RU2805012C2 - Method of data storage system management and data storage system - Google Patents

Abstract

FIELD: computer engineering.

SUBSTANCE: invention is proposed to increase the performance of the data storage system and increase its service life. It is achieved due to the fact that the data storage system includes RAM, a parallel SSD cache, divided into a first group of write blocks (RWC) and a second group of read blocks (RRC), a hard drive, a request detector and a frequency analyser requests to read data. The method allows you to divide requests into random and sequential both during write and read operations and, depending on the type of request, send them to different types of memory. When all free groups in the SSD cache are filled in the first group of write blocks or the second group of read blocks, data is evicted. Data that has not been used for a long time is evicted from the first group of write blocks (RWC). From the second group of read blocks (RRC), data from the data block that has the lowest usage intensity is released, taking into account the number of data evictions experienced from the block over a certain period of time.

EFFECT: increased performance and increased service life of the data storage system.

11 cl, 8 dwg

Description

TECHNICAL FIELD

The invention relates to a data storage system and management of this system, including operational memory, a parallel SSD cache and a hard drive.

BACKGROUND ART

The main advantages of storage systems with SSD cache and HDD are the high I/O performance provided by solid-state drives (SSDs) and the long-lasting capability provided by disk drives (HDDs). Moreover, such systems require special caching algorithms that take into account the characteristics of the cache device used - SSD.

Known patent US 9128847, publication 09/08/2015, IPC G06F-003/04, which describes a device for managing the cache memory of a system that includes a controller, an SSD cache and a hard drive, with the SSD cache divided into two memory blocks. The cache memory management method provides that when a certain threshold is exceeded, data from the first SSD cache block is rewritten either to the second SSD cache block or to the hard drive.

Known patent KR 101675694, publication November 23, 2016, IPC G06F-003/06. The patent describes a method for replacing an SSD block depending on the popularity of the block. The disk system contains SSD as cache and memory on HDD hard drives. The method involves determining the least popular SSD block based on the data residence time and the number of requests per block and removing it from the SSD queue. The invention prevents SSD cache contamination and replacement of an SSD block having a high hit rate. Additionally, SSD write operations are reduced by reducing the block exchange rate, thereby extending the life of the SSD.

The closest analogue is application US20180032266, publication 02/01/2018, IPC G06F-003/06, which describes a method for managing a storage system. The storage system contains an SSD cache and a hard disk drive, while the SSD cache is divided into memory blocks of the same size and from these blocks a first group of blocks for writing and a second group of blocks for reading are formed. If the I/O request is a read request, storing the target data from the storage device is performed in the first group of cache memory devices; if the I/O request is a write request, storing the target data in a second group of cache devices.

Using SSDs as cache devices with a limited number of write cycles can significantly accelerate their wear due to the use of traditional caching algorithms. Thus, hybrid storage systems using SSD and HDD memory require new caching algorithms to make optimal use of the hybrid storage structure.

SUMMARY OF THE INVENTION

The technical result of the invention is to increase the performance of the data storage system and increase its service life through optimal use of the SSD cache.

A method for managing a storage system that includes RAM, a parallel SSD cache, and a hard disk drive contains the operations presented below.

The SSD cache is divided into memory blocks of the same size and the first group of write blocks (RWC) and the second group of read blocks (RRC) are formed from these blocks.

Receive a request to write or read data.

- when requesting to record data, determine the type of request - sequential or random. Random request data is written to free memory blocks of the first group of blocks for writing the SSD cache, and sequential requests are written to the hard drive;

- when requesting to read data, determine the type of request - sequential or random. A sequential read request locates the data and reads it. In case of a random read request, the location of the data is determined and read, and if the data was stored in the hard drive, then they are analyzed for the frequency of requests to read data and, if the frequency of requests is more than two, the data from the hard drive is rewritten to the second group blocks for reading the SSD cache.

When all free groups of the first group of blocks for writing or the second group of blocks for reading are filled, the SSD cache performs data eviction,

- in this case, data is sequentially displaced into the hard drive from the first group of write blocks (RWC),

- and from the second group of read-readable blocks (RRC), the data that has the lowest write intensity is ousted, taking into account the number of data evictions experienced from the block over a certain period of time.

In particular, the SSD cache is pre-divided into memory blocks of K megabytes in size.

In addition, data is written to the SSD cache memory blocks in the form of sequential recording.

