CN1302392C - Online method for reorganizing magnetic disk - Google Patents

Abstract

The present invention discloses an online method for reorganizing magnetic disks, which comprises the following steps that logic units (LUN) the depth size of strips of which is similar in an independent redundant disk array (RAID) can be in advance divided into at least one group; when the reorganizing process of magnetic disks begins, a block of buffer can be applied as a special buffer according to the LUN the strip depth of which is maximal in RAID, and the reorganization can be carried out according to the group; when the reorganization is carried out to any group, a block of buffer can be applied as a free buffer according to the LUN the strip depth of which is maximal in the group, and the application of an advance buffer is the same with the application of the free buffer. The proposal of the present invention can ensure that the release frequencies of internal memory application of a manager of a buffer area are less, and simultaneously, the waste of the internal memory is reduced.

Description

A Disk Online Reconfiguration Method

technical field

The invention relates to the field of an independent redundant disk array (RAID) system, in particular to an online reconfiguration method of a disk.

Background technique

With the continuous enhancement of computer data processing capabilities, the reading speed of hard disks can no longer meet the needs of large amounts of data reading, so independent redundant disk array technology has emerged, which has low cost, low power consumption, high transmission rate, and simple implementation. And other advantages, it is widely used in network servers and other equipment.

One of the advantages of RAID technology is the ability to perform online disk reconstruction when a disk is damaged. The current mainstream disk reconstruction methods are as follows:

Construct N processes, N-1 processes correspond to N-1 disks that are still working normally, and one process corresponds to the hot spare disk.

A process attached to a functioning disk proceeds as follows:

Step 1. Find the stripe unit with the lowest address on the disk;

Step 2. If the buffer (buffer) has enough space to receive data, send a low-priority read request to the stripe unit and read it into the buffer (skip the reconstructed stripe);

Step 3. Wait for the read operation to complete. If there is a buffer in the buffer manager that can accept the data, send the data to the buffer of the buffer manager for XOR, and return to step 1, otherwise wait , until all the stripe units are read.

Processes connected to the hot spare disk are processed as follows:

Step 4. If there is a buffer in the buffer manager that has completed the XOR of all the striping units, take it out and go to the next step, otherwise wait;

Step 5. Send a low-priority write request to the hot spare disk, write the data in the fetched buffer to the hot spare disk, return to step 4, and wait for the writing to complete until all data on the failed disk is reconstructed.

The current mainstream Raid technology includes the division of logical units (LUN, Logical Unit Number). The striping depth of the LUN varies from one another.

As can be seen from the above process, in the disk-oriented reconstruction algorithm, due to the random arrival of users on different disks, some processes may read more data blocks than other processes, and the cache management process must manage this information until the slowest in the system The process submits the data up. In this way, each process's demand for cache changes at any time, and there is no fixed value, but the system's cache is limited, so the cache demand of each process must be limited. However, in the reconstruction process, the address of the stripe unit is carried out from low to high, so it is difficult to determine the size of the application buffer in the buffer manager. However, if the size of the buffer is not fixed, it will increase the application release process of the cache and affect the system performance. In this regard, the current common practice is to make the size of the buffer in the buffer manager depend on the largest stripe unit. Although this unifies the size of the buffer, it may cause a waste of memory. Especially when the stripe depth differs greatly, for example, according to different applications, the stripe depth of a logical unit may differ by thousands of times, then if the cache size of the unified application is based on the maximum stripe depth, it will occupy the buffer manager in the buffer manager for a long time. Large block cache, resulting in a great waste of memory resources.

Contents of the invention

In view of this, the object of the present invention is to provide a cache management method based on online reconfiguration of disk logical units. Through this method, the waste of memory can be reduced while ensuring fewer buffer manager memory application release times.

In order to achieve the above object, the technical solution of the present invention is specifically realized in the following way:

A cache management method in online reconfiguration of a disk logical unit, comprising the following steps:

a) Divide logical units with similar stripe depths in the independent redundant disk array into at least one or more groups in advance;

b) When the disk reconstruction process starts, apply for a cache according to the LUN with the largest stripe depth in the RAID as a dedicated cache, and perform reconstruction according to the group. Apply for a cache as a free cache, and apply for a reserved cache according to the LUN with the largest stripe depth in the group.

