Distributed file system adaptive to GPU cluster training, creation method, electronic equipment and storage medium

Abstract

The invention relates to a distributed file system adaptive to GPU cluster training, a creation method, electronic equipment and a storage medium, and the system comprises a management module which is used for distributing a target GPU cluster to a target training task after the target GPU cluster is distributed to the target training task; a server is created for a plurality of first computer nodes included in the target GPU cluster, a client is created for a plurality of second computer nodes included in the target GPU cluster, and at least one hardware GPU is deployed on each computer node included in the target GPU cluster; the server is used for constructing a GPU cluster distributed file system adaptive to the target training task by utilizing a preset local memory space deployed on the corresponding first computer node; and the client is used for executing IO operation on the GPU cluster distributed file system in the process of executing the target training task by utilizing the target GPU cluster. According to the embodiment of the invention, IO requirements in a GPU cluster training scene can be effectively met.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a distributed file system and creation method, electronic device, and storage medium adapted for GPU cluster training. Background Art

[0002] In the scenario of large model training based on graphics processing unit (GPU) cluster, the use of file system mainly includes three types of input / output (IO) operations: data set reading, checkpoint data writing, and checkpoint data reading. Currently, GPU cluster training generally uses external storage systems, such as general network file systems, parallel file systems, object storage, etc. The data reading performance depends on the performance of the storage system, which may not meet the data reading performance requirements in GPU cluster training. Summary of the invention

[0003] The present invention discloses a distributed file system and a creation method, an electronic device, and a storage medium that are adapted to GPU cluster training.

[0004] According to one aspect of the present disclosure, a distributed file system adapted for GPU cluster training is provided, comprising: a management module, a server and a client; the management module is used to create a server for multiple first computer nodes included in the target GPU cluster and create a client for multiple second computer nodes included in the target GPU cluster after allocating a target GPU cluster for a target training task, wherein each computer node included in the target GPU cluster is deployed with at least one hardware GPU; the server is used to construct a GPU cluster distributed file system adapted for the target training task by using a preset local memory space deployed on the corresponding first computer node; the client is used to perform IO operations on the GPU cluster distributed file system in the process of executing the target training task by using the target GPU cluster.

[0005] In a possible implementation manner, the memory size of the preset local memory space deployed on each first computer node is the same.

[0006] In a possible implementation, the GPU cluster distributed file system is used to store the data set required in the target training task and the checkpoint data in the training process.

[0007] In a possible implementation, the client includes: a system call interception module and a client communication module; the server includes: a server communication module; the system call interception module is used to intercept the target file system call in the second computer node where the client is located, and convert the target file system call into an IO request, wherein the target file system call is used to access the GPU cluster distributed file system; the client communication module is used to send the IO request to the server communication module.

[0008] In a possible implementation, the client communication module includes: a remote procedure call (RPC) client module; the server communication module includes: an RPC server module; the RPC client module is used to send the IO request to the RPC server module using a non-blocking RPC network communication mode.

[0009] In a possible implementation, the client communication module includes: an IPC client module; the server communication module includes: an IPC server module; the IPC client module is used to send the IO request to the IPC server module when the first computing point where the server is located and the second computer node where the client is located are the same computer node.

[0010] In a possible implementation, the client includes: a file map; the file map is used to store access records of the client to the GPU cluster distributed file system.

[0011] In a possible implementation, the IO request includes: a metadata IO request; the client includes: a data segmentation module; the server includes: a metadata database; the data segmentation module is used to divide the target data that needs to be written to the GPU cluster distributed file system into blocks, determine multiple data blocks corresponding to the target data, and metadata information of each data block, wherein the metadata information corresponding to each data block is used to indicate the target server to which the data block needs to be written; the client communication module is used to send the metadata IO request to each server communication module, wherein the metadata IO request is used to request that the metadata information of each data block corresponding to the target data be written to each server; the metadata database in each server is used to respond to the metadata IO request and store the metadata information of each data block corresponding to the target data.

[0012] In a possible implementation, the IO request includes: a data IO request; the server includes: a persistence module; the communication client module is used to send data IO requests to multiple target server ends, wherein the data IO request corresponding to each target server end is used to request that a specified data block corresponding to the target data be written to the target server end; the persistence module in each target server end is used to respond to the received data IO request, and according to the metadata IO request stored in the metadata database in the target server end, store the specified data block to be written to the target server end in the preset local memory space corresponding to the target server end.

[0013] In a possible implementation, the persistence module in each target server stores each designated data block that needs to be written to the target server in a chunk file in the preset local memory space corresponding to the target server.

