WO2024187900A1 - Data storage method, system and device for distributed storage system, and storage medium - Google Patents

Abstract

The present application is applied to the technical field of storage. Disclosed are a data storage method, system and device for a distributed storage system, and a storage medium. The distributed storage system comprises a plurality of hard disk clusters, wherein one hard disk cluster comprises a plurality of hard disks which have the same model number and are all added in the same batch to the distributed storage system. The method comprises: receiving data to be written, and determining the data type thereof; when the data type of said data is an ith data type, determining whether there is a hard disk cluster in an ith state currently; if there is a hard disk cluster in the ith state, selecting one hard disk cluster in the ith state; and writing said data into the selected hard disk cluster in the ith state, wherein the lower an estimated value of the number of times that said data is modified within a first duration, the higher the wear degree of the hard disk cluster into which said data is written. By applying the solution of the present application, the global wear balance of a distributed storage system is realized, the durability of a hard disk is ensured, and the situation in which a bad disk is frequently replaced is avoided.

Description

Data storage method, system, device and storage medium of distributed storage system

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims priority to a Chinese patent application filed with the China Patent Office on March 15, 2023, with application number 202310247233.0 and application name “Data storage method, system, device and storage medium for distributed storage system”, the entire contents of which are incorporated by reference in this application.

Technical Field

The present application relates to the field of storage technology, and in particular to a data storage method, system, device and storage medium of a distributed storage system.

Background Art

Flash media has gone through four generations of development. Currently, TLC (Triple Level Cell) is widely used, and QLC (Quad Level Cell) is about to enter mass production. Compared with the previous generation, each generation has higher data density, cheaper price, lower durability, lower performance, and improved energy efficiency, which are all determined by the physical properties of flash memory. However, QLC has two main disadvantages: low performance and poor durability. Regarding the poor durability, that is, the P/E (Program/Erase cycle) is decreasing rapidly, the current method is to do wear leveling in an SSD (Solid-State Drives) disk. The current P/E of QLC has dropped to around 1000, but the effect of the above approach is limited in distributed systems. There are thousands of QLC SSD disks in the distributed system, and the amount of data written to each disk is uneven. In addition, due to the reasons of new and old disks, some disks break down quickly, while some disks can be used for a long time. The low-cost characteristics of QLC cannot be reflected, and the frequent replacement of bad disks also increases the operation and maintenance costs.

Summary of the invention

The purpose of the present application is to provide a data storage method, system, device and storage medium of a distributed storage system, so as to effectively improve the durability of hard disks and avoid frequent replacement of bad disks.

In order to solve the above technical problems, this application provides the following technical solutions:

A data storage method for a distributed storage system, wherein the distributed storage system includes multiple hard disk clusters, each hard disk cluster includes multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch. The data storage method for the distributed storage system includes:

Receive data to be written and classify the data types of the data to be written; wherein, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the ith data type within the first time length, and N is a positive integer not less than 2;

When the data type of the data to be written is the i-th data type, determining whether there is a hard disk cluster in the i-th state;

If there is currently a hard disk cluster in the i-th state, select a hard disk cluster in the i-th state;

The data to be written is written into the selected hard disk cluster in the i-th state; wherein, among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1th state is lower than the wear degree of the hard disk cluster in the i-th state, and i is a positive integer.

Preferably, the data types of the data to be written are divided into:

Based on the file name of the data to be written, the data type of the data to be written is divided.

Preferably, based on the file name of the data to be written, the data type of the data to be written is divided, including:

When the file name of the data to be written matches the preset j-th database, the data type of the data to be written is divided into the j-th data type; wherein j is a positive integer and 1≤j≤N-1;

When the file name of the data to be written does not match any of the preset N-1 databases, the data type of the data to be written is divided into the Nth data type.

Preferably, it also includes:

The file name is used as a training sample, and the modification count of the training sample in the first time period is used as a training label of the training sample to train the preset deep learning model;

After the deep learning model is trained, different file names are input into the trained deep learning model in sequence, and data of N-1 databases are updated based on the output results of the deep learning model.

Preferably, it also includes:

An adjustment instruction for the jth database is received, and a data item adding operation, a data item deleting operation, and/or a data item modifying operation is performed on the jth database according to the adjustment instruction.

Preferably, N=3, the first data type is a read-only data type, the second data type is a cold data type, the third data type is a hot data type, the first state of the hard disk cluster is a G _readonly state, the second state of the hard disk cluster is a G _cold state, and the third state of the hard disk cluster is a G _hot state.

Preferably, it also includes:

When the data to be written is of the first data type and it is determined that there is no hard disk cluster in the first state currently, the data to be written is written into a hard disk cluster in the G _hot state or the G _cold state.

Preferably, it also includes:

When the data to be written is of the second data type and it is determined that there is no hard disk cluster in the second state currently, the data to be written is written into the hard disk cluster in the G _hot state.

Preferably, it also includes:

When the data to be written is of the third data type and it is determined that there is no disk cluster in the third state currently, a prompt message indicating write failure is fed back.

Preferably, selecting a hard disk cluster in the i-th state includes:

According to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, a hard disk cluster in the i-th state is selected.

Preferably, selecting a hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, includes:

For each hard disk cluster in the i-th state, search in order from the smallest to the largest wear degree;

When the current cluster write queue depth VG _{cur_queue_depth} of any disk cluster in the i-th state is found to be less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the search is stopped and the disk cluster is used as the selected disk cluster in the i-th state;

After searching all disk clusters in the i-th state, if there is no disk cluster whose current cluster write queue depth VG _{cur_queue_depth} is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the disk cluster with the smallest cluster busyness VGbusy is selected as the disk cluster in the i-th state;

The cluster busyness VGbusy of the hard disk cluster represents the value obtained by dividing the current cluster write queue depth VG _{cur_queue_depth} of the hard disk cluster by the maximum cluster write queue depth VG _{max_queue_depth} .

Preferably, writing the data to be written into the selected hard disk cluster in the i-th state includes:

According to the rule that the lower the wear degree of the hard disk, the higher the priority, the target hard disk is selected from the selected hard disk cluster in the i-th state;

Write the data to be written into the selected target hard disk.

Preferably, selecting a target hard disk from the selected hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk, the higher the priority, includes:

Search each hard disk in the selected hard disk cluster in the i-th state in order of the wear degree from small to large;

When it is found that the current hard disk write queue depth VD _{cur_queue_depth} of any hard disk in the hard disk cluster in the i-th state is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the search is stopped and the hard disk is selected as the target hard disk;

After searching all hard disks in the hard disk cluster in the i-th state, if there is no hard disk whose current hard disk write queue depth VD _{cur_queue_depth} is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the hard disk with the smallest hard disk busyness VDbusy is selected as the target hard disk;

The hard disk busyness VDbusy of the hard disk represents a value obtained by dividing the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk by the maximum hard disk write queue depth VD _{max_queue_depth} .

Preferably, each hard disk cluster in the distributed storage system is arranged in the first medium layer, and an SCM medium layer is also arranged in the distributed storage system to store target type data through the SCM medium layer and process non-block aligned write data through the SCM medium layer.

Preferably, the first dielectric layer is a PLC dielectric layer or a QLC dielectric layer.

Preferably, in each hard disk cluster, data is stored in blocks of a set size, and the data storage method of the distributed storage system further includes:

Determine the P/E times of each block in the hard disk cluster in the i-th state, and calculate the average value of the P/E times of each block in the hard disk cluster;

When there are blocks in the disk cluster in the i-th state whose difference between the P/E times and the average value is lower than the set first value, all blocks whose difference between the P/E times and the average value is lower than the set first value are taken as blocks to be migrated;

If there is currently a hard disk cluster in the (i+1)th state, each block to be migrated in the hard disk cluster is migrated to the hard disk cluster in the (i+1)th state.

