Data output fault tolerance apparatus and method in a high performance computing environment

Abstract

The application discloses a kind of data output fault-tolerant device and fault-tolerant method in high-performance computing environment, and relates to data output fault-tolerant technical field.The application includes storage monitoring module, and with its communication connection of cutting module, output module and coding module, and check repair module;Cutting module is used to divide the data that client writes into data block in storage monitoring module;Coding module generates coding block by cross-group coding of data block through coding algorithm, and forms coding group;Check repair module is used to check coding group, and simultaneously repairs missing coding by tree-shaped repair method.The application is provided with multiple copies, can reduce the probability of content data block loss in the output process, and content data block is as the main part of data, strengthens the protection of content data, improves the security of data output, while reducing the cost of data fault-tolerance.

Description

Technical Field

[0001] This invention belongs to the field of data output fault tolerance technology, and relates to a data output fault tolerance device in a high-performance computing environment, as well as a fault tolerance method for the data output fault tolerance device in a high-performance computing environment. Background Technology

[0002] As computing power grows, the amount of data output in parallel by high-performance computers also increases. The scale of various nodes, network devices, and storage devices involved in the data output process also grows accordingly, leading to an increase in localized component failures per unit time. Data fault tolerance refers to the technology that, with the help of computers, can effectively preserve and reuse data in the event of partial or overall failures, reducing organizational information loss. Modern society is highly dependent on information technology, and the vast amounts of data forming information assets require storage using computers and networks. This necessitates strengthened technological research and management to improve the fault tolerance capabilities of network-stored data.

[0003] Currently, facing the common occurrence of output node failures, multi-replica technology, as a very mature fault-tolerance technique, is used by many systems. This involves creating multiple identical copies of a data object by copying it multiple times, and then distributing these copies across multiple nodes. This allows the use of other copies when some copies fail due to node failure. However, this significantly reduces storage space utilization. Therefore, more and more storage systems are using erasure coding technology for data fault tolerance. Compared to multi-replica technology, erasure coding can achieve the same or even higher fault tolerance with much lower storage space overhead. However, its higher network and disk resource overhead becomes the performance bottleneck of the entire system, leading to higher data encoding costs and lower data repair efficiency. Furthermore, during data output, in most cases only a small portion of the data is corrupted, resulting in the entire data being unreadable on the client. Repairing the data requires re-sending the entire output data, which consumes a significant amount of output resources.

[0004] Existing data output fault tolerance technologies in high-performance computing environments suffer from high data encoding costs and low data repair efficiency. Furthermore, data repair consumes significant system resources, leading to a heavy system load. To address this, we propose a data output fault tolerance device and method for high-performance computing environments. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a data output fault tolerance device and fault tolerance method in a high-performance computing environment, which solves the problems of high data encoding cost, low data repair efficiency, and high system resource consumption during data repair, resulting in a large system load in existing data output fault tolerance technologies in high-performance computing environments.

[0006] The technical solution adopted in this invention is as follows: a data output fault-tolerant device in a high-performance computing environment, including a storage monitoring module, and a slitting module, an output module, an encoding module, and a verification and repair module that establish communication connections with it;

[0007] Segmentation module: Used to divide the data written by the client to the storage monitoring module into data blocks;

[0008] Encoding module: Generates encoded blocks by cross-grouping data blocks using an encoding algorithm, and forms encoded groups;

[0009] Output module: Used to output the encoded data stored on a node to the client;

[0010] Verification and Repair Module: Used to verify the code group and repair missing codes using a tree-structured repair method.

[0011] Furthermore, the aforementioned storage monitoring module is used to store the data written by the client and monitor the frequency of request output, and to count the request frequency of the data in real time. Based on the request frequency of the data, it determines whether the data is hot data, and stores copies of the hot data and generates a corresponding number of temporary hot data copies based on the popularity of the hot data. For non-hot data, it stores redundant data and deletes its temporary hot data copies.