[0014] In a possible implementation, the number of designated data blocks corresponding to the target data written into each target server is the same.

[0015] In a possible implementation, the data IO request corresponding to each target server includes client memory area information corresponding to the specified data block that needs to be written to the target server; the persistence module in each target server reads the specified data block from the memory space specified by the client memory area information through remote direct memory access RDMA based on the client memory area information included in the received data IO request.

[0016] In a possible implementation, the GPU cluster distributed file system is further used to asynchronously send all or part of the checkpoint data to a PFS outside the target GPU cluster.

[0017] In a possible implementation, the management module is used to delete the client and the server after the target training task is completed using the target GPU cluster.

[0018] According to one aspect of the present disclosure, a method for creating a distributed file system adapted for GPU cluster training is provided, comprising: after allocating a target GPU cluster for a target training task, creating a server for multiple first computer nodes included in the target GPU cluster, and creating a client for multiple second computer nodes included in the target GPU cluster, wherein at least one hardware GPU is deployed on each computer node included in the target GPU cluster; based on the server, a GPU cluster distributed file system adapted for the target training task is constructed using a preset local memory space deployed on the corresponding first computer node; based on the client, in the process of executing the target training task using the target GPU cluster, IO operations are performed on the GPU cluster distributed file system.

[0019] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the above method.

[0020] According to one aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor.

[0021] In an embodiment of the present disclosure, a distributed file system adapted to GPU cluster training is designed, including: a management module, a server, and a client; after allocating a target GPU cluster for a target training task, the management module creates a server for multiple first computer nodes included in the target GPU cluster and creates a client for multiple second computer nodes included in the target GPU cluster according to the data reading and writing requirements of the target training task, so that the server can use the preset local memory space deployed on the corresponding first computer node to flexibly build a GPU cluster distributed file system adapted to the target training task, and then, in the process of using the target GPU cluster to execute the target training task, the client can perform IO operations on the GPU cluster distributed file system to effectively meet the IO requirements in the GPU cluster training scenario.

[0022] It should be understood that the above general description and the following detailed description are exemplary and explanatory only and do not limit the present disclosure. Other features and aspects of the present disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used to illustrate the technical solutions of the present disclosure together with the specification.

[0024] Figure 1 A block diagram of a distributed file system adapted for GPU cluster training according to an embodiment of the present disclosure is shown.

[0025] Figure 2 A schematic diagram showing the interaction between a client and a server according to an embodiment of the present disclosure is shown.

[0026] Figure 3 A schematic diagram showing the transmission of metadata information and data between a client and a server according to an embodiment of the present disclosure is shown.

[0027] Figure 4 A schematic diagram showing a client writing a file to a GPU cluster distributed file system according to an embodiment of the present disclosure.

[0028] Figure 5 A flowchart of a method for creating a distributed file system adapted for GPU cluster training according to an embodiment of the present disclosure is shown.

[0029] Figure 6 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0030] Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0031] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0032] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C can represent including any one or more elements selected from the set consisting of A, B, and C.

[0033] In addition, in order to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present disclosure can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present disclosure.

[0034] In the scenario of large model training based on a graphics processing unit (GPU) cluster, the use of the file system mainly includes three types of input / output (IO) operations: data set reading, checkpoint data writing, and checkpoint data reading.

[0035] Dataset reading: At the beginning of each iteration of GPU cluster training, the training framework (for example, pytorch) uses a data loader to load the dataset. The characteristic of dataset reading is a large amount of random IO, mainly 4k random reads. You can also use some tools to convert the data format of the dataset in advance, so that the dataset reading can convert random reads into sequential read operations.

[0036] Checkpoint data writing: After a certain number of rounds of iterative training of the GPU cluster, the training framework needs to persist the output results of the training process, that is, the checkpoint data. Checkpoint data is saved for two main purposes: on the one hand, it can be directly output to fine-tune the model and continue training; on the other hand, in the event of a failure in subsequent training, training can be continued based on the checkpoint data after failure recovery. Checkpoint data is generally large. For example, the size of the checkpoint data corresponding to a 70B model is about 980GB. Therefore, checkpoint data is usually sliced ​​and distributed for storage and reading. The characteristic of checkpoint data writing is a large number of sequential write operations, mainly 1M sequential writes.

[0037] Checkpoint data reading: In the process of training large models based on large-scale GPU clusters, some software or hardware failures will inevitably cause training interruptions. After the failure is recovered, training needs to continue, and at this time, the previously saved checkpoint data needs to be read. The characteristic of checkpoint data reading is a large number of sequential read operations, mainly 1M sequential reads.