Preferably, the sum of the stripe depths of all LUNs in each group divided by step a) of the method is similar.

The method further includes: setting a threshold, if the required cache ratio of the LUN with the largest stripe depth and the LUN with the smallest stripe depth in the RAID is less than the threshold, then make the number of groups in step a) be 1, otherwise, Let the number of groups in step a) be between 3 and 5.

The threshold value of this method is 2 or 3.

It can be seen from the above scheme that for the shortcomings of the traditional disk-oriented reconstruction algorithm, the present invention increases the distinction of LUNs, so that the buffer size of the buffer manager can be adjusted according to the change of the LUN group, thereby achieving low memory waste and Less application and release process of memory.

Description of drawings

FIG. 1 is a schematic diagram of an implementation of an embodiment of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Generally, RAID reconstruction cache includes three parts: dedicated cache, reserved cache, and free cache. The private cache is dedicated to reading disk data during the reconstruction process; the free cache is used to save the data to be XORed; the reserved cache is part of the free cache and is mainly used when forced reconstruction occurs, that is, the segmentation of user write requests In the reconstruction state, if the private cache cannot submit the free cache for the XOR operation, the private cache will submit the reserved cache to complete the XOR operation.

For the purpose of the invention described above, the present invention adopts the following strategy: first divide the LUNs with closer stripe depths into one group, and then perform reconstruction according to the group. The stripe depth applies for free cache and reserved cache, and releases them after the group reconstruction is completed, while the dedicated cache for disk data pre-reading is applied for according to the LUN with the largest stripe depth in the entire RAID. After the application , the private cache is not freed until the refactoring is complete. In this way, by grouping LUNs, it not only reduces the occupation of cache resources, but also ensures that the number of cache application releases is as small as possible.

During actual operation, it is also necessary to consider that if the stripe depths of the LUNs in the RAID are not much different, there is no need to group them.

Therefore, a threshold M is given first. And assume that N=MAX_Memery_Requied/MIN_Memery_Requied. Here Memery_Requied is the cache size required for each LUN reconstruction; MAX_Memery_Requied is the maximum value of the LUN in the RAID; MIN_Memery_Requied is the minimum value.

There are two situations:

1. N is smaller than M. At this time, because the gap between the stripe depths is small, you can apply for a cache pool instead of grouping them. The dedicated cache, free cache, and reserved cache can all be allocated according to the LUN in the RAID. Unified application for maximum slitting depth. Of course, it can also be considered that the entire RAID is divided into a group. This not only reduces the difficulty of memory management, but also reduces the frequency of memory release. For example, there are LUN1-LUN5, and the maximum stripe unit depth is 45 sectors, then the required reconstruction cache will not exceed 1MB. At this time, LUN1-LUN5 can completely apply for a cache pool for unified use.

2. N is greater than M, and the difference in the depth of the strips is large at this time. You need to perform disk reconstruction in groups according to the above method.

The threshold M is generally between 2 and 3. For its specific value, M can be reasonably set according to the application occasion of the RAID group and in combination with the system configuration. For example, for multimedia applications, since the size of a stripe unit is relatively large, generally between 512K and 4M, M can be set slightly larger, such as 3.

Referring to FIG. 1, RAID5 is taken as an example to illustrate the LUN-based online reconstruction method.

In general, the stripe depth difference between LUNs is not very uniform. As shown in Figure 1, you can first determine an allowable difference δ, compare the stripe depths of all LUNs in the RAID group, gather the LUNs satisfying |LNi-LUNj|≤δ, and apply for cache respectively according to the grouping situation. for its reconstruction.

For example, if δ=50 sectors (Sectors), LUN1-LUN5 are divided into three groups, group 1: LUN1, LUN3; group 2: LUN2, LUN4; group 3: LUN5. During the reconstruction process, the private cache size is constant at 800 Sectors; the free cache or reserved cache is taken as 250 Sectors, 350 Sectors, and 800 Sectors in sequence as the reconstruction proceeds. When reconfiguring, follow the order of LUN1 and LUN3 in the first group, LUN2 and LUN4 in the second group, and LUN5 in the third group. After the reconstruction is complete, a group of LUNs is re-applied for free cache.

