Intelligent training simulation transaction acceleration method and apparatus for data center, device, and medium

Abstract

The present invention relates to the technical field of electrical digital data processing, and in particular to an intelligent training simulation transaction acceleration method and apparatus for a data center, a device, and a medium. The method comprises: acquiring the current computing power of a CPU and the current computing power of GPUs to be simulated; determining the current distributed simulation strategy on the basis of the current computing power of the CPU and the current computing power of said GPUs, and when simulation is carried out on the basis of the current distributed simulation strategy, determining whether a deterministic simulation transaction is present in a transaction queue in the current distributed simulation strategy; and if a deterministic simulation transaction is present, deleting the deterministic simulation transaction, and carrying out accelerated simulation on the basis of the remaining simulation transactions. In this way, the problems in the related art of low efficiency and long duration in simulation systems are solved; and by means of the structures of large model clusters and the features of model training, distributed simulation of multiple CPU cores is employed to reduce the number of transactions of the entire simulation system, improving the simulation efficiency.

Description

Data center intelligent training simulation transaction acceleration method, device, equipment and medium

[0001] Cross-reference of Related Applications

[0002] The present application is based on the Chinese patent application with the application number 202410466627.X and the application date of April 18, 2024, and claims the priority of the Chinese patent application, the entire contents of which are hereby incorporated by reference into the present application. TECHNICAL FIELD

[0003] The present application relates to the technical field of digital data processing, in particular to a data center intelligent training simulation transaction acceleration method, device, equipment and medium. BACKGROUND

[0004] With the rapid development of cloud computing, big data analysis and artificial intelligence technologies, data centers play a crucial role in modern society, not only providing flexible computing and storage resources for enterprises, but also supporting the development of innovative applications and services. At the same time, with the rise of the AI (Artificial Intelligence) field, the scale of model training is gradually increasing, and due to the excessive deployment overhead, it is necessary to evaluate the training overhead and verify the design and optimization strategies of each layer in model training under lower device overhead. Therefore, a simulation system for intelligent training is necessary.

[0005] In related technologies, NS-3 is one of the most commonly used network simulators, but it is subject to the simulation efficiency of the NS-3 simulator itself. The deep learning (large model) training simulation system using NS-3 as the underlying simulator is usually inefficient, and such simulation systems rely too much on network simulators. In the case of a large number of simulation transactions, it is usually impossible to complete the simulation task in a short time, and this problem needs to be solved urgently. SUMMARY

[0006] The present application provides a data center intelligent training simulation transaction acceleration method, device, equipment and medium to solve the problem of low efficiency of the simulation system in related technologies, long time consumption, etc.

[0007] The first aspect of the present application provides a data center intelligent training simulation transaction acceleration method, comprising the following steps:

[0008] Obtaining the current CPU (Central Processing Unit) computing power and the current GPU (Graphics Processing Unit) computing power to be simulated;

[0009] determine a current distributed simulation strategy according to the current CPU computing power and the current GPU computing power to be simulated, and determine whether there is a deterministic simulation transaction in a transaction queue in the current distributed simulation strategy when simulating based on the current distributed simulation strategy;

[0010] If the deterministic simulation transaction exists, the deterministic simulation transaction is deleted, and accelerated simulation is performed based on the remaining simulation transactions.

[0011] According to an embodiment of the present application, the current distributed simulation strategy is determined according to the current CPU computing power and the current GPU computing power to be simulated, comprising:

[0012] determining whether the current CPU computing power is greater than the current GPU computing power to be simulated;

[0013] If the current CPU computing power is greater than the current GPU computing power to be simulated, a first inter-core communication amount using all CPU cores on average and a second inter-core communication amount using CPU cores meeting a preset quantity requirement are calculated;

[0014] If the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed simulation strategy is to use the CPU cores meeting the preset quantity requirement for simulation, otherwise, the current distributed simulation strategy is to use all CPU cores on average for simulation.