[0012] Furthermore, the above-mentioned segmentation module segments the data according to the data format as a partitioning strategy, forming three types of data blocks: file header data blocks, content data blocks, and file number data blocks. Except for the content data blocks containing text data, the remaining data blocks are regarded as redundant data blocks. The number of redundant data blocks is much greater than the number of content data blocks. At the same time, the temporary hotspot copies of hot data are segmented, and the content data blocks in the temporary hotspot copies are saved as temporary content copy data blocks.

[0013] Furthermore, the cross-group coding used in the above coding module divides the segmented redundant data blocks, content data blocks, and temporary content copy data blocks into several subgroups. The number of subgroups does not exceed the number of content data blocks. Each subgroup generates a check block that is only related to the data blocks within the group. In the cross-group coding, each subgroup contains the same number of data blocks, and some of the data blocks in each group are the same. The same data blocks are selected from all content data blocks and temporary content copy data blocks in the coding group. Each subgroup consists of several redundant data blocks, at least one content data block, and one temporary content copy block. The number of cross-content data blocks in each subgroup is determined by the bandwidth load capacity of the output node and the new node (the stronger the bandwidth load capacity of the output node and the new node, the more cross-content data blocks in each subgroup, but it does not exceed the node load capacity threshold).

[0014] Furthermore, the above-mentioned verification and repair module verifies the output data through verification blocks, excludes lost content copy data blocks, and verifies whether there are any lost content data blocks. The verification block is the part of all the encoding blocks generated by the encoding module other than the blocks obtained by data splitting. The encoding group of a (k,m,k′)-encoding module contains the sum of k redundant data blocks, content data blocks and temporary content data block copies, collectively referred to as data blocks, m verification blocks, and k′ is the remaining data block.

[0015] A data output fault tolerance method for high-performance computing environments, comprising the following steps:

[0016] S1. The client requests to write file data to the data server and stores the file data in the storage monitoring module. At the same time, based on the access frequency of the file data over a period of time, it determines whether it is hot data and backs up the hot data to generate a corresponding number of temporary hot data copies.

[0017] S2. The client requests output file data from the data server. The data server returns a list of data block locations, queries the storage node of the file to be output, and retrieves the file data.

[0018] S3. The splitting module splits the data and temporary hot copy according to the data format as the partitioning strategy, and finally obtains redundant file data blocks, content data blocks and temporary content copy data blocks.

[0019] S4. The encoding module cross-divides the data blocks generated by the segmentation module into several subgroups. Each subgroup generates a check block that is only related to the data blocks within the group. All subgroups form a complete encoding group.

[0020] S5. The output module outputs the encoding group to the client, where the verification and repair module verifies the encoding group and repairs lost data using a tree-structured repair method.

[0021] Furthermore, in step S4 above, the encoding module uses a centralized conversion method to encode the data block. The encoding method includes the following specific steps:

[0022] S4.1 The encoding module downloads all data blocks in each subgroup from the nodes where each subgroup is stored;

[0023] S4.2 Calculate and generate check blocks for these subgroup data blocks;

[0024] S4.3 Send the check block to the node where the corresponding subgroup is stored.

[0025] Furthermore, the implementation method of the encoding group of a (k,m,k′)-encoding module in the above cross-coding includes the following specific steps:

[0026] a. Divide the k data blocks into y overlapping subgroups, each subgroup containing n data blocks;

[0027] b. The check block of each subgroup is generated by operating on n data blocks within the subgroup (the check block is generated by matrix operation through the encoding module).

[0028] Furthermore, the tree-based repair method used in step S5 above includes the following specific steps:

[0029] S5.1. Construct a spanning tree, called the repair tree, within the verification and repair module, with the replacement node as the root and covering all provided nodes.

[0030] S5.2 Repairing a leaf node in the tree first multiplies its own encoded block by the corresponding repair coefficient, and then sends the resulting intermediate encoded block to its parent node.

[0031] S5.3 Repair the internal nodes of the tree to receive intermediate encoded blocks from all child nodes, combine these intermediate encoded blocks with the encoded blocks they store, and then send the combined intermediate encoded blocks to their parent nodes.