The specific steps are:

Step 1: Apply for a private cache with a size of 800 Sectors.

Step 2: Read the data in LUN1 to the dedicated cache, apply for a free cache size of 250 sectors, submit the data of LUN1 to the free cache, perform an XOR operation, and send the data to the hot spare disk after completion. At this time, keep the free cache, and continue to reconstruct LUN3 according to the above steps, that is, read the data in LUN3 to the dedicated cache, submit the data of LUN3 to the free cache, and wait for the data of all the stripe units of LUN3 to be completely different or later, send the data to the hot spare disk, thus completing the reconstruction of group 1.

In this way, after the reconstruction of group 1 is completed, release the free cache; re-apply for the free cache with a size of 350 Sectors, and reconstruct LUN2 and LUN4 of group 2 once according to the above steps. After completion, release the free cache; apply again for free The size of the cache is 800 sectors, and LUN5 in group 3 is reconstructed.

When forced reconstruction occurs, if the data cannot be submitted from the private cache to the free cache for XOR operation, then the reserved cache is submitted to complete the XOR operation. The steps and cache application method are the same as those of the free cache.

Here, δ is better to make the number of groups between 3 and 5. If it is too large or too small, it will not be able to achieve the purpose of grouping and categorization. And when grouping, it is best to make the sum of all LUN sizes in each group similar, so that the size of each group to be reconstructed is similar. If the access frequency of each group of the host is similar, the reconstruction time of each group is similar. , so that the reconstruction time of a certain group is not too long, and the reconstruction time of a certain group is too short. When a LUN is reconfigured, the access efficiency of the host to the LUN will always be affected. Reducing the reconstruction time of the group where a LUN is located is good for improving the access efficiency of the host to the LUN.

If the stripe depths of the LUNs in the RAID system happen to be relatively uniform, such as the same RAID system as shown in Figure 1, and the stripe depths of the LUNs from LUN1 to LUN5 happen to be: 100, 200, 300, 400, and 500 sectors, At this time, it is difficult to determine an appropriate δ size. Therefore, according to the principle that the sum of the LUN sizes in each group is similar, divide LUN1 to LUN3 into a group, LUN4 into a group, and finally LUN5 into a group. The refactoring process is the same as above.

In addition, for another situation, for example, the RAID structure in FIG. 1 is still taken as an example. If LUN1 to LUN5 are 100, 210, 300, 410, and 500 in turn, although they can be grouped by setting a difference value δ, the principle of similar sum of LUN sizes in each group cannot be satisfied. At this time, LUN1 can also be grouped Group LUN3 into a group, LUN4 into a group, and finally LUN5 into a group.

In short, the principle of grouping is to keep the number of groups not too large, usually between 3 and 5, so as to ensure a small number of cache application releases, and to make the sum of the LUN sizes in each group similar. In this way, a better grouping scheme can be determined by comprehensively considering the above two principles according to the actual situation. Of course, the specific implementation is not limited to the above grouping method. That is to say, other grouping methods that ensure less cache application release times and achieve less waste of memory can also achieve the purpose of the present invention.

Claims (4)

1, a kind of disk on-line reorganization method is characterized in that, may further comprise the steps:

A) in advance the close logical block LUN of segment depth size among the raid-array RAID is divided at least more than one group;

B) the disk restructuring procedure begins, LUN by segment depth maximum among the RAID applies for that a block cache is as proprietary buffer memory, be reconstructed by group, when reconstruct proceeds to a certain group, LUN by segment depth maximum in this group applies for a block cache as free buffer memory, and reserves the application of buffer memory according to the LUN of segment depth maximum in this group.

2, method according to claim 1 is characterized in that, the segment depth size sum of all LUN is close in each group that step a) is divided.

3, method according to claim 1, it is characterized in that, further comprise: set a threshold value, if the ratio of the required buffer memory of LUN of the LUN of segment depth maximum and segment depth minimum is less than this threshold value among the RAID, then making the grouping number in the step a) is 1, otherwise making the grouping number in the step a) is between 3 to 5.

4, method according to claim 3 is characterized in that, described threshold value is 2 or 3.

Images