[0015] According to an embodiment of the present application, after determining whether the current CPU computing power is greater than the current GPU computing power to be simulated, further comprising:

[0016] If the current CPU computing power is less than or equal to the current GPU computing power to be simulated, a third inter-core communication amount of a single CPU core simulating all GPUs and a fourth inter-core communication amount of each CPU core corresponding to simulating the same type of GPU are calculated;

[0017] If the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed simulation strategy is to use a single CPU core to simulate all the current GPUs to be simulated, otherwise, the current distributed simulation strategy is to use each CPU core to simulate the same type of GPU.

[0018] According to an embodiment of the present application, the determination of whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy comprises:

[0019] obtaining a plurality of simulation transactions of the transaction queue in the current distributed simulation strategy;

[0020] determining whether there is an association between at least two simulation transactions in the plurality of simulation transactions.

[0021] if there is no correlation between the at least two simulation transactions, determining that the transaction queue in the current distributed simulation strategy has the deterministic simulation transaction.

[0022] According to an embodiment of the present application, after determining that the transaction queue in the current distributed simulation strategy has the deterministic simulation transaction, further comprising:

[0023] determining whether other target transactions are inserted into the transaction queue;

[0024] if the other target transactions are not inserted, simulating the last simulation transaction in the transaction queue.

[0025] The data center intelligent training simulation transaction acceleration method according to the embodiment of the present application acquires current CPU computing power and current to-be-simulated GPU computing power, determines a current distributed simulation strategy according to the current CPU computing power and the current to-be-simulated GPU computing power, and determines whether a deterministic simulation transaction exists in a transaction queue in the current distributed simulation strategy when simulation is performed based on the current distributed simulation strategy. If the deterministic simulation transaction exists, the deterministic simulation transaction is deleted, and accelerated simulation is performed based on remaining simulation transactions. In this way, the problems of low simulation system efficiency and long time consumption in the related art are solved. Through the structure of a large model cluster and the characteristics of model training, distributed simulation of multiple CPU cores is used to reduce the number of transactions of the entire simulation system, thereby improving simulation efficiency.

[0026] The second aspect embodiment of the present application provides a data center intelligent training simulation transaction acceleration device, comprising:

[0027] an acquisition module, configured to acquire current CPU computing power and current to-be-simulated GPU computing power;

[0028] a determination module, configured to determine a current distributed simulation strategy according to the current CPU computing power and the current to-be-simulated GPU computing power, and determine whether a deterministic simulation transaction exists in a transaction queue in the current distributed simulation strategy when simulation is performed based on the current distributed simulation strategy;

[0029] a simulation module, configured to delete the deterministic simulation transaction if the deterministic simulation transaction exists, and perform accelerated simulation based on remaining simulation transactions.

[0030] According to an embodiment of the present application, the determination module is specifically configured to:

[0031] determine whether the current CPU computing power is greater than the current to-be-simulated GPU computing power;

[0032] if the current CPU computing power is greater than the current GPU computing power to be emulated, calculating a first inter-core communication amount using all CPU cores and a second inter-core communication amount using CPU cores satisfying a preset number requirement;

[0033] if the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed emulation strategy is to use the CPU cores satisfying the preset number requirement for emulation, otherwise, the current distributed emulation strategy is to use all the CPU cores for emulation.

[0034] According to one embodiment of the present application, after judging whether the current CPU computing power is greater than the current GPU computing power to be emulated, the judging module is further used for:

[0035] if the current CPU computing power is less than or equal to the current GPU computing power to be emulated, calculating a third inter-core communication amount of a single CPU core emulating all GPUs and a fourth inter-core communication amount of each CPU core emulating the same type of GPU;

[0036] if the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed emulation strategy is to use a single CPU core to emulate all the current GPUs to be emulated, otherwise, the current distributed emulation strategy is to use each CPU core to emulate the same type of GPU.

[0037] According to one embodiment of the present application, the judging module is specifically used for:

[0038] obtaining a plurality of emulation transactions of a transaction queue in the current distributed emulation strategy;

[0039] judging whether there is an association between at least two emulation transactions in the plurality of emulation transactions;

[0040] if there is no association between the at least two emulation transactions, it is determined that the transaction queue in the current distributed emulation strategy has the deterministic emulation transaction.