In addition, when a write request is made, it is checked whether the request data is stored in the SSD cache and, if available, the data is updated.

At the same time, random requests of more than M kilobytes are written to the hard drive.

In addition, when receiving a read request, they first check where the request is located and if there are parts of the request simultaneously in RAM, SSD cache, or HDD, they split the request into parts and consider each part as a separate request.

In addition, during a sequential read request, data is read from either RAM or the hard drive.

In particular, data is released from the second group of read blocks (RRC) of the SSD cache when there are no free memory blocks other than the current one.

In addition, free memory blocks of the first or second group of SSD cache memory blocks are transferred to another group of SSD cache memory blocks.

In addition, data is evicted from the SSD cache in entire blocks.

The storage system includes a write and read request detector connected to RAM connected to a parallel SSD cache and hard drive storage. At the same time, the SSD cache is divided into memory blocks of the same size and includes the first group of write blocks (RWC) and the second group of read blocks (RRC). The SSD cache is also connected to the hard drive. At the same time, the system contains an analyzer of the frequency of requests to read data, connected at the input to the detector of requests for writing and reading data, and at the outputs to the RAM and SSD cache.

The following terms are used in this application.

Block - Logical space of the SSD cache, divided into blocks of equal size K megabytes.

The cache hit count is the number of requests for specific data that resides in a data block in the SSD cache.

BRIEF DESCRIPTION OF THE DRAWINGS

In FIG. Figure 1 shows a block diagram of the data storage system.

In FIG. Figure 2 shows the structure of the SSD cache, containing the first group of write blocks (RWC) and the second group of read blocks (RRC).

In FIG. Figure 3 shows the structure of the request.

In FIG. Figure 4 shows the structure of the first group of write blocks (RWC) and the second group of read blocks (RRC).

In FIG. Figure 5 shows a flow diagram of the read request processing process.

In FIG. Figure 6 shows a flowchart of the write request processing process.

In FIG. Figure 7 shows a flow diagram of the process of eviction from the second group of read blocks (RRC).

In FIG. 8 shows a flow diagram of the process of eviction from the first group of write blocks (RWC).

OPTIONS FOR IMPLEMENTING THE INVENTION

SSD caching is a technology that uses solid state drives as a buffer for frequently accessed data. The system determines how often data is accessed and moves it to a fast storage device. Storage with SSD cache complements HDD arrays to improve performance by optimizing random queries. The design of HDD drives allows them to successfully cope with sequential requests, but has a natural limitation for working with random requests. The size of the SDD cache is usually about 5-10% of the capacity of the main disk subsystem.

The claimed data storage system and its management method based on RAIDIX technology implements an SSD cache parallel to RAM, dividing incoming requests into categories RRC (Random Read Cache) - a read request, and RWC (Random Write Cache) - a write request, and the use of a new method of eviction from a group of read blocks (RRC). In addition, increasing the performance of the storage system and increasing its service life due to the optimal use of the SSD cache can be further increased by redistributing the ratio of the number of free and used RRC and RWC blocks depending on the current load on the storage system.

The data storage system (Fig. 1) includes RAM 2, a parallel SSD cache 4 based on solid state drives, and hard disk drives (HDD) 3. In a system with parallel SSD cache 4, RAM 2 is used separately from SSD cache 4.

At the system input there is a request detector 1, which determines the type of request - random or sequential.

The SSD cache space 4 is divided into blocks of equal size K megabytes. In terms of SSD cache efficiency, the smaller the block size, the better, but the more metadata needs to be stored in RAM. The most optimal block size is 1 GB. Data is written to the SSD cache memory blocks in the form of a sequential (log-structured) record.

The data storage system (Fig. 1) also contains an analyzer 5 of the frequency of requests to read data and the size of an incoming write request, connected at the input to the detector 1 of requests for writing and reading data, and at the outputs to RAM 2 and SSD cache 4 With the help of analyzer 5, decisions are made about the need to write new requests to the SSD cache, while only random requests of size less than M get into the RWC. The RRC gets requests that were repeatedly accessed at a certain point in time, that is, with a frequency of more than two.

The RAM 2 input is connected to detector 1 of requests for writing and reading data.