Preferably, it also includes:

After determining each block to be migrated, if there is no hard disk cluster in the i+1th state, the K blocks with the largest P/E times in the hard disk cluster in the i-th state are exchanged with the data in the K blocks to be migrated in the hard disk cluster in the i-th state, so as to complete the internal migration of the K blocks to be migrated in the hard disk cluster in the i-th state;

Here, K represents the number of blocks to be migrated determined in the hard disk cluster in the i-th state.

A data storage system of a distributed storage system, the distributed storage system includes multiple hard disk clusters, the hard disk clusters include multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch, the data storage system of the distributed storage system includes:

A type classification module, used for receiving the data to be written and classifying the data types of the data to be written; wherein, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the ith data type within the first time length, and N is a positive integer not less than 2;

The hard disk cluster status judgment module is used to judge whether there is a hard disk cluster in the i-th state when the data type of the data to be written is the i-th data type; if there is a hard disk cluster in the i-th state, the hard disk cluster selection module is triggered;

A hard disk cluster selection module is used to select a hard disk cluster in the i-th state;

The writing module is used to write the data to be written into the selected hard disk cluster in the i-th state; wherein, in the set Among the hard disk clusters in N states, the wear degree of the hard disk cluster in the i+1th state is lower than the wear degree of the hard disk cluster in the ith state, where i is a positive integer.

A data storage device of a distributed storage system, comprising:

Memory for storing computer programs;

The processor is used to execute a computer program to implement the steps of the data storage method of the distributed storage system as described above.

A computer non-volatile readable storage medium stores a computer program, which implements the steps of the data storage method of the above-mentioned distributed storage system when executed by a processor.

Applying the technical solution provided by the embodiment of the present application, considering that in order to fully realize the high cost-effectiveness of flash memory media, it is not limited to wear leveling within one hard disk, but it is necessary to perform global wear leveling based on a distributed storage system. Specifically, in the solution of the present application, the distributed storage system is divided into multiple hard disk clusters. Compared with directly managing each hard disk, the management data required for the hard disk cluster is lower, that is, the amount of metadata, which is easy to implement. In addition, for any one hard disk cluster, the hard disk cluster includes multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch. Through such a setting, it is conducive to conveniently realizing the wear leveling of each hard disk in the hard disk cluster. Then, by realizing the wear leveling between the hard disk clusters, the global wear leveling of the entire distributed storage system can be realized, which also ensures the durability of the hard disk and can avoid the situation of frequent replacement of bad disks.

Specifically, when performing wear leveling between hard disk clusters, the present application divides the hard disk cluster into N states. After receiving the data to be written, the data type of the data to be written will be divided. When the data to be written is divided into the i-th data type, it is determined whether there is a hard disk cluster in the i-th state at present. If so, a hard disk cluster in the i-th state will be selected. Since different data types reflect the different frequencies of future modification of the data to be written, and specifically, among the N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the i-th data type within the first time length. And among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1th state is lower than the wear degree of the hard disk cluster in the i-th state. It can be seen that for data that almost does not need to be modified, that is, when the estimated number of modifications of the data to be written within the first time length is very low, the data to be written will be divided into the first data type, and therefore will be written into the hard disk cluster in the first state. The wear degree of the hard disk cluster in the first state is the highest, which means that the hard disk cluster in the first state has written a large amount of data, so the data written is data that almost does not need to be modified. Writing to the hard disk will wear out the hard disk, but reading will not. It can be seen that, since the data written to the hard disk cluster with a high degree of wear is data that hardly needs to be modified, even if the hard disk cluster has a high degree of wear, it can still be read, thus giving full play to its residual value. Correspondingly, the more frequently the data to be written needs to be modified, that is, the higher the estimated number of modifications to the data to be written within the first time period, the data to be written will be written to the hard disk cluster with a lower degree of wear, so that the hard disk cluster with a lower degree of wear can be used more fully, thus realizing the global wear leveling of the distributed storage system.

In summary, the present application divides the distributed storage system into multiple hard disk clusters, which is conducive to conveniently realizing global wear leveling of the distributed storage system, thereby ensuring the durability of the hard disks in the distributed storage system and avoiding the frequent replacement of bad disks.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

FIG1 is a flowchart of an implementation method of a distributed storage system in the present application;

FIG2 is a schematic diagram showing a principle framework for implementing global wear leveling in a specific implementation of the present application;

FIG3 is a schematic diagram of a multi-layer flash memory architecture of a distributed storage system in a specific implementation manner of the present application;

FIG4 is a schematic diagram of the structure of a data storage system of a distributed storage system in the present application;

FIG5 is a schematic diagram of the structure of a data storage device of a distributed storage system in the present application;

FIG6 is a schematic diagram of the structure of a computer non-volatile readable storage medium in the present application.

DETAILED DESCRIPTION

The core of this application is to provide a data storage method for a distributed storage system, which divides the distributed storage system into multiple hard disk clusters, which is conducive to conveniently realizing global wear leveling of the distributed storage system, thereby ensuring the durability of the hard disks in the distributed storage system and avoiding the frequent replacement of bad disks.

In order to enable those skilled in the art to better understand the present application, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

Please refer to FIG. 1, which is a flowchart of an implementation of a data storage method of a distributed storage system in the present application. The distributed storage system includes multiple hard disk clusters, and the hard disk clusters include multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch. The data storage method of the distributed storage system may include the following steps:

Step S101: receiving data to be written, and classifying the data types of the data to be written;

Among them, among the set N data types, the estimated number of modifications to the data to be written of the i+1th data type within the first time period is higher than the estimated number of modifications to the data to be written of the i-th data type within the first time period, and N is a positive integer not less than 2.

Specifically, in the solution of the present application, the concept of a hard disk cluster is defined. A distributed storage system may include multiple hard disk clusters, and one hard disk cluster may include multiple hard disks of the same model.

It should be noted that for different hard disk clusters, the number and models of hard disks included can be different, that is, the disks can be from different manufacturers, of different sizes, and added to the cluster at different times. They can be set according to actual needs and will not affect the implementation of this application. The hard disks in the same hard disk cluster have the same model, that is, the same hard disk cluster is composed of a group of hard disks with the same capacity and the same properties.

All hard disks in the same hard disk cluster are added to the distributed storage system in the same batch, and since the present application realizes global wear balancing of the distributed storage system, the wear degree of each hard disk in the same hard disk cluster is similar, and if replacement is required, they are replaced at the same time.

The specific types of hard disks described in this application can be set and adjusted as needed, and are usually SSD hard disks. Of course, in other specific implementations, they can be other types of hard disks. However, it should be pointed out that since mechanical hard disks do not have the problem of friction loss, there is basically no limit on the number of writes. The wear leveling implemented in this application is for solid-state hard disks with a limit on the number of writes. Therefore, the hard disks in the solutions of this application are usually solid-state hard disks with a limit on the number of writes.

After receiving the data to be written, the data type of the data to be written needs to be divided. Please refer to Figure 2, which is a principle framework diagram for realizing global wear leveling. The medium user in Figure 2 represents the division of the data type of the data to be written, and then transmits it to the flash allocation module in Figure 2 through a tag, so that the flash allocation module knows the data type of the data to be written according to the tag, and then decides which hard disk cluster the data to be written should be written to.

In the solution of the present application, in order to achieve global wear leveling, the more frequently the data needs to be modified, the more it should be written to the hard disk cluster with low wear. In other words, the data types of the data to be written in the present application are divided according to the frequency of future modification of the data to be written. That is, different data types reflect the different frequency of modification of the data to be written. Therefore, the data types of the data to be written are divided according to the different frequency of future modification of the data to be written.