[0041] According to one embodiment of the present application, after determining that the transaction queue in the current distributed emulation strategy has the deterministic emulation transaction, the judging module is further used for:

[0042] judging whether other target transactions are inserted into the transaction queue;

[0043] if the other target transactions are not inserted, the last emulation transaction in the transaction queue is emulated.

[0044] According to the data center intelligent training simulation transaction acceleration device provided in the embodiments of the present application, the current CPU computing power and the current to-be-simulated GPU computing power are obtained; the current distributed simulation strategy is determined according to the current CPU computing power and the current to-be-simulated GPU computing power, and when simulation is performed based on the current distributed simulation strategy, it is judged whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy, if there is a deterministic simulation transaction, the deterministic simulation transaction is deleted, and the remaining simulation transactions are used for accelerated simulation. In this way, the problems of low simulation system efficiency and long time consumption of related technologies are solved, and through the structure of the large model cluster and the characteristics of the model training, the distributed simulation of the multi-CPU core is used to reduce the number of transactions of the entire simulation system, thereby improving the simulation efficiency.

[0045] The third aspect of the present application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the data center intelligent training simulation transaction acceleration method as described in the above embodiments.

[0046] The fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the data center intelligent training simulation transaction acceleration method as described in the above embodiments.

[0047] The fifth aspect of the present application provides a computer program product comprising computer programs / instructions, which, when executed by a processor, implement the steps of the method as described in the above embodiments.

[0048] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0049] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the appended drawings.

[0050] Fig. 1 is a flowchart of a data center intelligent training simulation transaction acceleration method according to an embodiment of the present application;

[0051] Fig. 2 is a schematic diagram of a multi-core parallel division module structure according to an embodiment of the present application;

[0052] Fig. 3 is a schematic diagram of a deterministic transaction acceleration module structure according to an embodiment of the present application;

[0053] Fig. 4 is a block diagram of a data center intelligent training simulation transaction acceleration device according to an embodiment of the present application;

[0054] FIG. 5 is a structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which like or similar elements are denoted by the same or similar reference signs, and examples of embodiments are shown in the drawings. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0056] A data center intelligent training simulation transaction acceleration method according to an embodiment of the present application is described below with reference to the accompanying drawings. In view of the low efficiency and long time consumption of the simulation system of the related technology mentioned in the background art, the present application provides a data center intelligent training simulation transaction acceleration method. In the method, the current CPU computing power and the current GPU computing power to be simulated are obtained. The current distributed simulation strategy is determined according to the current CPU computing power and the current GPU computing power to be simulated. When simulation is performed based on the current distributed simulation strategy, it is determined whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy. If there is a deterministic simulation transaction, the deterministic simulation transaction is deleted, and accelerated simulation is performed based on the remaining simulation transactions. Thus, the problems of low efficiency and long time consumption of the simulation system of the related technology are solved. Through the structure of the large model cluster and the characteristics of model training, the distributed simulation of multiple CPU cores is used to reduce the number of transactions of the entire simulation system, thereby improving the simulation efficiency.

[0057] Specifically, FIG. 1 is a flowchart of a data center intelligent training simulation transaction acceleration method according to an embodiment of the present application.

[0058] As shown in FIG. 1, the data center intelligent training simulation transaction acceleration method includes the following steps:

[0059] In step S101, the current CPU computing power and the current GPU computing power to be simulated are obtained.

[0060] Specifically, in the embodiment of the present application, in order to solve the problem that the training of a large number of models in the related technology consumes a large amount of GPU resources, thereby reducing the simulation efficiency and consuming a long simulation time and cost, the distributed simulation method of multiple CPU cores is fully used to reduce the number of transactions of the simulation system, thereby realizing accelerated simulation.