The SSD cache 4 (Fig. 2) is divided into two functional categories: blocks of the first write block group 7 (RWC) and the second read block group 6 (RRC). Each of these groups of blocks has its own hit and displacement rules. In FIG. Figure 3 shows that request 8 consists of two parts. The first part of the request is occupied by metadata - information about the data block itself and the state of the data that is recorded in it. The rest is the data itself. For the first group of write blocks (RWC), they are records that store information in the form of a hash table about where on the hard disk drive (HDD) this cached data is located.

In FIG. 4 shows a view of the blocks of the first group of blocks 6 for writing (RWC) and the second group of blocks 7 for reading (RRC).

- An occupied block that is completely full.

- The current block in which requests are currently written.

- A free block that contains no data.

- The one being displaced, from which the displacement procedure is currently taking place.

Free main blocks 7 of the second group for reading RRC can, if necessary, become free blocks of the first group for writing RWC. If necessary, free blocks 6 of the first group of RWC blocks can become free blocks of the second group for reading RRC. This depends on the selected policy - with or without read and write priority. If there is no priority, free blocks are not moved between RRC and RWC. With read or write priority, free RRC/RWC blocks become RWC/RRC blocks if needed. If priority is set for both read and write operations, then free blocks are not initially owned by the RWC/RRC, and free blocks are assigned to the first requests to arrive.

A flowchart of operations when processing a read request is shown in FIG. 5. When requesting to read 101 data, determine the type of request 102 - sequential or random. In a sequential read request, the location of the data 103 is determined and read from either the SSD 104 or RAM 105 or the HDD 106.

With a random read request, the location of the data is determined. They check if there is this random request in RAM 107 and read from the RAM 108. They check if there is a random request in SSD 109 and read it from SSD 110, or from HDD 112. If the data was read from the SSD, the block metric changes in RRC , which was hit.

If the data was stored on the HDD, then they are analyzed 113 for the frequency of requests to read data using analyzer 5 (Fig. 1) and, if the frequency of requests is more than two, the data from the hard drive 3 is rewritten 114 to the second group of blocks 7 for reading SSD -cache. If there are no free blocks 115 in the second group of sides 7 for reading, operations are performed to release blocks 116. In this case, the block with the smallest WE metric is released and the metrics of all occupied blocks are recalculated. Processing of the read request ends 117.

It should also be noted that the read request may refer to data that is located in different parts of the database, for example, in RAM, SSD cache, or HDD. In this case, before all operations, the primary request is divided into parts, depending on which memory they relate to, and each part is considered a separate request.

A flowchart of operations when processing a write request is shown in FIG. 6. When requesting 201 to record data, determine the type of request 202 - sequential or random. If the request is sequential, write request 203 to HDD 3. Next, check 204 whether this request is already in SSD cache 4. If it is, update 205 parts of the request in SSD cache 4. Next, check whether 206 the size of the request exceeds recording parameter M. If the request exceeds this size M, which is set, for example, to 32 KB - 64 KB, then the request is written 211 to the HDD. If the request is small, then 207 is checked, if there is free space in the current block of the first group of RWC blocks in SSD cache 4, then 208 is written to the current block. If there is no space, then they check 209 whether there is a free block, and if there is a 210 free block, write to it, otherwise write to HDD 3.

The process of eviction from the second group of read blocks 7 (RRC) is shown in FIG. 7. The process of eviction from the second group of blocks begins when there are no free blocks in the group except the current one.

Initially, the two blocks with the fewest cache hits per block are determined. The intensity of use of each of the selected blocks for a certain period of time is determined. The intensity of use is determined based on the well-known metric “Write Efficiency”, proposed in [1]

Where:

The number of cache hits is the number of requests for data located in a specific, selected data block in the SSD cache.

Number of entries in the cache block - the number of entries in the data block in the SSD cache at the time the WE metric is determined.

All blocks are ordered by the WE metric and block 301 with the lowest value is selected. It is now considered free block 302.

After this, the WE metric of value 303 is recalculated for all occupied blocks. To do this, the WE metric for all occupied blocks is divided by the number N - this is the query aging parameter. In the implemented algorithm, it is equal to 2. This ends the eviction process 304.

The process of eviction from the first group of write blocks (RWC) is disclosed in FIG. 8. From the first group of write blocks (RWC), long-unused data is ousted in accordance with the 401 eviction policy. In this case, eviction algorithms based on the “queue” principle, for example, FIFO (“first in, first out”) can be used ( https://ru.wikipedia.org/wiki/FIFO). When eviction is performed, the requests in the block are ordered 402 by increasing addressing number, then the ordered requests are written to the HDD - 403. After this, the 404 freed block receives the status of free.