[0038] The storage systems used for reading and writing data sets and checkpoint data for current GPU cluster training can mainly be the following three solutions: 1. General network file systems or distributed file systems, such as NFS, CephFS, etc., can ensure data reliability, but require independent deployment and operation and maintenance, and have relatively poor performance. 2. Parallel file systems, such as lustrefs, gpfs, beegfs, etc., have good performance, but require independent deployment and operation. In terms of physical form, they are mainly divided into all-in-one form (single storage device) and distributed form (multiple storage servers). 3. Object storage, which has good scalability but poor performance. Some scenarios can be accelerated by combining distributed cache, such as juicefs, alluxio, etc., but they need to be downloaded to the local file system in the GPU cluster in advance, or the GPU cluster locally provides a client file system similar to S3FS to provide POSIX semantic access.

[0039] The storage system used for data set reading and checkpoint data reading and writing in current GPU cluster training is large and complex, and most of its functions are not required by the three core IO business scenarios in the GPU cluster training scenario. Therefore, the embodiments of the present disclosure provide a distributed file system adapted to GPU cluster training, which can adapt to the three core IO business scenarios in the GPU cluster training scenario, and has simple operation and maintenance, saving costs while improving performance. The following is a detailed description of the distributed file system adapted to GPU cluster training provided by the embodiments of the present disclosure.

[0040] Figure 1 A block diagram of a distributed file system adapted for GPU cluster training according to an embodiment of the present disclosure is shown. Figure 1 As shown, the system includes: a management module, a server and a client.

[0041] A management module is used to create a server for multiple first computer nodes included in the target GPU cluster and create a client for multiple second computer nodes included in the target GPU cluster after allocating a target GPU cluster for a target training task, wherein at least one hardware GPU is deployed on each computer node included in the target GPU cluster.

[0042] A GPU cluster is a computer cluster that includes multiple computer nodes. Each computer node is not only equipped with at least one hardware GPU, but also with single-core CPU, multi-core CPU, or even multi-CPU computing resources. A GPU cluster uses the computing power of hardware GPUs to perform very fast computing tasks.

[0043] For the target training task, a target GPU cluster can be allocated for the target training task based on the Kubernetes platform according to the task requirements. The number of computer nodes included in the target GPU cluster depends on the task requirements of the target training task, and this disclosure does not specifically limit this.

[0044] After allocating a target GPU cluster for a target training task and before starting to execute the target training task, the management module may create a server for a plurality of first computer nodes included in the target GPU cluster according to actual IO requirements of the target training task.

[0045] The management module can reuse the operator mechanism for cluster management in the Kubernetes platform to create and distribute the server to multiple first computer nodes in the target GPU cluster.

[0046] The server is used to build a GPU cluster distributed file system adapted to the target training task by using the preset local memory space deployed on the corresponding first computer node.

[0047] Multiple servers can communicate through a remote direct memory access (RDMA) network, so that the preset local memory space deployed on multiple first computer nodes is used as the file system backend to jointly build a GPU cluster distributed file system adapted to the target training task.

[0048] The number of first computer nodes can be determined according to the actual IO requirements of the target training task and the size of the preset local memory space deployed on each first computer node, as long as the total size of the preset local memory space deployed on all first computer nodes, that is, the storage space size of the GPU cluster distributed file system constructed above, can meet the actual IO requirements of the target training task. The first computer nodes can be all computer nodes in the target GPU cluster, or only some computer nodes in the target GPU cluster, and the present disclosure does not make specific limitations on this.

[0049] like Figure 1 As shown, only some of the computer nodes in the target GPU cluster are used as first computer nodes and are deployed with a server.

[0050] In a possible implementation manner, the memory size of the preset local memory space deployed on each first computer node is the same.

[0051] In order to ensure the balance of the GPU cluster distributed file system constructed above, the memory of the preset local memory space deployed on each first computer node can be set to the same size.

[0052] In one example, the preset local memory space deployed on the first computer node can be a local NVME (Non-Volatile Memory Express) disk space deployed on the first computer node, or a shared memory space deployed on the first computer node (for example, shared memory / dev / shm in the Linux operating system), or other local memory space deployed on the first computer node. The embodiments of the present disclosure do not specifically limit this.

[0053] For example, a target GPU cluster is assigned to a target training task, and the target GPU cluster includes N computer nodes. Before the target training task is started, the management module creates a server for each computer node included in the target GPU cluster, that is, each computer node included in the target GPU cluster is used as a first computer node. The size of the preset local memory space deployed on each first computer node is 15.36TB. At this time, the N servers can communicate through the RDMA network and jointly build a GPU cluster distributed file system that is adapted to the target training task and has a storage space size of 15.36TB×N.

