able to.
This makes it possible to perform prefetching when accessing non-contiguous blocks (random access),
Further, the number of prefetch blocks can be made variable. In the past, since there was only one READ block number storage area, only the last block number that the AP requested to read was always stored, and there was no prefetch block number storage area. Therefore, prefetching can be executed only in continuous block access (sequential access), and the number of prefetching blocks is only one.

Next, the look-ahead algorithm of the present invention is
This will be described with reference to the flowchart of FIG. The maximum number of prefetch blocks is eight.

In FIG. 1, when the AP 21 issues a READ request with a block number M by a READ system call, the access pattern analysis unit 220 (hereinafter
(Simply referred to as OS) determines whether the block number M-1 is registered in any of the READ block number storage areas 251 in each storage area of the specific storage area 25 (step 401). If registered, this REA
Since the D request is an access to a block that the AP 21 has read in the past and a continuous block, it is predicted that the next continuous block is likely to be accessed in the future, and the prefetch process is performed. First, the block number of the READ block number storage area 251 of the storage area is rewritten from M-1 to M (step 402). Next, R of the storage area
It is determined whether the number of blocks registered in the prefetch block number storage area 252 corresponding to the EAD block number storage area 251 is less than 8 (step 403). If it is less than 8, 1 is added to the number of blocks in the prefetch block number storage area 252 (step 404). Next, the prefetch is executed from the block number M + 1 for the number of blocks registered in the prefetch block number storage area 252 (step 4).
05). If the number of blocks in the prefetch block number storage area 252 has reached 8, step 404 is skipped. That is, the number of prefetch blocks is gradually increased to prefetch up to a maximum of 8 blocks.

On the other hand, when the block number M-1 is not registered in any of the READ block number storage areas, it is determined whether the block number M is registered next (step 406). If registered, past R
Since the same block is being read by the EAD system call and the prefetching process is being executed at that time, prefetching is not performed this time (step 409). Also,
If it is not registered, the read request this time does not have continuity with the block read by the past READ system call, and therefore, prefetching is not performed in this case either. However,
In this case, the block number of the READ block number storage area that has not been updated the longest of the storage areas in the specific storage area 25 is rewritten to M (step 407), and the prefetch corresponding to the READ block number storage area of the storage area is performed. Register 0 in the block number storage area (step 40).
8).

The data read in advance from the secondary storage device 30 is held in the system buffer 24 managed by the OS 22, and copied from the system buffer 24 to the AP buffer 23 when the READ system call for the data is issued from the AP 21. To do. Since the number of system buffers is limited, the prefetched system buffer is temporarily released. Therefore, if the prefetched system buffer is not read by the READ system call for a long time, the system buffer is reused and the prefetched data disappears. That is, even in the case of READ of continuous blocks, it is not worth reading ahead if the time interval between blocks is open, and thus it is preferable to reuse the READ block storage area that has not been updated for the longest time. It is reasonable. From the above, the number N of storage areas of the storage area 25 is determined by the total number of system buffers. By prefetching, it is useless to reuse the system buffer that has been prefetched in the past and has not been accessed. Therefore, the number of storage areas is equal to the number of storage areas N = number of system buffers / maximum number of prefetch blocks. It is reasonable. A so-called LRU algorithm may be used to determine the storage area that has not been updated for the longest time.

FIG. 5 shows an example of pre-reading in the sequential access by applying the access pattern analysis algorithm according to the present invention. For the sake of convenience, FIG. 5 also shows the conventional case in comparison with the present invention. Here, the block number is the serial number from the beginning of the file on the secondary storage device, and the access number is the order of blocks requested by the AP for READ. FIG. 5 is an example of sequential access, and it is assumed that READ system calls are performed one block at a time from the beginning of the file. RE
The AD request block number is the number of the block that the AP READ requested and is 1, 2, 3, 4, ... The prefetch block number is the block number prefetched when the AP issues a READ request, and is all listed in comparison with the case of the conventional example and the embodiment of the present invention. Since the block numbers in <> have already been read ahead in the past READ request, they are not subject to read ahead. In the present invention, the pre-read block numbers are overlapped because the pre-reading is performed up to 8 blocks from the next block for which the READ request is issued from the AP. Conventionally, the storage area has only one READ block number storage area, and the storage area is indicated by M. On the other hand, in the present invention, the storage area includes the READ block storage area 251 and the prefetch block number storage area 252 described in FIG. 3, and there are N sets. M1 indicates one of N storage areas. The contents of the storage areas M and M1 are
This is the value after updating with the EAD system call.