Among the N data types that are set, the estimated number of modifications to the data to be written of the i+1th data type within the first time period is higher than the estimated number of modifications to the data to be written of the i-th data type within the first time period, and N is a positive integer not less than 2, that is, at least 2 data types need to be set, and N represents the total number of data types.

That is to say, when the data to be written is of the i-th data type, it means that the estimated number of modifications of the data to be written within the first time length is in the i-th estimated value interval, and the N estimated value intervals are sorted from small to large in order from 1 to N.

It can be seen that when the data to be written is of the first data type, the estimated number of modifications of the data to be written within the first time length is the smallest, indicating that the data to be written hardly needs to be modified. When the data to be written is of the Nth data type, the estimated number of modifications of the data to be written within the first time length is the largest, indicating that the data to be written needs to be modified frequently.

In order to divide the data types of the data to be written according to the modification frequency of the data to be written, there may be multiple specific implementations. For example, in a specific implementation of the present application, the data types of the data to be written described in step S101 may specifically include:

Based on the file name of the data to be written, the data type of the data to be written is divided.

As described above, the present application divides the data types of the data to be written according to the frequency of future modification of the data to be written. This implementation method takes into account whether the data to be written needs to be modified frequently, which can be reflected to a certain extent in the file name. Therefore, the frequency of modification of the data to be written can be predicted based on the file name of the data to be written, that is, the data type of the data to be written can be divided, which will be more convenient in implementation.

Further, in a specific implementation of the present application, based on the file name of the data to be written, the data type of the data to be written is divided, which may specifically include:

When the file name of the data to be written matches the preset j-th database, the data type of the data to be written is divided into the j-th data type; wherein j is a positive integer and 1≤j≤N-1;

When the file name of the data to be written does not match any of the preset N-1 databases, the data type of the data to be written is divided into the Nth data type.

N in this application is a positive integer not less than 2, which can represent the total number of data types. At the same time, N in this application is also the total number of states of the hard disk cluster, that is, the number of data types set is the number of states of the hard disk cluster set. By judging whether the file name of the data to be written matches the corresponding database, it is very simple to implement, that is, this implementation method can easily and quickly predict the specific type of the data to be written.

This implementation takes into account that N-1 databases can be pre-set, so that when the file name of the data to be written matches a certain database, the data type of the data to be written can be determined to be the data type corresponding to the database, that is, if the file name of the data to be written matches the j-th database among the N-1 databases, the data type of the data to be written is classified as the j-th data type.

In addition, it should be pointed out that the reason why N-1 databases are set up instead of N databases in this implementation is that no matter whether N-1 databases or N databases are set up, there may still be a situation where the file name of the data to be written does not match any database, that is, the set database is difficult to comprehensively cover the file names of various data to be written in actual applications.

If no matching database is found, it means that the modification frequency of the data to be written cannot be predicted. Therefore, in order to avoid excessive wear on the hard disk cluster with high wear, in this implementation, the data to be written in this situation is divided into the Nth data type, that is, the subsequent data to be written will be written to the hard disk cluster in the Nth state. The hard disk cluster in the Nth state is the hard disk cluster with the lowest wear among the hard disk clusters in various states.

Since in this implementation mode, it is necessary to classify the data type of the data to be written whose file name does not match any database as the Nth data type, there is no need to set up N databases in this implementation mode, but only N-1 databases. That is, as long as the file name of the data to be written does not match the preset N-1 databases, no matter whether the frequency of modification of the data to be written in the future cannot be predicted or whether it is indeed necessary to modify it very frequently in the future, it is directly classified as the Nth data type.

In addition, the specific content in the N-1 databases can be set by the staff based on experience, and also supports dynamic adjustment of the database content to better meet actual needs. For example, in a specific implementation of the present application, it can also include:

An adjustment instruction for the jth database is received, and a data item adding operation, a data item deleting operation, and/or a data item modifying operation is performed on the jth database according to the adjustment instruction.

The adjustment instruction may be issued by the staff. When the adjustment instruction for the j-th database is received, the data items of the j-th database may be added, and/or deleted, and/or modified according to the adjustment instruction.

It should also be noted that whether the file name of the data to be written matches the preset j-th database, the specific matching rules can also be set according to actual needs. For example, in one scenario, one or more file name suffixes are set in the j-th database. As long as the suffix of the file name of the data to be written is consistent with any file name suffix in the j-th database, the file name of the data to be written is considered to match the preset j-th database.

For example, in a specific case, N=3, then two databases need to be set up, which are called the first database and the second database. For example, when the file name of the data to be written has a suffix of avi, bmp, etc., it can be determined that the file name of the data to be written matches the first database, and the data to be written is divided into the first data type, and the data to be written needs to be written to the hard disk cluster in the first state. For example, when the file name of the data to be written has a suffix of bak, log, etc., it can be determined that the file name of the data to be written matches the second database, and the data to be written is divided into the second data type, and the data to be written needs to be written to the hard disk cluster in the second state.

Of course, matching based on the file name suffix is only an example of a relatively simple implementation method. In other specific implementations, other more complex matching methods can also be set. For example, the header of the file name can be analyzed, and the analysis result of the file name header and the file name suffix can be combined to determine whether the file name of the data to be written matches the corresponding database. For another example, the overall pattern of the file name can be analyzed to determine whether the file name of the data to be written matches the corresponding database.

In a specific implementation of the present application, it may also include:

The file name is used as a training sample, and the modification count of the training sample in the first time period is used as a training label of the training sample to train the preset deep learning model;

After the deep learning model is trained, different file names are input into the trained deep learning model in sequence, and data of N-1 databases are updated based on the output results of the deep learning model.

As described above, the specific contents in the N-1 databases all support dynamic adjustment. This implementation method takes into account that if the specific contents in the N-1 databases are set and adjusted based on the experience of the staff, the workload will be large, and the adjustment effect will also be affected by the staff's business level.

In this regard, in this implementation, data updates of N-1 databases can be implemented based on the deep learning model. A deep learning model can be established and trained. During the training process, the training samples are the file names, and the training labels are the statistical values of the number of modifications of the training samples in the first time period. After the deep learning model is trained, the file name is input to the trained deep learning model, and the deep learning model can output the prediction result corresponding to the file name. The prediction result represents the estimation of the number of modifications of the data with the file name in the first time period in the future.

After inputting each different file name into the trained deep learning model in turn, the data of N-1 databases can be updated based on the output results of the deep learning model, that is, different file names are placed in the corresponding databases according to the output results of the deep learning model.

Step S102: when the data type of the data to be written is the i-th data type, determine whether there is a hard disk cluster in the i-th state; if there is a hard disk cluster in the i-th state, execute step S103.

Step S103: Select a hard disk cluster in the i-th state.

After the data types of the data to be written are classified, if the data to be written is of the i-th data type, as long as there is a hard disk cluster in the i-th state, the data to be written can be subsequently written into the hard disk cluster in the i-th state.

As described above, in the scheme of the present application, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time period is higher than the estimated number of modifications of the data to be written of the i-th data type within the first time period, that is, the data to be written of the 1st data type hardly needs to be modified, while the data to be written of the Nth data type needs to be modified most frequently.

Among the N states of hard disk clusters, the wear degree of the hard disk cluster in the i+1th state is lower than the wear degree of the hard disk cluster in the ith state, where i is a positive integer. That is, the hard disk cluster in the ith state reflects that the wear degree of the hard disk cluster is in the ith wear degree interval. Therefore, there are N wear degree intervals in total, and the N wear degree intervals are sorted from large to small in the order from 1 to N.