[0054] The management module is further used to create clients for multiple second computer nodes included in the target GPU cluster.

[0055] The number of second computer nodes can be determined according to the actual data reading requirements of each computer node in the target GPU cluster. In the process of executing the target training task in the target GPU cluster, each computer node in the target GPU cluster executes a different training process. According to the training process executed by each computer node, it can be determined whether the computer node has a data reading requirement. The management module can create clients only for multiple second computer nodes that have data reading requirements in the target GPU cluster, without creating clients for computer nodes that do not have data reading requirements in the target GPU cluster. The second computer nodes can be all computer nodes in the target GPU cluster, or only some computer nodes in the target GPU cluster, and the present disclosure does not make specific limitations on this.

[0056] like Figure 1 As shown, only some of the computer nodes in the target GPU cluster are used as second computer nodes and are deployed with a server.

[0057] In addition, some computer nodes in the target GPU cluster can be used as both the first computer node and the second computer node, that is, both the server and the client are deployed in these computer nodes.

[0058] like Figure 1 As shown, both the server and the client are deployed on some of the computer nodes included in the target GPU cluster.

[0059] The client is used to perform IO operations on the GPU cluster distributed file system when using the target GPU cluster to execute the target training task.

[0060] Integrate the client into the target training task entry so that the client can access the GPU cluster distributed file system built to adapt to the target training task. Then, in the process of using the target GPU cluster to execute the target training task, the client can perform IO operations on the GPU cluster distributed file system.

[0061] In a possible implementation, a GPU cluster distributed file system is used to store data sets required in a target training task and checkpoint data during the training process.

[0062] Since the GPU cluster distributed file system can store the data sets required for the target training task and the checkpoint data during the training process, the client can perform IO operations such as data set reading, checkpoint data writing, and checkpoint data reading on the GPU cluster distributed file system, adapting to the three core IO business scenarios in the GPU cluster training scenario.

[0063] In an embodiment of the present disclosure, a distributed file system adapted to GPU cluster training is designed, including: a management module, a server, and a client; after allocating a target GPU cluster for a target training task, the management module creates a server for multiple first computer nodes included in the target GPU cluster and creates a client for multiple second computer nodes included in the target GPU cluster according to the data reading and writing requirements of the target training task, so that the server can use the preset local memory space deployed on the corresponding first computer node to flexibly build a GPU cluster distributed file system adapted to the target training task, and then, in the process of using the target GPU cluster to execute the target training task, the client can perform IO operations on the GPU cluster distributed file system to effectively meet the IO requirements in the GPU cluster training scenario.

[0064] In one possible implementation, the client includes: a system call interception module and a client communication module; the server includes: a server communication module; a system call interception module, which is used to intercept the target file system call in the second computer node where the client is located, and convert the target file system call into an IO request, wherein the target file system call is used to access the GPU cluster distributed file system; and a client communication module, which is used to send the IO request to the server communication module.

[0065] Figure 2 A schematic diagram of the interaction between the client and the server according to an embodiment of the present disclosure is shown. Figure 2 As shown, each client includes a system call interception module.

[0066] When the second computer node where the client is located has a data reading demand during the training process, it will initiate a target file system call to the GPU cluster distributed file system constructed above. The system call interception module included in the client can intercept the target file system call, and then, through random hashing, the target file system call can be converted into an IO request, and the client communication module can be used to send the IO request to the server communication module included in the server.

[0067] In one possible implementation, the client communication module includes: a remote procedure call (RPC) client module; the server communication module includes: an RPC server module; an RPC client module, which is used to adopt a non-blocking RPC network communication mode to send IO requests to the RPC server module.

[0068] The client and server can communicate across computer nodes through a high-performance RPC network. In addition, in order to further improve performance, the client and server use a non-blocking RPC network communication mode for communication.

[0069] The non-blocking RPC network communication mode means that after the RPC client module initiates a remote call to the RPC server module, it will not wait for the remote call to complete, but will immediately return and continue to execute subsequent code. When the remote call process is completed, the RPC client will obtain the remote call result in some way (for example, callback function, event notification, etc.).

[0070] The RPC client module in the embodiment of the present disclosure can communicate with any RPC server module using a non-blocking RPC network communication mode.

[0071] like Figure 2 As shown, each client includes an RPC client module, and each server includes an RPC server module. Each RPC client module can communicate with each RPC server module based on a high-performance network (for example, an RPC network).