As shown in FIG. 5, the next one block is always read ahead in the prior art, whereas in the present invention, the number of read-ahead blocks is increased by one block, and the maximum number is eight in the embodiment. I understand. That is, in sequential access, prefetching is performed both in the conventional method and in the present invention. However, the number of prefetch blocks is different.

FIG. 6 shows an example of pre-reading by applying the access pattern analysis algorithm according to the present invention in random access. FIG. 6 also shows the conventional case in comparison with the present invention. The meanings in FIG. 6 are the same as those in FIG. However, three storage areas M1 to M3 are used in the present invention. That is, a maximum of N sets of storage areas can be used, but with the random access pattern shown in FIG.

As shown in FIG. 6, in random access, the block numbers viewed in the order of access numbers are not continuous. However, if you look at only a part of the file, you can see that the accessed blocks are continuous. For example, blocks 1, 2, 3, 4, and 5 are accessed by READ access requests with access numbers 1, 4, 7, 10, and 13. According to the present invention, prefetching can be performed on the random access pattern in this manner, but conventionally, prefetching cannot be performed at all.

FIG. 7 is a diagram showing the control system in the system of FIG. In this embodiment, HD is used as the secondary storage device. As shown in FIG. 7, the entire system is AP,
The control that is configured by the OS and the software of the HD driver that is a part of the OS, and the control that is configured by the HD hardware of the secondary storage device can be classified into four layers in total. The most important feature of the present invention is the access pattern analysis of the OS layer, which is as described with reference to FIGS. 4 to 6.
The other part is a conventional configuration for the AP to read data from the secondary storage device.

AP is SEEK system call, READ
Issue a system call. The SEEK system call specifies the offset of the data to be accessed in the file, and the READ system call specifies the AP buffer that stores the file data and the data transfer size. O
S includes access pattern analysis, system buffer control mechanism and READ request issue. In the access pattern analysis, the flowchart of FIG. 4 is executed to analyze the access pattern of the AP file data and predict whether or not prefetching will be performed. Further, when prefetching, the amount of prefetch data is determined. The system buffer control mechanism is
It controls a system buffer located between the secondary storage device and the AP buffer. That is, the file data on the secondary storage device is once read into the system buffer and copied from the system buffer to the AP buffer. In issuing a READ request, a read request for a data block requested by the AP and a read-ahead data block is predicted to be issued to the HD driver when the access pattern analysis predicts that read-ahead will be performed. As described with reference to FIG. 2, one system buffer can hold one block of data. The size of one block is usually 8 KB. The HD driver is a part that controls the HD of the secondary storage device, and issues an I / O request for a data block for which the OS has made a READ request to the HD. The I / O request here is HD
Data transfer request (data transfer from HD to system buffer). The HD driver has a block merge function. The block merge function collects READ requests of a plurality of physically continuous blocks into one,
This is a function of issuing an I / O request to the HD. Also,
The HD driver has a system buffer queue management function. As described in FIG. 2, a plurality of system buffers are queued, and the HD driver I
Registration of I / O issuance, I / O processing completion, etc. is performed.

The HD is a device that stores file data and has a disk cache mechanism. After positioning the disk head, the HD temporarily puts the data on the disk cache and transfers the data from the disk cache to the system buffer. DMA for data transfer
Since (Direct Memory Access) is used, the CPU can execute other processing in parallel during data transfer. The disk cache is FIFO (First In First Out)
When data is transferred to the system buffer, the next continuous data is placed on the disk cache even if there is no next I / O request from the HD driver. However, some HDs do not have a disk cache mechanism.
The present invention is effective regardless of the presence or absence of the disk cache mechanism.