It can be seen that the more frequently the data to be written needs to be modified, that is, the higher the estimated number of modifications of the data to be written within the first period of time will be, the harder it will be to select a hard disk cluster with a lower degree of wear to store the data to be written.

The wear degree of the hard disk cluster is the wear degree of the hard disk cluster. The higher the wear degree of the hard disk cluster, the lower the number of subsequent write operations that can be performed on the hard disk cluster, that is, the lower the life of the hard disk cluster. In a specific case, the wear degree of the hard disk cluster can be represented by how much data has been written in the hard disk cluster. That is, in a specific case, the wear degree Vwl of the hard disk cluster can be defined as: the sum of the amount of written data of the hard disk cluster/the total amount of writable data of all hard disks in the hard disk cluster. In other words, the more data has been written to the hard disk cluster, the higher the wear degree Vwl of the hard disk cluster, and the highest is 100%.

In addition, as described above, N is the total number of data types set, and is also the total number of states of the hard disk cluster. In practical applications, considering that the present application divides the data types of the data to be written, it is equivalent to predicting the frequency of modification of the data to be written in the future. It can be understood that after the data to be written is written into the distributed storage system, the actual frequency of modification may be different in different occasions and time periods. Therefore, in practical applications, it is not necessary to divide the data types of the data to be written too finely, that is, the total number of states of the hard disk cluster does not need to be divided too much, and too many total states of the hard disk cluster are not easy to manage. Therefore, in a specific implementation, N can be set to 3. At this time, the first data type is specifically a read-only data type, the second data type is specifically a cold data type, and the third data type is specifically a hot data type. The first state of the hard disk cluster is the G _readonly state, the second state of the hard disk cluster is the G _cold state, and the third state of the hard disk cluster is the G _hot state.

It can be seen that in this implementation, for data that will hardly be modified in the future after being written, the data type is 1 data type, that is, read-only data type, will be written to the hard disk cluster in the G _readonly state. For example, backup files belong to this type of data.

The hard disk cluster in the G _readonly state has a high degree of wear, but it can still be read. The read operation will not affect the life of the hard disk. Writing data that will hardly be modified in the future into the hard disk cluster in the G _readonly state is conducive to giving full play to the value of this type of hard disk cluster.

Correspondingly, in this example, data that will be modified after being written but will not be modified too frequently has a data type of the second data type, ie, a cold data type, and will be written to a hard disk cluster in the G _cold state.

Data that will be modified frequently after being written is of the third data type, that is, the hot data type, and will be written to the hard disk cluster in the G _hot state.

In the following text of this application, the implementation method of N=3 is also used as an example for explanation. Since N=3, when the wear degree of the hard disk cluster Vwl＜V1, the state of the hard disk cluster is G _hot state; when V1≤Vwl＜V2, the state of the hard disk cluster is G _cold state; when Vwl≥V2, the state of the hard disk cluster is G _readonly state. V1 and V2 described here are two preset parameter thresholds, which can be set and adjusted by the staff as needed. It can be understood that the set V1＜V2.

It should also be noted that when the data to be written is of the i-th data type, there may be multiple hard disk clusters in the current i-th state. In this case, it is necessary to select one hard disk cluster in the i-th state according to the set rules to write the data to be written into the selected hard disk cluster in the i-th state. The specific selection rules can be set and adjusted according to actual needs.

For example, in a specific implementation of the present application, considering that there are multiple hard disk clusters in the i-th state, although the wear degrees of these hard disk clusters belong to the same wear degree range, there are still differences in the wear degrees of each hard disk cluster in the i-th state. In this regard, in order to more effectively achieve global wear balancing of the distributed storage system in this implementation, considering that the lower the wear degree of the hard disk cluster, the higher the priority, a hard disk cluster in the i-th state can be selected. In other words, for each hard disk cluster in the i-th state, they can be sorted from low to high according to the wear degree, and the hard disk cluster with low wear degree can be selected first, so as to more effectively achieve global wear balancing of the distributed storage system.

In the implementation of FIG2 , there are three G _hot state hard disk clusters, which are marked as G1, G2 and G3, six SSDs are set in G1, which are marked as d1 to d6, four SSDs are set in G2, which are marked as d1 to d4, and seven SSDs are set in G3, which are marked as d1 to d7. In FIG2 , the three G _hot state hard disk clusters are arranged in order from small to large according to the wear degree Vwl, so as to give priority to the hard disk cluster with low wear degree.

In addition, it is understandable that, since the wear degree of the hard disk cluster may change, the Vwl sorting of each hard disk cluster in the same state may also be dynamically updated in real time or periodically.

Further, in a specific implementation of the present application, selecting a hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, may specifically include:

For each hard disk cluster in the i-th state, search in order from the smallest to the largest wear degree;

When the current cluster write queue depth VG _{cur_queue_depth} of any disk cluster in the i-th state is found to be less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the search is stopped and the disk cluster is used as the selected disk cluster in the i-th state;

After searching all disk clusters in the i-th state, if there is no disk cluster whose current cluster write queue depth VG _{cur_queue_depth} is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the disk cluster with the smallest cluster busyness VGbusy is selected as the disk cluster in the i-th state;

The cluster busyness VGbusy of the hard disk cluster represents the value obtained by dividing the current cluster write queue depth VG _{cur_queue_depth} of the hard disk cluster by the maximum cluster write queue depth VG _{max_queue_depth} .

In this implementation, not only the wear of the hard disk cluster is taken into consideration, but also the current cluster write queue depth of the hard disk cluster is taken into consideration, which is conducive to ensuring high concurrency, that is, it is conducive to achieving high IOPS (Input/Output Operations Per Second) and high bandwidth of the distributed storage system.

Specifically, for each disk cluster in the i-th state, searches are performed in order of wear degree from small to large. The purpose of the search is to determine whether the current cluster write queue depth VG _{cur_queue_depth} of the disk cluster is less than the preset maximum cluster write queue depth VG _{max_queue_depth} .

Taking Figure 2 as an example, among the three hard disk clusters in the G _hot state, the first search is for the hard disk cluster G1 with the lowest wear. If the current cluster write queue depth VG _{cur_queue_depth} of G1 is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , it means that G1 is not busy at present, so the search can be stopped and G1 is selected as the i-th state hard disk cluster according to the preset rules, so that the data to be written can be written to G1 later.

However, if G1 is currently very busy and the current cluster write queue depth VG _{cur_queue_depth} is equal to or even exceeds the maximum cluster write queue depth VG _{max_queue_depth} , even if G1's wear is lower than G2 and G3, if G1 is selected to write the data to be written, G1's busyness will increase, causing queue congestion, which is not conducive to achieving high IOPS and high bandwidth of the distributed storage system.

Therefore, in this implementation, G2 will continue to be searched. If the current cluster write queue depth VG _{cur_queue_depth} of G2 is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , it means that G2 is not busy at present, so the search can be stopped and G2 will be selected as the hard disk cluster in the i-th state according to the preset rules, so that the data to be written can be written to G2 later.

If G1, G2, and G3 are all searched and are all busy, that is, there is no disk cluster whose current cluster write queue depth VG _{cur_queue_depth} is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , then the disk cluster with the smallest cluster busyness VGbusy is selected as the disk cluster in the i-th state according to the preset rule.

The cluster busyness VGbusy of the hard disk cluster represents the value obtained by dividing the current cluster write queue depth VG _{cur_queue_depth} of the hard disk cluster by the maximum cluster write queue depth VG _{max_queue_depth} , that is, VGbusy=VG _{cur_queue_depth} /VG _{max_queue_depth} .

The maximum cluster write queue depth VG _{max_queue_depth} is hardware-related and indicates the maximum number of write requests that a disk cluster can process simultaneously, which depends on the computing resources and storage resources of the disk cluster.