[0072] In one possible implementation, the client communication module includes: an IPC client module; the server communication module includes: an IPC server module; an IPC client module, used to send an IO request to the IPC server module when the first computing point where the server is located and the second computer node where the client is located are the same computer node.

[0073] When the client and the server are located in the same computer node, in order to improve performance, inter-process communication (IPC) can be used within the computer node to communicate.

[0074] like Figure 2 As shown, the client includes an IPC client module, and the server includes an IPC server module. For the client and server inside the same computer node, the internal communication of the computer node can be performed based on the IPC client module and the IPC server module.

[0075] In a possible implementation, the client includes: a file map; the file map is used to store the access record of the client to the GPU cluster distributed file system.

[0076] The file graph in the client stores the client's access records to the GPU cluster distributed file system, for example, records which files of the GPU cluster distributed file system the client currently accesses.

[0077] In one example, the data stored in the file map may be a map data structure.

[0078] In the embodiment of the present disclosure, the high-performance network is abstracted and divided into an RPC layer and a bull layer. Figure 3 FIG. 2 shows a schematic diagram of the transmission of metadata information and data between a client and a server according to an embodiment of the present disclosure. Figure 3 As shown in the figure, the client and the server use the RPC layer to transmit metadata information point-to-point, and use the bulk layer to transmit actual data. The bulk layer can use RDMA to transmit actual data. The bulk layer uses RDMA to transmit data

[0079] In one example, the underlying implementation of RPC transmission can be to call the OFED interface through the Fabric API, call the Verbs or TCP interface, and finally transmit data through a high-performance network card driver. The specific transmission process can refer to the RPC transmission process in the relevant technology, and the embodiments of the present disclosure do not specifically limit this.

[0080] In one example, data transmission is performed based on the RPC layer, and a non-blocking RPC communication mode can be used to improve data transmission performance.

[0081] In one example, a thread pool may be pre-built, and then threads in the thread pool may be used to implement data transmission between the client and the server to improve data transmission performance.

[0082] The following is a detailed description of the data transmission between the client and the server.

[0083] In a possible implementation, the IO request includes: a metadata IO request; the client includes: a data segmentation module; the server includes: a metadata database; the data segmentation module is used to divide the target data that needs to be written to the GPU cluster distributed file system into blocks, determine multiple data blocks corresponding to the target data, and metadata information of each data block, wherein the metadata information corresponding to each data block is used to indicate the target server to which the data block needs to be written; the client communication module is used to send the metadata IO request to each server communication module, wherein the metadata IO request is used to request that the metadata information of each data block corresponding to the target data be written to each server; the metadata database in each server is used to respond to the metadata IO request and store the metadata information of each data block corresponding to the target data.

[0084] In the disclosed embodiment, the GPU cluster distributed file system is a fully distributed system, that is, each client can communicate with each server in the GPU cluster distributed file system, so that each client can independently parse each server in the GPU cluster distributed file system without the need for a central data structure that can track metadata information or the location of data.

[0085] In order to achieve balanced data distribution of large files, before requesting to write the target data to the GPU cluster distributed file system for storage, the client can use the data segmentation module to segment the target data into multiple data blocks of the same size and determine the metadata information of each data block. The metadata information corresponding to each data block is used to indicate the target server to which the data block needs to be written.

[0086] like Figure 2 As shown, each server includes a metadata database for storing metadata information.

[0087] In a possible implementation, the number of designated data blocks corresponding to the target data written to each target server is the same.

[0088] In order to achieve balanced data distribution for large files, the number of designated data blocks corresponding to the target data written to each target server is the same.

[0089] In one example, for any client, all data blocks corresponding to the target data may be evenly distributed and stored on all servers, or all data blocks corresponding to the target data may be evenly distributed and stored on some servers.

[0090] For example, the GPU cluster distributed file system corresponds to 10 servers, and the client stores all data blocks corresponding to the target data in a balanced distributed manner on all servers, that is, all 10 servers are target clients. Furthermore, the client can use the data segmentation module to segment the target data into 20 data blocks of the same size, and then store 2 data blocks in each server.

[0091] For example, the GPU cluster distributed file system corresponds to 10 servers, and the client stores all data blocks corresponding to the target data in a balanced distributed manner on 8 servers, that is, 8 of the 10 servers are selected (for example, randomly selected, specified based on configuration information, etc.) as target clients. Furthermore, the client can use the data segmentation module to segment the target data into 8 data blocks of the same size, and subsequently store 1 data block in each server.