FIG. 8 shows R in the layer structure described in FIG.
6 shows a flow of file data of an EAD system call. File data flows in the order of disk track, disk cache, system buffer, AP buffer. FIG. 8 is an example in which the AP requests READ of one block to the AP buffer (denoted as 231) by the READ system call, and the OS prefetches four blocks as a result of the access pattern analysis.

The OS first sets the system buffer (241
Information of the block for which the READ request is issued from the AP (system buffer allocation), and issues a READ request to the HD driver. After the HD driver issues an I / O request for the system buffer (241) to the HD,
Immediately returns control to the OS. After positioning the disk head, the HD transfers the data of the disk track 301 to the disk cache 310 and DMAs the data from the disk cache 310 to the system buffer 241.
Forward. During that time, the OS reads the block information of the read ahead, here, from the system buffer 242 to the system buffer 24.
Set to 5 and issue a READ request to the HD driver.
The HD driver merges the blocks of the system buffer 245 from the system buffer 242, and when the HD notifies the completion of the I / O processing of the system buffer 241, the merged system buffer I / O requests are collectively sent to the HD. Issue. Then, the control returns to the OS again. OS
The data in the system buffer 241 to the AP buffer 23
1 and copy to AP buffer 231
The system call returns. Meanwhile, the I / O processing from the system buffer 242 to the system buffer 245 is executed in parallel by the HD layer.

Next, when the AP issues a READ system call to the AP buffer 232, for example, the READ request for the system buffer 242 has already been issued to the AP buffer 23.
The OS has been issued by the READ system call for 1. Therefore, the REA of the system buffer 242
Issuing the D request is omitted, and I / O of the system buffer 242 is omitted.
The READ system call returns only by waiting for the end of the O processing. Issuing a READ system call to AP buffer 231 and READ to AP buffer 232.
If system calls are issued in time, AP
When the READ system call is issued to the buffer 232, the I / O processing of the system buffer 242 has already been completed and the system buffer 242 is released. However, since the number of system buffers is sufficiently larger than the number of system buffers read ahead, the system buffers read ahead are not immediately reused for other purposes. If not reused, the OS can return the READ system call simply by transferring data from the system buffer to the AP buffer.

Next, the role of the disk cache 310 will be described. There is a slight time interval from the end of the I / O of the system buffer 241 to the issuance of the I / O of the system buffer 242, during which the disk slightly rotates. Therefore, if there is no discache 310, it is necessary to wait one rotation to read the data in the system buffer 242. When the disk cache 310 is present, the data in the system buffer 242 is also automatically transferred from the disk track 301 to the disk cache 310 following the data transfer in the system buffer 241, so that the system buffer can be transferred from the disk cache 310 without waiting for one revolution. Data can be transferred to 242.

The block merging function of the HD driver is most effective when there is no disk cache 310. That is, in the example of FIG. 8, the disk cache 310
If there is not, the system waits for one rotation, but without the block merge function of the HD driver, the system buffer 242 to the system buffer 245 will issue divided I / O processing requests, so wait a total of four rotations. Become.

9 and 10 show the entire flow chart of the READ system call. The READ system call process is a process in the extension of the READ system call (top half process) and an HD driver process in the extension of the interrupt notified at the end of the I / O process (bottom half process).
Divided into Here, the I / O processing means processing at the HD hardware level, and the I / O processing end means HD processing.
Means that the data transfer from to the system buffer is completed.