Step S104: writing the data to be written into the selected hard disk cluster in the i-th state.

After a hard disk cluster for storing the data to be written is selected, the data to be written can be written into the selected hard disk cluster.

In a specific implementation of the present application, step S104 may specifically include:

According to the rule that the lower the wear degree of the hard disk, the higher the priority, the target hard disk is selected from the selected hard disk cluster in the i-th state;

Write the data to be written into the selected target hard disk.

This implementation method takes into account that after determining the hard disk cluster for storing the data to be written, the specific hard disk for storing the data to be written is selected from the hard disk cluster. Although the wear degree of each hard disk in the same hard disk cluster is generally the same, there are still some differences. In order to more effectively ensure the wear balance of each hard disk in the hard disk cluster, in this implementation method, the hard disk with low wear degree is preferentially selected as the selected target hard disk. In other words, the hard disks in the selected hard disk cluster can be sorted from low to high according to the wear degree, and the hard disk with low wear degree is preferentially selected to more effectively achieve global wear balance of the distributed storage system.

In addition, it is understandable that, since the wear degree of the hard disk may change, the wear degree ranking of each hard disk in the hard disk cluster may also be dynamically updated in real time or periodically.

Further, in a specific implementation of the present application, selecting a target hard disk from the selected hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk, the higher the priority, includes:

Search each hard disk in the selected hard disk cluster in the i-th state in order of the wear degree from small to large;

When it is found that the current hard disk write queue depth VD _{cur_queue_depth} of any hard disk in the hard disk cluster in the i-th state is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the search is stopped and the hard disk is selected as the target hard disk;

After searching all hard disks in the hard disk cluster in the i-th state, if there is no hard disk whose current hard disk write queue depth VD _{cur_queue_depth} is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the hard disk with the smallest hard disk busyness VDbusy is selected as the target hard disk;

The hard disk busyness VDbusy of the hard disk represents a value obtained by dividing the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk by the maximum hard disk write queue depth VD _{max_queue_depth} .

In this implementation, the principle of selecting one hard disk cluster from each hard disk cluster in the i-th state is the same as in the previous implementation, that is, for each hard disk in the hard disk cluster, when selecting the target hard disk, not only the wear of the hard disk is considered, but also the current hard disk write queue depth of the hard disk is considered, which is conducive to ensuring high concurrency, that is, it is conducive to achieving high IOPS and high bandwidth of the distributed storage system.

Specifically, after selecting a hard disk cluster in the i-th state, each hard disk in the selected hard disk cluster in the i-th state will be searched in order from small to large wear degree. The purpose of the search is to determine whether the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} .

The first hard disk to be searched is the one with the lowest wear. If the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , it means that the hard disk is not busy at present, so the search can be stopped and the hard disk can be used as the target hard disk so that the data to be written can be written to the hard disk later.

However, if the hard disk with the lowest wear is currently very busy, and the current hard disk write queue depth VD _{cur_queue_depth} is equal to or even exceeds the maximum hard disk write queue depth VD _{max_queue_depth} , if this hard disk is selected to write the data to be written, it will increase the busyness of the hard disk and cause queue congestion, which is not conducive to achieving high IOPS and high bandwidth of the distributed storage system.

Therefore, in this implementation mode, other hard disks will continue to be searched at this time. The principle is the same as above and will not be repeated.

Similarly, if all hard disks in the hard disk cluster are searched and are all busy, that is, there is no hard disk whose current hard disk write queue depth VD _{cur_queue_depth} is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , then the hard disk with the smallest hard disk busyness VDbusy is used as the target hard disk.

The hard disk busyness VDbusy represents the value obtained by dividing the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk by the maximum hard disk write queue depth VD _{max_queue_depth} , that is, VDbusy=VD _{cur_queue_depth} /VD _{max_queue_depth} .

The maximum disk write queue depth VD _{max_queue_depth} is hardware-related and indicates the maximum number of write requests that a hard disk can process simultaneously, which depends on the computing resources and storage resources of the hard disk.

It is understandable that in the present application, the state of the hard disk cluster will change over time. Still taking N=3 as an example, at the beginning, the state of a hard disk cluster newly added to the distributed storage system is G _hot . As the data wears out, the wear degree Vwl of the hard disk cluster will become higher and higher. When it is greater than or equal to V1, its state is changed to G _cold . When the wear level Vwl of the hard disk cluster is greater than or equal to V2, its state is changed to G _readonly state. In practical applications, when there are too many hard disk clusters in the G _readonly state, the system can alarm and prompt to update the SSD disk.

In a specific implementation of the present application, it may also include:

When the data to be written is of the second data type and it is determined that there is no hard disk cluster in the second state currently, the data to be written is written into the hard disk cluster in the G _hot state.

As described above, the state of the hard disk cluster will change over time. When the distributed storage system is just established, since no data has been written to each hard disk cluster, the wear degree is very low, that is, the state of each hard disk cluster is in the G _hot state. At this time, if the data to be written is classified as the second data type, since there is currently no hard disk cluster in the second state, that is, there is no hard disk cluster in the G _cold state, the data to be written can be written to the hard disk cluster in the G _hot state.

In a specific implementation of the present application, it may also include:

When the data to be written is of the first data type and it is determined that there is no hard disk cluster in the first state currently, the data to be written is written into a hard disk cluster in the G _hot state or the G _cold state.

This implementation takes into account that when the data to be written is classified as the first data type, if there is currently no hard disk cluster in the first state, that is, there is no hard disk cluster in the G _readonly state, the data to be written can be written to a hard disk cluster in the G _hot state or the G _cold state.

In addition, it should be noted that there is currently no hard disk cluster in the first state. It may be that the wear of each hard disk cluster is relatively low. For example, in the example of N=3, it may be that the wear of each hard disk cluster is relatively low, and there is only a hard disk cluster in the G _hot state or the G _cold state. It may also be that there is a hard disk with a very high wear degree, but the hard disk is completely full and no new data can be written. In this case, it can also be regarded as that there is currently no hard disk cluster in the first state, and the data to be written needs to be written to the hard disk cluster in the G _hot state or the G _cold state.

In a specific implementation of the present application, it may also include:

When the data to be written is of the third data type and it is determined that there is no disk cluster in the third state currently, a prompt message indicating write failure is fed back.

This implementation takes into account that when the data to be written is of the third data type, if there is currently no hard disk cluster in the third state, that is, there is no hard disk cluster in the G _hot state, it means that the storage space of the distributed storage system has been used in large quantities and there is insufficient remaining space. Therefore, in order to avoid data loss, this implementation will directly feedback a prompt message of write failure.

In addition, in actual applications, an alarm can be sent to the system to remind staff to add additional resources to the distributed storage system in a timely manner.

In a specific embodiment of the present application, each hard disk cluster in the distributed storage system is arranged in the first medium layer, and an SCM medium layer is also arranged in the distributed storage system to store target type data through the SCM medium layer and process non-block-aligned write data through the SCM medium layer.

The flash memory medium used in the solution of the present application can generally be QLC or PLC (Penta Level Cell, five-layer storage unit). Since global wear leveling is achieved, the service life is guaranteed while having high data density, high energy efficiency ratio and low price, that is, a high cost performance is achieved.

This implementation further considers that the data in the hard disk cluster is usually stored in blocks of a set size, and an SCM (Storage-Class Memory) medium layer can also be set in the distributed storage system. By processing the non-block-aligned write data through the SCM medium layer, the IOPS can be effectively improved. After the large block IO (Input/Output) is divided into blocks, there may be a remaining part, which is the non-block-aligned write data, that is, the IO with no aligned boundaries. In addition, some small block IO also belongs to non-block-aligned write data.