[0032] S5.4 The root node can obtain the lost code block by further combining the intermediate code blocks received from all its child nodes;

[0033] Among them, the lost coded block C n By calculating the remaining coded blocks C1, C2, C3, and C4 n-1 linear groups ...

[0035] The combination of these elements can be used to repair the damage, and the combination relationship is as follows: C n =γ1C1+γ2C2+γ3C3+...+γ n-1 C n-1 γ1, γ2, γ3...γn-1 This is the repair coefficient.

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] 1) This invention provides a novel fault tolerance mechanism for output data by combining a replica fault tolerance method with an error correction coding fault tolerance method, thus overcoming the shortcomings of each method. By monitoring the data to be output and identifying frequently requested data as "hot data" based on its request frequency, this invention backs up the hot data, reducing the risk of data loss during frequent outputs. Simultaneously, a data segmentation module segments the data according to its format, resulting in roughly three categories: file header data blocks and file count data blocks coexisting as redundant data blocks with the content data blocks. The redundant data blocks are further divided into... Normally, data will not be lost. The content data block, which is the main body of the data, improves its fault tolerance efficiency through backup. That is, redundant data blocks, content data blocks, and content copy data blocks are encoded by the encoding module as data blocks. The encoding module adopts the method of cross-group encoding to encode the data blocks. In this way, if some data blocks are lost, only the data blocks within the cross subgroup need to be repaired, instead of repairing all data blocks. This reduces the repair cost by repairing a large number of data blocks within the group. Furthermore, by using content data blocks and content copy data blocks as cross data blocks, it is equivalent to setting up multiple copies of the content data blocks, which reduces the risk of content data block loss.

[0038] 2) In this invention, the output data is verified through a verification and repair module. During the verification process, lost content copy data blocks are ignored, and only the content data blocks and redundant data blocks are verified. If any lost data blocks are found, they are immediately repaired, which reduces the workload of data repair. Multiple copies of the content data blocks are made, which reduces the probability of content data block loss during the output process. Furthermore, as the main part of the data, the content data blocks strengthen the protection of the content data, improve the security of data output, and reduce the cost of data fault tolerance. Attached Figure Description

[0039] Figure 1 A system block diagram of a data output fault-tolerant device in a high-performance computing environment;

[0040] Figure 2 A flowchart for a data output fault tolerance method in a high-performance computing environment. Detailed Implementation

[0041] The present invention will be further described below with reference to specific embodiments.

[0042] Example 1: See Figure 1 The present invention is a data output fault tolerance device in a high-performance computing environment, including a storage monitoring module, and a slitting module, an output module, an encoding module, and a verification and repair module that establish communication connections with it.

[0043] Segmentation module: Used to divide the data written by the client to the storage monitoring module into data blocks;

[0044] Encoding module: Generates encoded blocks by cross-grouping data blocks using an encoding algorithm, and forms encoded groups;

[0045] Output module: Used to output the encoded data stored on a node to the client;

[0046] Verification and Repair Module: Used to verify the code group and repair missing codes using a tree-structured repair method.

[0047] The storage monitoring module stores the data written by the client and monitors the frequency of request output. It calculates the request frequency of the data in real time, determines whether the data is hot data based on the request frequency, stores copies of the hot data, and generates a corresponding number of temporary hot data copies based on the popularity of the hot data. For non-hot data, it stores redundant data and deletes its temporary hot data copies.

[0048] In the data splitting module, the data is split into three types of data blocks based on the data format as the partitioning strategy: file header data block, content data block, and file number data block. Except for the content data block containing text data, the other data blocks are regarded as redundant data blocks. The number of redundant data blocks is much greater than the number of content data blocks. At the same time, the temporary hot spot copy of the hot data is split, and the content data block in the temporary hot spot copy is saved as the temporary content copy data block.

[0049] The cross-group coding used in the encoding module divides the split redundant data blocks, content data blocks, and temporary content copy data blocks into several subgroups. The number of subgroups does not exceed the number of content data blocks. Each subgroup generates a check block that is only related to the data blocks within the group. Each subgroup in the cross-group coding contains the same number of data blocks, and some of the data blocks in each group are the same. The same data blocks are selected from all content data blocks and temporary content copy data blocks in the coding group. Each subgroup consists of several redundant data blocks, at least one content data block, and one temporary content copy block. The number of cross-content data blocks in each subgroup is determined by the bandwidth load capacity of the output node and the new node.