[0061] Specifically, as shown in FIG. 2, in the process of implementing accelerated simulation, first, the simulation system (for example, the NS-3 simulator) receives the function configurations such as the cluster configuration of model training, the model structure configuration, the training data volume configuration, and the network configuration; second, after the simulation system loads various configuration parameters, the architecture and model training behavior of the AI (Artificial Intelligence) cluster are determined, at this time, the multi-core parallel division module calculates the predictability of the simulation performed in the multi-core CPU of the physical machine running the simulation, and the multi-core parallel division module determines the number of CPU cores of the distributed simulation in combination with the physical machine configuration of the running simulation system and the complexity of the simulation, that is, the number of CPU cores of the distributed simulation is determined by obtaining the current CPU computing power and the current GPU computing power to be simulated, since the computing power of the current GPU to be simulated and the current CPU computing power of the physical server are known, the multi-core parallel division module can roughly estimate the number of GPU cores that can be simulated on a single physical CPU core.

[0062] In step S102, the current distributed simulation strategy is determined according to the current CPU computing power and the current GPU computing power to be simulated, and when simulation is performed based on the current distributed simulation strategy, it is judged whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy.

[0063] According to one embodiment of the present application, the current distributed simulation strategy is determined according to the current CPU computing power and the current GPU computing power to be simulated, which includes: judging whether the current CPU computing power is greater than the current GPU computing power to be simulated; if the current CPU computing power is greater than the current GPU computing power to be simulated, calculating the first inter-core communication amount using all the CPU cores and the second inter-core communication amount using the CPU cores meeting the preset quantity requirement; if the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed simulation strategy is to use the CPU cores meeting the preset quantity requirement for simulation, otherwise, the current distributed simulation strategy is to use all the CPU cores for simulation.

[0064] The preset quantity requirement can be set by a person skilled in the art according to actual simulation requirements, or can be set through a limited number of computer simulations, which is not specifically limited here.

[0065] Specifically, in an implementable manner, after the current CPU computing power and the current to-be-simulated GPU computing power are acquired, the multi-core parallel division module first calculates a first inter-core communication amount using all the CPU cores on average and a second inter-core communication amount using the CPU cores satisfying a preset quantity requirement, so as to evaluate and compare the performance gap therebetween, when the current CPU computing power owned is greater than the current to-be-simulated GPU computing power and the network transaction; secondly, if using all the CPU cores on average results in an excessively large inter-core communication amount, that is, the first inter-core communication amount is greater than the second inter-core communication amount, then the CPU cores satisfying the preset quantity requirement are selected for simulation based on the principle of minimum inter-core communication amount, otherwise, all the CPU cores are used for simulation.

[0066] For example, in the embodiment of the present application, if a 64-core physical machine is used to simulate an AI cluster training of 4 machines and 8 cards, compared with using a 64-core physical machine, the inter-core communication amount of using an 8-core physical machine will be much smaller than that of using a 64-core physical machine, therefore, a distributed simulation strategy of using one CPU core to run one GPU is selected.

[0067] According to an embodiment of the present application, after judging whether the current CPU computing power is greater than the current to-be-simulated GPU computing power, it further includes: if the current CPU computing power is less than or equal to the current to-be-simulated GPU computing power, calculating a third inter-core communication amount of simulating all the GPUs by a single CPU core and a fourth inter-core communication amount of simulating the same type of GPU corresponding to each CPU core; if the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed simulation strategy is to simulate all the current to-be-simulated GPUs by a single CPU core, otherwise, the current distributed simulation strategy is to simulate the same type of GPU corresponding to each CPU core.