FIG. 9 is a flowchart of the top half processing. When the AP issues a READ request to the block number M by the READ system call, the OS searches the queue of the system buffer and determines whether the I / O processing of the block number M is already completed (step 901). The process is completed when the block number M has been prefetched in the past and the data has already been transferred to the system buffer. At this time, the prefetch process is immediately executed. If not completed, the system buffer queue is searched in the same manner, and READ of block number M is performed.
It is determined whether the request has already been issued from the OS to the HD driver (step 902). The block number M is issued when the block number M has been prefetched in the past. At this time, the prefetch process is immediately executed. The block number M is not issued when the block number M has not been prefetched in the past, the system buffer is allocated, and the block number M
Is issued from the OS to the HD driver (step 903). The HD driver determines whether the HD is busy (step 904). Busy means that the HD is in the process of I / O for another block. If it is not busy, the I / O request of the block number M is issued from the HD driver to the HD (step 906). If busy,
The block number M is registered in the I / O issue waiting queue (step 905). The processing of the I / O request registered in the queue is executed by the bottom half processing described in FIG.

Next, it is determined whether or not to prefetch (step 9).
07). The pre-reading determination algorithm is as described in FIG. When pre-reading, the number of pre-reading blocks is determined (step 908). Next, the system buffer queue is searched to determine whether the I / O processing of the prefetch block has already been completed (step 909). It is completed when it overlaps with the past read-ahead block and the data has already been transferred to the system buffer. The overlap of prefetch blocks is as described in FIGS. 5 and 6. If it is not completed, the system buffer queue is searched in the same manner to read the R of the look-ahead block.
It is determined whether the EAD request has already been issued from the OS to the HD driver (step 910). It is issued when it overlaps with the past look-ahead block.
If it has not been issued, the system buffer is allocated and the READ request for the prefetch block is issued from the OS to the HD driver (step 911). HD driver is H
It is determined whether D is busy (step 912). If it is not busy, an I / O request for the prefetch block is issued (step 914). If it is busy, the prefetch block is registered in the I / O issue waiting queue (step 91).
3).

Next, it is judged whether the I / O processing of the block number M has been completed (step 915). If it is not completed, the I / O processing of the block number M is waited (step 916). If it is completed, the data of the block number M is copied from the system buffer to the AP buffer (step 917). The call returns.

FIG. 10 is a flowchart of the bottom half processing. The bottom half processing is executed by the HD driver triggered by the I / O end interrupt. First, the HD driver notifies the top half of the end of I / O (step 1001). By this notification, the waiting state is released in step 916 of FIG. Next, the system buffer queue is searched to determine whether there is an entry in the I / O issue waiting queue (step 1002). The I / O issue waiting queue is entered in steps 905 and 913 of FIG. If there is an entry, the I / O request is HD
(Step 1003). At this time, if there are a plurality of consecutive block entries in the queue, the blocks are merged and an I / O request is issued.

FIGS. 11 to 14 verify that the wait for HD rotation and the number of seeks are reduced by prefetching according to the present invention.

FIG. 11 shows verification in the present invention in sequential access. The READ request block is A
The block requested by P in the READ system call and the prefetch block are the blocks prefetched at that time. I
I / O is issued by extension of the READ system call.
The number of merge blocks is the number of blocks merged when the I / O process is issued. The rotation wait shows whether there was a disk rotation wait when the data was read from the HD in the I / O processing, and shows whether there is a disk cache or not. The blank column means that there was no waiting for rotation.

In the case of READ request block 1, since it is the first access, only block number 1 is issued as an I / O processing request. At this time, there is a rotation wait regardless of the presence or absence of the disk cache. The seek is ignored during sequential access. In the case of the READ request block 2, after issuing the I / O processing request of the block number 2, the block number 3 is registered in the system buffer queue as an I / O processing request issue wait. At this time, if there is a disk cache, there is no wait for rotation, but if there is no disk cache, there is a wait for rotation. Hereinafter, when an I / O processing request is issued, there is no rotation wait if there is a disk cache, but there is a rotation wait if there is no disk cache. Immediately after the I / O processing of the block number 2 ends, the I / O processing request of the block number 3 registered in the system buffer queue is issued. READ
In the case of the request block 3, the block number 3 is already under I / O processing, and the prefetch blocks 4 and 5 are registered in the queue. When the I / O processing of the block number 3 is completed, the block numbers 4 and 5 are merged and the I / O processing request is issued. In the case of the READ request block 4, the block number 4 is already under I / O processing, and the prefetch blocks 6 and 7 are registered in the system buffer queue. Block number 4 I /
When the O processing is completed, the lock numbers 6 and 7 are merged and an I / O processing request is issued. In the case of the READ request block 5, since the block number 5 has been merged with the block number 4, the I / O processing has already been completed, and the prefetch blocks 8 and 9 are registered in the system buffer queue. At this time, the block numbers 6 and 7 are in the process of I / O, and the I / O process request is not issued by the extension of this system call.