The set of all operations (dividing the parallel SSD cache into memory blocks of the same size and forming from these blocks the first group of write blocks (RWC) and the second group of read blocks (RRC); processing of write and read requests and data eviction processes, different for groups of SSD cache blocks for reading and writing) method of managing the data storage system (including RAM, parallel SSD cache and hard drive) can significantly increase the performance of the storage system and extend the life of SSD drives.

Comparative experiments with known designs of hybrid data storage systems, which include RAM, parallel SSD cache and hard drives, showed that the performance increase for different types of queries ranges from 5 to 46 times. Moreover, the number of required rewrites on SSD drives in the proposed method will be approximately 6 times less than in many traditional data storage systems. Accordingly, in this system the resource of solid-state drives is used much more efficiently, increasing their lifespan.

INDUSTRIAL APPLICABILITY

The claimed method and device can be used to effectively work with a fairly wide range of tasks and loads in various areas of data storage.

Thanks to query sorting and new data eviction methods, random load peaks are smoothed out faster and with less impact on overall system performance.

[1] WEC: Improving Durability of SSD Cache Drives by Caching Write Efficient Data / Yunpeng Chai, Zhihui Du, Xiao Qln, David A Bader // IEEE Transactions on Computers. - 2015. - Vol. 64, no. 11. - P. 3304-3316.

Claims (20)

1. A method of managing a data storage system that includes RAM, parallel SSD cache and hard drive storage, in which: - divide the SSD cache into memory blocks of the same size and form from these blocks the first group of write blocks (RWC) and the second group of read blocks (RRC); - receive a request to write or read data; - when requesting to write data, the type of request is determined - sequential or random, and the data of random requests is written to free memory blocks of the first group of blocks for writing the SSD cache, and sequential requests are written to the hard drive; - when requesting to read data, determine the type of request, sequential or random; - with a sequential read request, they determine the location of the data and read it; - in the event of a random read request, the location of the data is determined and read, and if the data was stored in the hard drive, then they are analyzed for the frequency of requests to read data, and if the frequency of requests is more than two, the data from the hard drive is rewritten to the second group blocks for reading SSD cache; - when all free groups of the first group of blocks for writing or the second group of blocks for reading of the SSD cache are filled, data is evicted, - in this case, data is sequentially displaced onto the hard disk drive from the first group of write blocks (RWC), - and from the second group of read-readable blocks (RRC), the data that has the lowest write intensity is ousted, taking into account the number of data evictions experienced from the block over a certain period of time. 2. The method according to claim 1, characterized in that the SSD cache is first divided into memory blocks of K megabytes in size. 3. The method according to claim 1, characterized in that the data is written in the form of sequential recording into the memory blocks of the SSD cache. 4. The method according to claim 1, characterized in that when a write request is made, it is checked whether the request data is stored in the SSD cache, and if so, the data is updated. 5. The method according to claim 1, characterized in that random requests of more than M kilobytes are written to the hard drive. 6. The method according to claim 1, characterized in that upon receiving a read request, they first check where the request data is located and if parts of the request are simultaneously in RAM, SSD cache or HDD, the request is divided into parts and each part is considered a separate request. 7. The method according to claim 1, characterized in that during a sequential read request, data is read either from RAM or from a hard drive. 8. The method according to claim 1, characterized in that data is evicted from the second group of read-readable blocks (RRC) of the SSD cache when there are no free memory blocks other than the current one. 9. The method according to claim 1, characterized in that free memory blocks of the first or second group of SSD cache memory blocks are transferred to another group of SSD cache memory blocks. 10. The method according to claim 1, characterized in that data is evicted from the SSD cache in entire blocks. 11. A data storage system including a data write and read request detector connected to RAM connected to a parallel SSD cache and a hard disk drive, wherein the SSD cache is divided into memory blocks of the same size and includes a first group of blocks for writing (RWC) and a second group of blocks for reading (RRC), the SSD cache is also connected to the hard drive, and the system contains an analyzer of the frequency of requests for reading data, connected at the input to a detector of requests for writing and reading data, and via outputs to RAM and SSD cache.