[0092] After determining the metadata information of each data block obtained by segmenting the target data, the client can generate a metadata IO request, and use the client communication module and the server communication module to send the metadata IO request to each server to request that the metadata information of each data block corresponding to the target data be written to each server. In other words, the metadata information of each data block corresponding to the target data is fully stored in each server in the GPU cluster distributed file system, so that the GPU cluster distributed file system becomes a fully distributed system.

[0093] Since the GPU cluster distributed file system provided by the embodiment of the present disclosure can only support three core IO business scenarios in the GPU cluster training scenario, the size of metadata information can be reduced, thereby achieving high-performance processing of metadata information in scenarios with hundreds or even thousands of computer nodes.

[0094] In a possible implementation, the IO request includes: a data IO request; the server includes: a persistence module; a communication client module, which is used to send data IO requests to multiple target servers, wherein the data IO request corresponding to each target server is used to request that a specified data block corresponding to the target data be written to the target server; the persistence module in each target server is used to respond to the received data IO request, and according to the metadata IO request stored in the metadata database in the target server, store the specified data block that needs to be written to the target server in the preset local memory space corresponding to the target server.

[0095] After the metadata information of each data block corresponding to the target data is fully stored in each server in the GPU cluster distributed file system, the client generates a data IO request according to the metadata information of each data block corresponding to the target data, and uses the client communication module and the server communication module to send the data IO request to each target server, so that according to the metadata IO request stored in the metadata database in each target server, the specified data block that needs to be written to the target server is persistently stored in the preset local memory space corresponding to the target server using the persistence module.

[0096] Figure 4 A schematic diagram showing a client writing a file to a GPU cluster distributed file system according to an embodiment of the present disclosure. Figure 4 As shown, when the training process in the computer node where the client is located needs to write a file to the GPU cluster distributed file system, first, the target data is written to the cache in the client, and then the target data is divided into 6 data blocks of the same size using the data segmentation module: data block 0 to data block 5. Then, data block 0 and data block 5 are packaged and sent to the preset local memory space in server 1 through thread 1, data block 2 and data block 4 are packaged and sent to the preset local memory space in server 2 through thread 2, and data block 1 and data block 3 are packaged and sent to the preset local memory space in server 3 through thread 3.

[0097] In a possible implementation, the data IO request corresponding to each target server includes client memory area information corresponding to the specified data block that needs to be written to the target server; the persistence module in each target server reads the specified data block from the memory space specified by the client memory area information through RDMA based on the client memory area information included in the received data IO request.

[0098] If supported by the underlying network structure protocol, the client can disclose the client memory area information corresponding to the specified data block that needs to be written to the target server to the target server through a data IO request, so that the target server can read the specified data block directly from the memory space specified by the client memory area information through RDMA, thereby improving data transmission performance.

[0099] In a possible implementation, the persistence module in each target server needs to write each specified data block of the target server and store it in a shard chunk file in a preset local memory space corresponding to the target server.

[0100] The persistence module in each target server will need to write each specified data block of the target server to a shard chunk file in the preset local memory space corresponding to the target server, in the underlying node structure domain storage.

[0101] In a possible implementation, the GPU cluster distributed file system is also used to asynchronously send all or part of the checkpoint data to a parallel file system (Parallel File Storage, PFS) outside the target GPU cluster.

[0102] In actual use, the GPU cluster distributed file system can be used only for temporary data storage. During or after the execution of the target training task, all or part of the checkpoint data temporarily stored in the GPU cluster distributed file system can be asynchronously sent to the PFS outside the target GPU cluster (for example, GPFS, Lustre, etc.) to prevent these data from being affected by the failure of the target GPU cluster.

[0103] In a possible implementation, the management module is used to delete the client and the server after the target training task is completed using the target GPU cluster.

[0104] After the target GPU cluster is used to execute the target training task, the management module can delete the client and the server to release resources.

[0105] In one example, the management module can reuse the operator mechanism for cluster management in the Kubernetes platform to create and delete servers on computer nodes according to target training tasks.

[0106] In one example, the management module is also used to store the configuration information of each server and the monitoring information during the training process to ensure the smooth execution of the target training task based on the target GPU cluster.

[0107] In an embodiment of the present disclosure, a distributed file system adapted to GPU cluster training is designed, including: a management module, a server, and a client; after allocating a target GPU cluster for a target training task, the management module creates a server for multiple first computer nodes included in the target GPU cluster and creates a client for multiple second computer nodes included in the target GPU cluster according to the data reading and writing requirements of the target training task, so that the server can use the preset local memory space deployed on the corresponding first computer node to flexibly build a GPU cluster distributed file system adapted to the target training task, and then, in the process of using the target GPU cluster to execute the target training task, the client can perform IO operations on the GPU cluster distributed file system to effectively meet the IO requirements in the GPU cluster training scenario.