[0050] In the verification and repair module, the output data is verified through verification blocks. The verification block is the part of all the encoding blocks generated by the encoding module except for the blocks obtained by data splitting. The encoding group of a (k,m,k′)-encoding module contains the sum of k redundant data blocks, content data blocks and temporary content data block copies, collectively referred to as data blocks, m verification blocks, and k′ as the remaining data blocks.

[0051] Example 2: See Figure 2 A data output fault tolerance method for high-performance computing environments includes the following steps:

[0052] S1. The client requests to write file data to the data server and stores the file data in the storage monitoring module. At the same time, based on the access frequency of the file data over a period of time, it determines whether it is hot data and backs up the hot data to generate a corresponding number of temporary hot data copies.

[0053] S2. The client requests output file data from the data server. The data server returns a list of data block locations, queries the storage node of the file to be output, and retrieves the file data.

[0054] S3. The splitting module splits the data and temporary hot copy according to the data format as the partitioning strategy, and finally obtains redundant file data blocks, content data blocks and temporary content copy data blocks.

[0055] S4. The encoding module cross-divides the data blocks generated by the segmentation module into several subgroups. Each subgroup generates a check block that is only related to the data blocks within the group. All subgroups form a complete encoding group.

[0056] The encoding module uses a centralized conversion method to encode data blocks, including the following specific steps:

[0057] S4.1 The encoding module downloads all data blocks in each subgroup from the nodes where each subgroup is stored;

[0058] S4.2 Calculate and generate check blocks for these subgroup data blocks;

[0059] S4.3 Send the check block to the node where the corresponding subgroup is stored;

[0060] S5. The output module outputs the encoding group to the client, where the verification and repair module verifies the encoding group and repairs lost data using a tree-structured repair method.

[0061] The encoding group of a (k,m,k′)-encoding module in cross-coding includes the following specific steps:

[0062] a. Divide the k data blocks into y overlapping subgroups, each subgroup containing n data blocks;

[0063] b. The check block of each subgroup is generated by processing n data blocks within the subgroup;

[0064] The tree-shaped restoration method employed includes the following specific steps:

[0065] S5.1. Construct a spanning tree, called the repair tree, within the verification and repair module, with the replacement node as the root and covering all provided nodes.

[0066] S5.2 Repairing a leaf node in the tree first multiplies its own encoded block by the corresponding repair coefficient, and then sends the resulting intermediate encoded block to its parent node.

[0067] S5.3 Repair the internal nodes of the tree to receive intermediate encoded blocks from all child nodes, combine these intermediate encoded blocks with the encoded blocks they store, and then send the combined intermediate encoded blocks to their parent nodes.

[0068] S5.4 The root node can obtain the lost code block by further combining the intermediate code blocks received from all its child nodes;

[0069] Among them, the lost coded block C n By calculating the remaining coded blocks C1, C2, C3, and C4 n-1 linear groups ...

[0071] The combination of these elements can be used to repair the damage, and the combination relationship is as follows: C n =γ1C1+γ2C2+γ3C3+...+γ n-1 C n-1 γ1, γ2, γ3...γ n-1 This is the repair coefficient.

[0072] In this invention, for a (k,m,k′)-corrected code, let k data blocks form a vector, and the generated k+m coded blocks form a vector. Then the coded block C... i The generation can be represented as its encoding coefficient vector α i =(α i,j )=(α i,1 α i,2 ...α i,k The product of ) and vector D, 1≤i≤k+m , 1≤j≤k, as shown in equation (1):

[0073]

[0074] Where, α i,j It is a size of 2 q finite field F qAll operations on the elements of the finite field F are also performed within the finite field F. q The above is performed. Matrix G = (α1α2...α) k+m ) T It is the encoding matrix (or generator matrix). For a (k,m,k′) erasure code, its encoding proof needs to satisfy that any k′×k submatrix has a left inverse matrix. The upper part of the encoding matrix is ​​generally a k×k identity matrix, and the lower part is an m×k matrix. In this way, the generated m×k coded blocks contain the original k data blocks, which facilitates data reading. Such erasure codes are called systematic codes, and the erasure codes used for storage systems are basically erasure codes.