After that, the same operation is repeated to increase the number of merge blocks, and the merge blocks are converged at 4 and 5 blocks. The number of I / O processing requests issued is reduced by the number of merge blocks, and the number of rotation waits is also reduced accordingly. At the time of the READ request block 21, the number of rotation waits is once when there is a disk cache and 9 times when there is no disk cache. The states of the READ request blocks 12 and 21 are all the same, and thereafter, the cycle of 12 to 21 is repeated, and the number of I / O processing request issuances is twice the 9 READ request blocks.

FIG. 12 shows the verification in the conventional prefetch in the sequential access. Conventionally, since the number of prefetch blocks is one, the I / O processing request for the prefetch block is always issued when the I / O processing of the READ request block ends. Therefore, the rotation wait is once if there is a disk cache, and RE if there is no disk cache.
The number of AD requested blocks has occurred.

In the sequential access, comparing the present invention with the conventional one, when the disk cache is present, the number of rotation waiting times is one, and when the disk cache is not present, the present invention is 9 at the time of the READ request block 21. Times, 21 times in the past. Looking at the READ request block 13 and thereafter, the number is 2 in the present invention and 9 in the conventional case.

FIG. 13 shows verification in the present invention in random access. This is the data of block numbers 1 to 11 and 50 to 59, alternately REA one block at a time.
This is an example of D request. That is, FIG. 6 shows an example in which data is distributed at three points, whereas FIG. 13 shows an example in which data is distributed at two points. The look-ahead algorithm is the same as in FIG.

In random access, a seek occurs when an I / O processing request for blocks that are not continuous is issued. RE
In the case of the AD request block 2, I of block number 2 is first
The / O processing request is issued, and block number 3 of the prefetch block is registered in the system buffer queue. When the I / O processing of block number 2 ends, the I / O of block number 3
O processing request is issued. In the case of the READ request block 51, the block number 51 is registered in the system buffer queue because the I / O processing of the block number 3 has not been completed yet. Then, the block number 52 of the prefetch block is also registered in the system buffer queue. When the I / O processing of the block number 3 is completed, the block numbers 51 and 52 are merged and an I / O processing request is issued.

When the I / O processing of the block numbers 51 and 52 is completed, the processing shifts to the READ request block 3.
Therefore, when the processing of the READ request block 3 is started,
O There is no block being processed or a block entered in the queue. Since the I / O processing of the block number 3 has already been completed, the I / O processing request of the block number 4 of the prefetch block is immediately issued. And block number 5
Is registered in the system buffer queue. Since the block number 4 is the prefetch block, the processing shifts to the processing of the READ request block 52 without waiting for the end of the I / O processing. Since the I / O processing of the block number 52 has already been completed, the block numbers 53 and 54 are registered in the system buffer queue. Therefore, in the processing of the READ request block 52, I
No / O processing request is issued. READ request block 4
Since the block number 4 is still in the process of I / O at the start of the processing (1), the prefetch blocks 6 and 7 are registered in the system buffer queue. When the I / O processing of the block number 4 is completed, the prefetch blocks 5, 6, and 7 are merged, and the I / O processing request is issued. At this time, block number 4 I /
Seek does not occur due to O processing and continuous blocks.
If there is a disk cache, rotation wait does not occur.

Similarly, in the example of FIG. 13, I / O
The number of processing request issuance is 11 times, seek is 9 times, rotation wait is 9 times if there is a disk cache, 11 if not
Times.