In addition, in some implementations, data that needs to be modified very frequently may not be stored in the first medium layer, that is, not stored in the SSD, but may be directly placed in the SCM to achieve high-speed reading and writing of such data, which is also beneficial to further improve the life of the distributed storage system.

Please refer to Figure 3, which is a schematic diagram of a multi-layer flash memory architecture of a distributed storage system in a specific implementation. In the implementation of Figure 3, a first medium layer and an SCM medium layer are set. The first medium layer in Figure 3 is specifically a PLC medium layer, which is also the solution usually selected in practical applications. The PLC medium layer has the advantages of high data density, high energy efficiency ratio, and low price. Through the global wear leveling strategy of this application, the service life of the PLC medium layer is effectively guaranteed and its cost performance is improved. In addition, in some occasions, the first medium layer can also be a QLC medium layer.

In a specific implementation of the present application, in each hard disk cluster, data is stored in blocks of a set size, and the data storage method of the distributed storage system further includes:

Determine the P/E times of each block in the hard disk cluster in the i-th state, and calculate the average value of the P/E times of each block in the hard disk cluster;

When there are blocks in the disk cluster in the i-th state whose difference between the P/E times and the average value is lower than the set first value, all blocks whose difference between the P/E times and the average value is lower than the set first value are taken as blocks to be migrated;

If there is currently a hard disk cluster in the (i+1)th state, each block to be migrated in the hard disk cluster is migrated to the hard disk cluster in the (i+1)th state.

This implementation further takes into account that in the aforementioned implementation, the data type of the data to be written is divided, which is equivalent to predicting the frequency of future modification of the data to be written. It is understandable that there will be deviations in the prediction results, and even for the same data to be written, the actual frequency of modification may be different in different occasions and different time periods. Therefore, in this implementation, data migration will be performed.

The P/E number can also be called the number of erase cycles. If the P/E number of a block is very low, it means that the data in the block is not frequently modified. Conversely, if the P/E number of a block is very high, it means that the data in the block is frequently modified.

Specifically, for the disk cluster in the i-th state, the P/E times of each block in the disk cluster in the i-th state will be determined, and the average value of the P/E times of each block in the disk cluster will be calculated. If the P/E times of a block are much lower than the average value, the block will be used as a block to be migrated. If there is currently a disk cluster in the i+1th state, each block to be migrated in the disk cluster in the i-th state will be migrated to the disk cluster in the i+1th state.

For example, in G1 of FIG2 , the P/E times of two blocks are particularly low, indicating that the modification frequency of these two blocks is very low, so the data of these two blocks are migrated to the hard disk cluster in the G _cold state.

Furthermore, in a specific implementation of the present application, it may also include:

After determining each block to be migrated, if there is no hard disk cluster in the i+1th state, the K blocks with the largest P/E times in the hard disk cluster in the i-th state are exchanged with the data in the K blocks to be migrated in the hard disk cluster in the i-th state, so as to complete the internal migration of the K blocks to be migrated in the hard disk cluster in the i-th state;

Here, K represents the number of blocks to be migrated determined in the hard disk cluster in the i-th state.

For example, a hard disk cluster in the G _readonly state cannot be migrated, that is, a hard disk cluster in the Nth state cannot be migrated because there is no hard disk cluster in the N+1th state.

For another example, when the i-th state is the G _cold state, for a certain hard disk cluster in the G _cold state, after determining a number of blocks to be migrated, it is detected that the storage space of each hard disk cluster in the G _readonly state has been exhausted, and it is also deemed that there is no hard disk cluster in the i+1 state, and the operation of this implementation method will be executed, that is, migration will be performed within the hard disk cluster. In other words, if the migration destination does not exist or the space of the migration destination is full, the blocks to be migrated will be migrated within the source hard disk cluster. shift.

When performing internal migration, the data in the K blocks with the largest P/E times in the hard disk cluster in the i-th state are exchanged with the data in the K blocks to be migrated in the hard disk cluster, thereby further ensuring the wear balance of each block in the hard disk cluster.

For example, there are 100 blocks in a hard disk cluster, among which the P/E times of block 1 is 2, the P/E times of block 2 is 3, the P/E times of block 3 is 90, the P/E times of block 4 is 80, and the P/E times of the remaining 96 blocks are, for example, all 30. In this implementation, block 1 and block 2 are blocks to be migrated, and the two blocks with the largest P/E times in the hard disk cluster are block 3 and block 4. Therefore, the data of block 1 and block 2 need to be exchanged with the data of block 3 and block 4. For example, the data of block 1 can be exchanged with the data of block 3, and the data of block 2 can be exchanged with the data of block 4, so as to complete the internal migration of the two blocks to be migrated in the hard disk cluster.

Applying the technical solution provided by the embodiment of the present application, considering that in order to fully realize the high cost-effectiveness of flash memory media, it is not limited to wear leveling within one hard disk, but it is necessary to perform global wear leveling based on a distributed storage system. Specifically, in the solution of the present application, the distributed storage system is divided into multiple hard disk clusters. Compared with directly managing each hard disk, the management data required for the hard disk cluster is lower, that is, the amount of metadata, which is easy to implement. In addition, for any one hard disk cluster, the hard disk cluster includes multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch. Through such a setting, it is conducive to conveniently realizing the wear leveling of each hard disk in the hard disk cluster. Then, by realizing the wear leveling between the hard disk clusters, the global wear leveling of the entire distributed storage system can be realized, which also ensures the durability of the hard disk and can avoid the situation of frequent replacement of bad disks.

Specifically, when performing wear leveling between hard disk clusters, the present application divides the hard disk cluster into N states. After receiving the data to be written, the data type of the data to be written will be divided. When the data to be written is divided into the i-th data type, it is determined whether there is a hard disk cluster in the i-th state at present. If so, a hard disk cluster in the i-th state will be selected. Since different data types reflect the different frequencies of future modification of the data to be written, and specifically, among the N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the i-th data type within the first time length. And among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1th state is lower than the wear degree of the hard disk cluster in the i-th state. It can be seen that for data that almost does not need to be modified, that is, when the estimated number of modifications of the data to be written within the first time length is very low, the data to be written will be divided into the first data type, and therefore will be written into the hard disk cluster in the first state. The wear degree of the hard disk cluster in the first state is the highest, which means that the hard disk cluster in the first state has written a large amount of data, so the data written is data that almost does not need to be modified. Writing to the hard disk will wear out the hard disk, but reading will not. It can be seen that, since the data written to the hard disk cluster with a high degree of wear is data that hardly needs to be modified, even if the hard disk cluster has a high degree of wear, it can still be read, thus giving full play to its residual value. Correspondingly, the more frequently the data to be written needs to be modified, that is, the higher the estimated number of modifications to the data to be written within the first time period, the data to be written will be written to the hard disk cluster with a lower degree of wear, so that the hard disk cluster with a lower degree of wear can be used more fully, thus realizing the global wear leveling of the distributed storage system.

In summary, the present application divides the distributed storage system into multiple hard disk clusters, which is conducive to conveniently realizing global wear leveling of the distributed storage system, thereby ensuring the durability of the hard disks in the distributed storage system and avoiding the frequent replacement of bad disks.

Corresponding to the above method embodiment, the embodiment of the present application further provides a data storage system of a distributed storage system, which can be referenced in correspondence with the above.