[0075] The decoding process of linear erasure codes involves matrix operations. Without loss of generality, suppose we want to use coded blocks C1C2...C k′ To decode the original k data blocks, let G′ represent (α1α2...α... k′ ) T Let C′ represent (C1C2...C k′ ) T According to equation (1), we can obtain G′×D=C′. Assume G′ L -1 G is the left inverse matrix of G′. Multiply both sides of the equation G′×D=C′ by G′ on the left. L -1 Therefore, the original k data blocks can be obtained as follows:

[0076]

[0077] In this invention, let the reliability of the cloud storage site be Rstorage, the reliability of the storage nodes be Rchunk, and the data redundancy factor be p. Then, the reliability model of a cloud storage site composed of a large number of storage nodes is:

[0078] R storage =1-(1-R) chunk ) p (2)

[0079] Let the reliability of a cloud storage site using fault-tolerant technology based on the RS codec algorithm be RRS, the reliability of the storage node be Rchunk, the number of file blocks be k, the number of encoded data blocks be n, and the redundancy factor be s (s = n / k). Then the reliability model of the cloud storage site using fault-tolerant technology based on the RS codec algorithm is:

[0080]

[0081] Let the reliability of a cloud storage site employing a combination of two fault-tolerance mechanisms be RcStor. Substituting equation (2) into equation (3), we can obtain the reliability model of RcStor as follows:

[0082]

[0083] Similarly, taking the reliability of a single storage node as Rchunk = 0.9, the number of block replicas as p = 2, the number of file blocks as k = 1~20, and the file encoding data blocks encoded according to the encoding redundancy multiple as s = 1~4, and knowing that the number of replicas in the block replica storage is 2, we can see the relationship between the cloud storage site using the two fault tolerance mechanisms and the encoding redundancy multiple and the number of file blocks. When the encoding redundancy multiple is equal to 1 (i.e., no encoding), RcStor decreases as the number of file blocks increases; when the encoding redundancy multiple is greater than or equal to 2, RcStor increases as the number of file blocks increases; when the number of file blocks reaches 5 or more, RcStor is greater than 0.999. However, in practical applications, when the encoding redundancy multiple is greater than 1, RcStor will increase as the number of file blocks increases, and the effect of RcStor increasing with the number of file blocks becomes more significant as the encoding redundancy multiple increases. Therefore, considering the reliability of the cloud storage site and the utilization rate of storage space resources, the number of replicas in the block replica storage is set to 2, the coding redundancy multiple is set to 1.25, and the number of file blocks is 8, that is, the file coding data is divided into 10 blocks. At this time, the reliability of the cloud storage site can reach 0.9999.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the claims.

Claims

1. A data output fault-tolerant device for a high-performance computing environment, characterized in that: It includes a storage monitoring module, as well as a slicing module, an output module, and an encoding module that establish communication connections with it, and a verification and repair module; Segmentation module: Used to divide the data written by the client to the storage monitoring module into data blocks; Encoding module: Generates encoded blocks by cross-grouping data blocks using an encoding algorithm, and forms encoded groups; Output module: Used to output the encoded data stored on a node to the client; Verification and Repair Module: Used to verify the code group and repair missing codes using a tree-structured repair method; The storage monitoring module is used to store the data written by the client and monitor the request output frequency. It counts the requested frequency of the data in real time, determines whether it is hot data based on the request frequency, stores hot data in duplicate, and generates a corresponding number of temporary hot data copies based on the popularity of the hot data. Non-hot data is stored redundantly and its temporary hot data copies are deleted. The segmentation module segments the data according to the data format as a partitioning strategy, forming three types of data blocks: file header data blocks, content data blocks, and file number data blocks. Except for the content data blocks containing text data, the remaining data blocks are considered as redundant data blocks. The number of redundant data blocks is much greater than the number of content data blocks. At the same time, the temporary hotspot copies of hot data are segmented, and the content data blocks in the temporary hotspot copies are saved as temporary content copy data blocks. The cross-group coding used in the encoding module divides the split redundant data blocks, content data blocks, and temporary content copy data blocks into several subgroups. The number of subgroups does not exceed the number of content data blocks. Each subgroup generates a check block that is only related to the data blocks within the group. Each subgroup in the cross-group coding contains the same number of data blocks, and some of the data blocks in each group are the same. The same data blocks are selected from all content data blocks and temporary content copy data blocks in the coding group. Each subgroup includes several redundant data blocks, at least one content data block, and one temporary content copy block. The number of cross-content data blocks in each subgroup is determined by the bandwidth load capacity of the output node and the new node.