FIG. 14 shows the verification of the conventional look-ahead in random access. With the conventional read-ahead algorithm, random access patterns cannot be read ahead at all, so seek and rotation wait occur for each block as shown in FIG. Therefore, in this example, 21 seeks and 21 rotation waits occur regardless of the presence or absence of the disk cache.

In the random access of FIGS. 13 and 14, comparing the present invention with the prior art, the number of seeks is 9 in the present invention, 21 in the conventional, and waiting for rotation is 9 in the present invention when there is a disk cache. If the conventional method is 21 times and the disk cache is not provided, the present invention is 11 times and the conventional method is 21 times.

[0048]

As is apparent from the above description, the data prefetching method of the present invention has the following effects.

(1) Since prefetching can be performed for an access pattern which has been conventionally regarded as random access and which has not been prefetched, it becomes possible to execute I / O processing and AP processing in parallel. Throughput can be improved.

(2) Even in sequential access, batch I / O processing of a plurality of blocks can be realized by pre-reading up to, for example, a maximum of 8 blocks that were previously read in advance for only 1 block. as a result,
For example, if the secondary storage device is a hard disk (HD), the number of times I / O processing is issued from the HD driver to the HD is reduced, the number of disk seeks and rotation waits is reduced to less than half, and file access is accelerated. You can

(3) By changing the number of pre-read blocks, the number of pre-read blocks can be reduced once with respect to the total number of blocks read by the AP, and the pre-read block is not accessed by the AP. It is possible to reduce the damage caused by hitting.

(4) By realizing the read-ahead function at the OS level without changing the AP interface of the READ system call, the AP can obtain the read-ahead effect without changing the conventional file access method. .

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram of an entire system according to an embodiment of the present invention.

FIG. 2 is a conceptual diagram of a system buffer in FIG.

FIG. 3 is a diagram showing a configuration example of a specific storage area in FIG.

FIG. 4 is a flowchart of a data prefetch process according to the present invention.

FIG. 5 is an example in which the present invention and the conventional read-ahead are executed for sequential access.

FIG. 6 is an example in which the present invention and the conventional read-ahead are performed for random access.

FIG. 7 is a diagram illustrating an example of a layer structure of control for realizing prefetching according to the present invention.

8 is a diagram showing a data flow of a READ system call in FIG.

FIG. 9 is an overall flowchart of a top half process of a READ system call.

FIG. 10 is an overall flowchart of a bottom half process of a READ system call.

FIG. 11 is an example of verifying the number of times of waiting for rotation of a hard disk in sequential access according to the present invention.

FIG. 12 is an example of verifying the number of times of waiting for rotation of a hard disk in conventional sequential access by prefetching.

FIG. 13 is an example of verifying the number of seeks and rotation waiting times of a hard disk in random access according to the present invention.

FIG. 14 is an example of verifying the number of seeks and rotation waiting times of a hard disk in conventional random access by prefetching.

[Explanation of symbols]

 10 Central Processing Unit (CPU) 20 Main Storage Device 21 Application Program (AP) 22 Operating System (OS) 220 Access Pattern Analysis Unit 23 AP Buffer 24 System Buffer 30 Secondary Storage Device 31 Disk Cache Mechanism

Claims (3)

[Claims] 1. A method for prefetching data stored in the secondary storage device into the main storage device in a computer system having a central processing unit, a main storage device, and a secondary storage device, the secondary storage device comprising: By analyzing the data access pattern for accessing the data stored in the storage device continuously / discontinuously, predicting the data to be accessed in the future, and prefetching the data from the secondary storage device to the main storage device, Characteristic data prefetching method. 2. The data pre-reading method according to claim 1, wherein the amount of data to be pre-readed is dynamically changed according to a prediction result of data to be accessed in the future. 3. The data pre-reading method according to claim 1, wherein a plurality of sets of an address that has accessed data in the secondary storage device in the past and the number of pre-reading data are held in a specific storage area. A data pre-reading method, which comprises predicting data to be accessed in the future and calculating the amount of data to be read in advance based on the contents of the specific storage area when an access request is made to the next storage device.