[0108] Compared with the use of main memory or traditional parallel file systems such as localfs, the distributed file system adapted for GPU cluster training provided by the embodiment of the present disclosure can provide a unified storage space, as well as a larger capacity and a linearly growing throughput. Moreover, the distributed file system adapted for GPU cluster training provided by the embodiment of the present disclosure can only adapt to the file system calls in the GPU cluster training scenario, abandoning the traditional file system and other capabilities of the traditional parallel file system, simplifying the call chain, so that the GPU cluster training can achieve more than 6-7 times the performance, while the operation and maintenance is very simple. In addition, since the distributed file system adapted for GPU cluster training provided by the embodiment of the present disclosure can support the shared memory on each computer node in the GPU cluster as the back-end storage medium, more than 10 times the performance can be achieved.

[0109] The distributed file system adapted to GPU cluster training provided by the embodiments of the present disclosure can flexibly create and integrate the deployment of a GPU cluster distributed file system that meets the task requirements as training tasks are started and deleted, making full use of the CPU, storage, PCIE bandwidth and network resources of the computing nodes in the GPU cluster, saving costs while improving performance.

[0110] Figure 5 A flow chart of a method for creating a distributed file system adapted for GPU cluster training according to an embodiment of the present disclosure is shown. Figure 5 As shown, the method includes:

[0111] In step S51, after allocating a target GPU cluster for a target training task, a server is created for multiple first computer nodes included in the target GPU cluster, and a client is created for multiple second computer nodes included in the target GPU cluster, wherein at least one hardware GPU is deployed on each computer node included in the target GPU cluster.

[0112] In step S52, based on the server, a GPU cluster distributed file system adapted to the target training task is constructed using the preset local memory space deployed on the corresponding first computer node.

[0113] In step S53, based on the client, in the process of using the target GPU cluster to execute the target training task, an IO operation is performed on the GPU cluster distributed file system.

[0114] For the specific process of the client performing IO operations on the GPU cluster distributed file system, please refer to the above Figures 1 to 5 The relevant description of the illustrated embodiment is not repeated here.

[0115] It can be understood that the above-mentioned various method embodiments mentioned in the present disclosure can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present disclosure will not repeat them. It can be understood by those skilled in the art that in the above-mentioned method of the specific implementation method, the specific execution order of each step should be determined according to its function and possible internal logic.

[0116] In addition, the present disclosure also provides electronic devices, computer-readable storage media, and programs, all of which can be used to implement any method provided by the present disclosure. The corresponding technical solutions and descriptions are referred to the corresponding records in the method part and will not be repeated here.

[0117] This method has a specific technical connection with the internal structure of the computer system, and can solve the technical problem of how to improve the hardware computing efficiency or execution effect (including reducing the amount of data storage, reducing the amount of data transmission, increasing the hardware processing speed, etc.), thereby obtaining the technical effect of improving the internal performance of the computer system in accordance with the laws of nature.

[0118] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0119] The embodiment of the present disclosure also provides a computer-readable storage medium on which computer program instructions are stored, and the computer program instructions implement the above method when executed by a processor. The computer-readable storage medium can be a volatile or non-volatile computer-readable storage medium.

[0120] An embodiment of the present disclosure further proposes an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above method.

[0121] The embodiments of the present disclosure also provide a computer program product, including a computer-readable code, or a non-volatile computer-readable storage medium carrying the computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above method.

[0122] The electronic device may be provided as a terminal, a server, or a device in other forms.

[0123] Figure 6 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. Figure 6 , the electronic device 1900 can be provided as a server or a terminal device. Figure 6 , the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions executable by the processing component 1922, such as an application. The application stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above method.

[0124] The electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output interface 1958. The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as a Microsoft Server operating system (Windows Server 2003). TM ), a graphical user interface operating system launched by Apple (Mac OS X TM ), a multi-user, multi-process computer operating system (Unix TM ), a free and open source Unix-like operating system (Linux TM ), an open source Unix-like operating system (FreeBSD TM ) or similar.

[0125] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to perform the above method.

[0126] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0127] Computer readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. Computer readable storage medium can be, for example, (but not limited to) an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a convex structure in a groove on which instructions are stored, and any suitable combination thereof. The computer readable storage medium used here is not interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated by a waveguide or other transmission medium (for example, a light pulse by an optical fiber cable), or an electrical signal transmitted by a wire.

[0128] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0129] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0130] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0131] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0132] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0133] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.

[0134] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (SDK) and the like.