2. The data output fault-tolerant device in a high-performance computing environment according to claim 1, characterized in that, The verification and repair module verifies the output data using verification blocks, excluding lost content copy data blocks, and checks for any missing content data blocks. The verification block is the portion of all encoded blocks generated by the encoding module excluding the blocks obtained from data splitting. The encoding module's encoding group consists of k redundant data blocks, content data blocks, and temporary content data block copies, collectively referred to as data blocks, and m check blocks. For the remaining data blocks.

3. A data output fault tolerance method in a high-performance computing environment, applied to the data output fault tolerance device in any one of claims 1-2, characterized in that, Includes the following steps: S1. The client requests to write file data to the data server and stores the file data in the storage monitoring module. At the same time, based on the access frequency of the file data over a period of time, it determines whether it is hot data and backs up the hot data to generate a corresponding number of temporary hot data copies. S2. The client requests output file data from the data server. The data server returns a list of data block locations, queries the storage node of the file to be output, and retrieves the file data. S3. The splitting module splits the data and temporary hot copy according to the data format as the partitioning strategy, and finally obtains redundant file data blocks, content data blocks and temporary content copy data blocks. S4. The encoding module cross-divides the data blocks generated by the segmentation module into several subgroups. Each subgroup generates a check block that is only related to the data blocks within the group. All subgroups form a complete encoding group. S5. The output module outputs the encoding group to the client. The verification and repair module verifies the encoding group, excludes lost content copy data blocks, verifies whether there are any lost content data blocks, and repairs the lost data using the tree-structured repair method.

4. The data output fault-tolerant method in a high-performance computing environment according to claim 3, characterized in that, In step S4, the encoding module uses a centralized conversion method to encode the data block. The encoding method includes the following specific steps: S4.1 The encoding module downloads all data blocks in each subgroup from the nodes where each subgroup is stored; S4.2 Calculate and generate check blocks for the data blocks of these subgroups; S4.3 Send the check block to the node where the corresponding subgroup is stored.

5. The data output fault-tolerant method in a high-performance computing environment according to claim 4, characterized in that, One of the cross-block coding The implementation method of the encoding module's encoding group includes the following specific steps: a. Divide the k data blocks into y overlapping subgroups, each subgroup containing n data blocks; b. The check block of each subgroup is generated by operating on n data blocks within the subgroup.

6. The data output fault-tolerant method in a high-performance computing environment according to claim 5, characterized in that, The tree-shaped repair method used in step S5 includes the following specific steps: S5.

1. Construct a spanning tree, called the repair tree, within the verification and repair module, with the replacement node as the root and covering all provided nodes. S5.2 Repairing a leaf node in the tree first multiplies its own encoded block by the corresponding repair coefficient, and then sends the resulting intermediate encoded block to its parent node. S5.3 Repair the internal nodes of the tree to receive intermediate encoded blocks from all child nodes, combine these intermediate encoded blocks with the encoded blocks they store, and then send the combined intermediate encoded blocks to their parent nodes. S5.4 The root node further combines the intermediate encoded blocks received from all its child nodes to obtain the lost encoded blocks. Among them, the lost coded blocks By calculating the remaining coded blocks , , ... The linear combination is used for repair, and the combination relationship is as follows: , This is the repair coefficient.

Images