The distributed storage system includes multiple hard disk clusters, each hard disk cluster includes multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch. The data storage system includes:

The type classification module 401 is used to receive the data to be written and classify the data types of the data to be written; wherein, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the ith data type within the first time length, and N is a positive integer not less than 2;

The hard disk cluster status judgment module 402 is used to judge whether there is a hard disk cluster in the i-th state when the data type of the data to be written is the i-th data type; if there is a hard disk cluster in the i-th state, the hard disk cluster selection module 403 is triggered;

The hard disk cluster selection module 403 is used to select a hard disk cluster in the i-th state;

The writing module 404 is used to write the data to be written into the selected hard disk cluster in the i-th state; wherein, among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1-th state is lower than the wear degree of the hard disk cluster in the i-th state, and i is a positive integer.

In a specific implementation of the present application, the type classification module 401 classifies the data types of the data to be written, including:

Based on the file name of the data to be written, the data type of the data to be written is divided.

In a specific implementation of the present application, the type classification module 401 classifies the data types of the data to be written based on the file names of the data to be written, including:

When the file name of the data to be written matches the preset j-th database, the data type of the data to be written is divided into the j-th data type; wherein j is a positive integer and 1≤j≤N-1;

When the file name of the data to be written does not match any of the preset N-1 databases, the data type of the data to be written is divided into the Nth data type.

In a specific implementation of the present application, a first update module is also included, which is used to:

The file name is used as a training sample, and the modification count of the training sample in the first time period is used as a training label of the training sample to train the preset deep learning model;

After the deep learning model is trained, different file names are input into the trained deep learning model in sequence, and the data of N-1 databases are updated based on the output results of the deep learning model.

In a specific implementation of the present application, a second updating module is further included, which is used to:

An adjustment instruction for the jth database is received, and a data item adding operation, a data item deleting operation, and/or a data item modifying operation is performed on the jth database according to the adjustment instruction.

In a specific implementation of the present application, N=3, the first data type is a read-only data type, the second data type is a cold data type, the third data type is a hot data type, the first state of the hard disk cluster is a G _readonly state, the second state of the hard disk cluster is a G _cold state, and the third state of the hard disk cluster is a G _hot state.

In a specific implementation of the present application, a first execution module is further included, which is used to:

When the data to be written is of the first data type and it is determined that there is no hard disk cluster in the first state currently, the data to be written is written into a hard disk cluster in the G _hot state or the G _cold state.

In a specific implementation of the present application, a second execution module is further included, which is used to:

When the data to be written is of the second data type and it is determined that there is no hard disk cluster in the second state currently, the data to be written is written into the hard disk cluster in the G _hot state.

In a specific implementation of the present application, a third execution module is further included, which is used to:

When the data to be written is of the third data type and it is determined that there is no disk cluster in the third state currently, a prompt message indicating write failure is fed back.

In a specific implementation of the present application, the hard disk cluster selection module 403 is specifically used to:

According to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, a hard disk cluster in the i-th state is selected.

In a specific implementation of the present application, the hard disk cluster selection module 403 is specifically used to:

For each hard disk cluster in the i-th state, search in order from the smallest to the largest wear degree;

When the current cluster write queue depth VG _{cur_queue_depth} of any disk cluster in the i-th state is found to be less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the search is stopped and the disk cluster is used as the selected disk cluster in the i-th state;

After searching all disk clusters in the i-th state, if there is no disk cluster whose current cluster write queue depth VG _{cur_queue_depth} is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the disk cluster with the smallest cluster busyness VGbusy is selected as the disk cluster in the i-th state;

The cluster busyness VGbusy of the hard disk cluster represents the value obtained by dividing the current cluster write queue depth VG _{cur_queue_depth} of the hard disk cluster by the maximum cluster write queue depth VG _{max_queue_depth} .

In a specific implementation of the present application, the writing module 404 is specifically used for:

According to the rule that the lower the wear degree of the hard disk, the higher the priority, the target hard disk is selected from the selected hard disk cluster in the i-th state;

Write the data to be written into the selected target hard disk.

In a specific implementation of the present application, the writing module 404 is specifically used for:

Search each hard disk in the selected hard disk cluster in the i-th state in order of the wear degree from small to large;

When it is found that the current hard disk write queue depth VD _{cur_queue_depth} of any hard disk in the hard disk cluster in the i-th state is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the search is stopped and the hard disk is selected as the target hard disk;

After searching all hard disks in the hard disk cluster in the i-th state, if there is no hard disk whose current hard disk write queue depth VD _{cur_queue_depth} is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the hard disk with the smallest hard disk busyness VDbusy is selected as the target hard disk;

The hard disk busyness VDbusy of the hard disk represents a value obtained by dividing the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk by the maximum hard disk write queue depth VD _{max_queue_depth} .

In a specific embodiment of the present application, each hard disk cluster in the distributed storage system is arranged in the first medium layer, and an SCM medium layer is also arranged in the distributed storage system to store target type data through the SCM medium layer and process non-block-aligned write data through the SCM medium layer.

In a specific implementation of the present application, the first dielectric layer is a PLC dielectric layer or a QLC dielectric layer.

In a specific implementation of the present application, in each hard disk cluster, data is stored in blocks of a set size, and a migration module is also included for:

Determine the P/E times of each block in the hard disk cluster in the i-th state, and calculate the average value of the P/E times of each block in the hard disk cluster;

When there are blocks in the disk cluster in the i-th state whose difference between the P/E times and the average value is lower than the set first value, all blocks whose difference between the P/E times and the average value is lower than the set first value are taken as blocks to be migrated;

If there is currently a hard disk cluster in the (i+1)th state, each block to be migrated in the hard disk cluster is migrated to the hard disk cluster in the (i+1)th state.

In a specific implementation of the present application, the migration module is also used to:

After determining each block to be migrated, if there is no hard disk cluster in the i+1th state, the K blocks with the largest P/E times in the hard disk cluster in the i-th state are exchanged with the data in the K blocks to be migrated in the hard disk cluster in the i-th state, so as to complete the internal migration of the K blocks to be migrated in the hard disk cluster in the i-th state;

Here, K represents the number of blocks to be migrated determined in the hard disk cluster in the i-th state.

Corresponding to the above method and system embodiments, the embodiments of the present application also provide a data storage device of a distributed storage system and a computer non-volatile readable storage medium, which can be referenced in correspondence with the above.

The data storage device of the distributed storage system may include:

Memory 501, used for storing computer programs;

The processor 502 is used to execute a computer program to implement the steps of the data storage method of the distributed storage system.

Referring to FIG6 , a computer program 61 is stored on the non-volatile computer readable storage medium 60. When the computer program 61 is executed by the processor, the steps of the data storage method of the distributed storage system in any of the above embodiments are implemented. The non-volatile computer readable storage medium 60 mentioned here includes a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the technical field.

It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

Specific examples are used herein to illustrate the principles and implementation methods of the present application, and the description of the above embodiments is only used to help understand the technical solution and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the present application.