[0135] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0136] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0137] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

[0138] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A distributed file system adapted for GPU cluster training, characterized in that: include: Management modules, servers and clients; The management module is used to create a server for a plurality of first computer nodes included in the target GPU cluster and a client for a plurality of second computer nodes included in the target GPU cluster after allocating a target GPU cluster for a target training task, wherein at least one hardware GPU is deployed on each computer node included in the target GPU cluster; The server is used to construct a GPU cluster distributed file system adapted to the target training task by using the preset local memory space deployed on the corresponding first computer node; The client is used to perform input and output IO operations on the GPU cluster distributed file system during the process of using the target GPU cluster to execute the target training task.

2. The system according to claim 1, characterized in that The memory size of the preset local memory space deployed on each first computer node is the same.

3. The system according to claim 1, characterized in that The GPU cluster distributed file system is used to store the data set required in the target training task and the checkpoint data during the training process.

4. The system according to claim 1, characterized in that The client includes: a system call interception module and a client communication module; the server includes: a server communication module; The system call interception module is used to intercept a target file system call in the second computer node where the client is located, and convert the target file system call into an IO request, wherein the target file system call is used to access the GPU cluster distributed file system; The client communication module is used to send the IO request to the server communication module.

5. The system according to claim 4, characterized in that The client communication module includes: a remote procedure call (RPC) client module; the server communication module includes: an RPC server module; The RPC client module is used to send the IO request to the RPC server module by adopting a non-blocking RPC network communication mode.

6. The system according to claim 4, characterized in that The client communication module includes: an IPC client module; the server communication module includes: an IPC server module; The IPC client module is used to send the IO request to the IPC server module when the first computing point where the server is located and the second computer node where the client is located are the same computer node.

7. The system according to any one of claims 1 to 6, characterized in that: The client includes: a file map; The file map is used to store the access record of the client to the GPU cluster distributed file system.

8. The system according to claim 4, characterized in that The IO request includes: a metadata IO request; the client includes: a data segmentation module; the server includes: a metadata database; The data segmentation module is used to divide the target data to be written into the GPU cluster distributed file system into blocks, determine multiple data blocks corresponding to the target data, and metadata information of each data block, wherein the metadata information corresponding to each data block is used to indicate the target server to which the data block needs to be written; The client communication module is used to send the metadata IO request to each server communication module, wherein the metadata IO request is used to request to write metadata information of each data block corresponding to the target data into each server; The metadata database in each server is used to respond to the metadata IO request and store metadata information of each data block corresponding to the target data.

9. The system according to claim 8, characterized in that The IO request includes: a data IO request; the server includes: a persistence module; The communication client module is used to send data IO requests to multiple target service terminals, wherein the data IO request corresponding to each target service terminal is used to request that a specified data block corresponding to the target data be written into the target service terminal; The persistence module in each target server is used to respond to the received data IO request and store the specified data block that needs to be written to the target server in the preset local memory space corresponding to the target server according to the metadata IO request stored in the metadata database in the target server.

10. The system according to claim 9, characterized in that The persistence module in each target server will store each specified data block that needs to be written to the target server in a shard chunk file in the preset local memory space corresponding to the target server.

11. The system according to any one of claims 8 to 10, characterized in that: The number of designated data blocks corresponding to the target data written to each target server is the same.

12. The system according to claim 9 or 10, characterized in that The data IO request corresponding to each target server includes the client memory area information corresponding to the specified data block to be written to the target server; The persistence module in each target server reads a specified data block from the memory space specified by the client memory area information through a remote direct memory access (RDMA) mode based on the client memory area information included in the received data IO request.

13. The system according to claim 3, characterized in that The GPU cluster distributed file system is also used to asynchronously send all or part of the checkpoint data to a parallel file system PFS outside the target GPU cluster.

14. The system according to claim 1, characterized in that The management module is used to delete the client and the server after the target training task is completed using the target GPU cluster.

15. A method for creating a distributed file system adapted for GPU cluster training, characterized in that: include: After allocating a target GPU cluster for a target training task, creating a server for a plurality of first computer nodes included in the target GPU cluster, and creating a client for a plurality of second computer nodes included in the target GPU cluster, wherein at least one hardware GPU is deployed on each computer node included in the target GPU cluster; Based on the server, a GPU cluster distributed file system adapted to the target training task is constructed by using a preset local memory space deployed on the corresponding first computer node; Based on the client, in the process of using the target GPU cluster to execute the target training task, an IO operation is performed on the GPU cluster distributed file system.

16. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; Wherein, the processor is configured to call the instructions stored in the memory to execute the method of claim 15.

17. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method of claim 15 is implemented.

Images