Claims (20)

1. A data storage method for a distributed storage system, characterized in that the distributed storage system includes multiple hard disk clusters, the hard disk clusters include multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch, and the data storage method for the distributed storage system includes:

   Receive data to be written, and classify the data types of the data to be written; wherein, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the ith data type within the first time length, and N is a positive integer not less than 2;

   When the data type of the data to be written is the i-th data type, determining whether there is a hard disk cluster in the i-th state currently;

   If there is a hard disk cluster in the i-th state, select a hard disk cluster in the i-th state;

   The data to be written is written into the selected hard disk cluster in the i-th state; wherein, among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1-th state is lower than the wear degree of the hard disk cluster in the i-th state, and i is a positive integer.
2. The data storage method of a distributed storage system according to claim 1, characterized in that said dividing the data type of the data to be written comprises:

   Based on the file name of the data to be written, the data type of the data to be written is divided.
3. The data storage method of a distributed storage system according to claim 2, characterized in that the data type of the data to be written is divided based on the file name of the data to be written, comprising:

   When the file name of the data to be written matches the preset j-th database, the data type of the data to be written is divided into the j-th data type; wherein j is a positive integer and 1≤j≤N-1;

   When the file name of the data to be written does not match any of the preset N-1 databases, the data type of the data to be written is divided into an Nth data type.
4. The data storage method of the distributed storage system according to claim 3, characterized in that it also includes:

   Using the file name as a training sample and the statistical value of the number of modifications of the training sample within the first time period as a training label of the training sample, training the preset deep learning model;

   After the deep learning model is trained, different file names are input into the trained deep learning model in sequence, and based on the output results of the deep learning model, data of N-1 databases are updated.
5. The data storage method of the distributed storage system according to claim 3, characterized in that it also includes:

   An adjustment instruction for the jth database is received, and a data item adding operation, a data item deleting operation, and/or a data item modifying operation is performed on the jth database according to the adjustment instruction.
6. The data storage method of a distributed storage system according to claim 1 is characterized in that N=3, the first data type is a read-only data type, the second data type is a cold data type, the third data type is a hot data type, the first state of the hard disk cluster is a G _readonly state, the second state of the hard disk cluster is a G _cold state, and the third state of the hard disk cluster is a G _hot state.
7. The data storage method of the distributed storage system according to claim 6, characterized in that it also includes:

   When the data to be written is of the first data type and it is determined that there is no hard disk cluster in the first state currently, the data to be written is written into a hard disk cluster in the G _hot state or the G _cold state.
8. The data storage method of the distributed storage system according to claim 6, characterized in that it also includes:

   When the data to be written is of the second data type and it is determined that there is no hard disk cluster in the second state currently, the data to be written is written into the hard disk cluster in the G _hot state.
9. The data storage method of the distributed storage system according to claim 6, characterized in that it also includes:

   When the data to be written is of the third data type and it is determined that there is no hard disk cluster in the third state currently, a prompt message indicating write failure is fed back.
10. The data storage method of a distributed storage system according to claim 1, wherein the step of selecting a hard disk cluster in the i-th state comprises:

    According to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, a hard disk cluster in the i-th state is selected.
11. The data storage method of a distributed storage system according to claim 10, characterized in that the step of selecting a hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk cluster, the higher the priority, comprises:

    For each hard disk cluster in the i-th state, search in order from the smallest to the largest wear degree;

    When the current cluster write queue depth VG _{cur_queue_depth} of any disk cluster in the i-th state is found to be less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the search is stopped and the disk cluster is used as the selected disk cluster in the i-th state;

    After searching all disk clusters in the i-th state, if there is no disk cluster whose current cluster write queue depth VG _{cur_queue_depth} is less than the preset maximum cluster write queue depth VG _{max_queue_depth} , the disk cluster with the smallest cluster busyness VGbusy is selected as the disk cluster in the i-th state;

    The cluster busyness VGbusy of the hard disk cluster represents a value obtained by dividing the current cluster write queue depth VG _{cur_queue_depth} of the hard disk cluster by the maximum cluster write queue depth VG _{max_queue_depth} .
12. The data storage method of the distributed storage system according to claim 1, characterized in that writing the data to be written into the selected hard disk cluster in the i-th state comprises:

    According to the rule that the lower the wear degree of the hard disk, the higher the priority, the target hard disk is selected from the selected hard disk cluster in the i-th state;

    The data to be written is written into the selected target hard disk.
13. The data storage method of a distributed storage system according to claim 12, characterized in that the step of selecting a target hard disk from the selected hard disk cluster in the i-th state according to the rule that the lower the wear degree of the hard disk, the higher the priority, comprises:

    Search each hard disk in the selected hard disk cluster in the i-th state in order of the wear degree from small to large;

    When it is found that the current hard disk write queue depth VD _{cur_queue_depth} of any hard disk in the hard disk cluster in the i-th state is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the search is stopped and the hard disk is selected as the target hard disk;

    After searching all hard disks in the hard disk cluster in the i-th state, if there is no hard disk whose current hard disk write queue depth VD _{cur_queue_depth} is less than the preset maximum hard disk write queue depth VD _{max_queue_depth} , the hard disk with the smallest hard disk busyness VDbusy is selected as the target hard disk;

    The hard disk busyness VDbusy of the hard disk represents a value obtained by dividing the current hard disk write queue depth VD _{cur_queue_depth} of the hard disk by the maximum hard disk write queue depth VD _{max_queue_depth} .
14. The data storage method of a distributed storage system according to claim 1 is characterized in that each of the hard disk clusters in the distributed storage system is arranged in a first medium layer, and the distributed storage system is further provided with an SCM medium layer to store target type data through the SCM medium layer, and to process unprocessed data through the SCM medium layer. Block-aligned write data.
15. The data storage method of a distributed storage system according to claim 14 is characterized in that the first medium layer is a PLC medium layer or a QLC medium layer.
16. The data storage method of a distributed storage system according to any one of claims 1 to 15, characterized in that in each hard disk cluster, data is stored in blocks of a set size, and the data storage method of a distributed storage system further comprises:

    Determine the P/E times of each block in the hard disk cluster in the i-th state, and calculate the average value of the P/E times of each block in the hard disk cluster;

    When there are blocks in the hard disk cluster in the i-th state whose difference between the P/E times and the average value is lower than the set first value, all blocks whose difference between the P/E times and the average value is lower than the set first value are taken as blocks to be migrated;

    If there is currently a hard disk cluster in the (i+1)th state, each of the blocks to be migrated in the hard disk cluster is migrated to the hard disk cluster in the (i+1)th state.
17. The data storage method of the distributed storage system according to claim 16, characterized in that it also includes:

    After determining each of the blocks to be migrated, if there is no hard disk cluster in the i+1th state, then the K blocks with the largest P/E times in the hard disk cluster in the i-th state are exchanged with the data in the K blocks to be migrated in the hard disk cluster in the i-th state, so as to complete the internal migration of the K blocks to be migrated in the hard disk cluster in the i-th state;

    Here, K represents the number of blocks to be migrated determined in the hard disk cluster in the i-th state.
18. A data storage system of a distributed storage system, characterized in that the distributed storage system includes multiple hard disk clusters, the hard disk clusters include multiple hard disks of the same model, and each hard disk in the same hard disk cluster is added to the distributed storage system in the same batch, and the data storage system of the distributed storage system includes:

    A type classification module, used for receiving data to be written and classifying the data types of the data to be written; wherein, among the set N data types, the estimated number of modifications of the data to be written of the i+1th data type within the first time length is higher than the estimated number of modifications of the data to be written of the ith data type within the first time length, and N is a positive integer not less than 2;

    A hard disk cluster status judgment module, used for judging whether there is a hard disk cluster in the i-th state when the data type of the data to be written is the i-th data type; if there is a hard disk cluster in the i-th state, triggering the hard disk cluster selection module;

    The hard disk cluster selection module is used to select a hard disk cluster in the i-th state;

    A writing module is used to write the data to be written into the selected hard disk cluster in the i-th state; wherein, among the hard disk clusters in the set N states, the wear degree of the hard disk cluster in the i+1-th state is lower than the wear degree of the hard disk cluster in the i-th state, and i is a positive integer.
19. A data storage device of a distributed storage system, characterized by comprising:

    Memory for storing computer programs;

    A processor, configured to execute the computer program to implement the steps of the data storage method of the distributed storage system as described in any one of claims 1 to 17.
20. A computer non-volatile readable storage medium, characterized in that a computer program is stored on the computer non-volatile readable storage medium, and when the computer program is executed by a processor, it implements any one of claims 1 to 17 The steps of the data storage method of the distributed storage system described in item.