[0068] Specifically, in another implementable manner, when the current CPU computing power owned is less than or equal to the current to-be-simulated GPU computing power and network transaction, the embodiment of the present application trains a large model based on the current to-be-simulated GPU communication features of the same machine or adjacent machines, and the multi-core parallel division module first simulates all the current to-be-simulated GPUs of a single machine on a single CPU core as much as possible, and calculates the third inter-core communication amount of the single CPU core simulating all the GPUs, but when the scale of the AI cluster is much larger than the scale of the server running the simulation system (for example, a 64-core CPU machine simulates a 256-machine 2048-GPU model training), the multi-core parallel division module can reasonably divide the entire simulation system according to the communication amount between different machines in the simulation parameter input file, rather than uniform division, that is, calculating the fourth inter-core communication amount of each CPU core corresponding to the simulation of the same type of GPU, for example, frequent collective communication occurs between 64 GPU cards of 8 machines, and at the same time, frequent collective communication occurs between 16 GPU cards of other 2 machines, so the 64 GPUs are simulated on one CPU core, and the 16 GPUs are simulated on another CPU core.

[0069] According to an embodiment of the present application, determining whether there is a deterministic simulation transaction in a transaction queue in a current distributed simulation strategy comprises: obtaining a plurality of simulation transactions of the transaction queue in the current distributed simulation strategy; determining whether there is an association between at least two simulation transactions in the plurality of simulation transactions; and if there is no association between the at least two simulation transactions, determining that there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy.

[0070] Specifically, the embodiment of the present application starts the deterministic transaction acceleration module when the simulation run starts. Since the core logic of the NS-3 simulator is that each functional module inserts a simulation transaction into the transaction queue through the Schedule function interface as the logical time rolls, the simulation transaction will be truly called and executed when the logical clock of the simulation system rolls to its set execution time. However, due to the structure and communication mode of the AI cluster where the large model is deployed, many communication transactions defined as "deterministic transactions" (deterministic simulation transactions) will be generated in the simulation run, that is, when a communication simulation transaction does not intersect in space and time with all other communication simulation transactions, that is, there is no association between at least two simulation transactions in the plurality of simulation transactions, it is determined that there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy.

[0071] In step S103, if there is a deterministic simulation transaction, the deterministic simulation transaction is deleted, and an accelerated simulation is performed based on the remaining simulation transactions.

[0072] According to one embodiment of the present application, after determining that there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy, further comprising: determining whether other target transactions are inserted in the transaction queue; if no other target transactions are inserted, the last simulation transaction in the simulation transaction queue.

[0073] Specifically, as shown in FIG. 3, when there is a deterministic simulation transaction in the transaction queue according to the embodiment of the present application, the deterministic simulation transaction is deleted through the Cancel interface provided by the NS-3 simulator, and it is determined whether other target transactions are inserted in the transaction queue. If no other target transactions are inserted, the last simulation transaction in the simulation transaction queue. That is, when a node continuously sends a long piece of data to another node, it often means that multiple simulation transactions are inserted into the transaction queue through the Schedule interface. If no other target simulation transactions are inserted between these simulation transactions, only the last simulation transaction symbolizing the completion of transmission can be retained, because in this case, since only simulation transactions belonging to the same deterministic transaction queue need to be executed, it means that the long data monopolizes the link, and whether to simulate the long simulation transaction does not affect the simulation result. Therefore, only the last simulation in the continuous transaction needs to be simulated.

[0074] It should be noted that the life cycle of the deterministic transaction acceleration module is equal to the entire cycle of simulation execution. As the simulation time advances, it is continuously determined whether there are continuous transactions from the same deterministic transaction queue. If there are, it means that there are a number of deterministic simulation transactions in the transaction queue, and the deterministic transactions are deleted through the Cancel interface provided by the NS-3 simulator, thereby reducing the number of transactions of the entire simulation system, thereby achieving the purpose of accelerating simulation.

[0075] According to the data center intelligent training simulation transaction acceleration method according to the embodiment of the present application, the current CPU computing power and the current to-be-simulated GPU computing power are obtained; the current distributed simulation strategy is determined according to the current CPU computing power and the current to-be-simulated GPU computing power, and when simulation is performed based on the current distributed simulation strategy, it is determined whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy. If there is a deterministic simulation transaction, the deterministic simulation transaction is deleted, and the remaining simulation transaction is accelerated based on the remaining simulation transaction. Thus, the simulation system efficiency of related technologies is low, and the time-consuming problem is solved. Through the structure of the large model cluster and the characteristics of model training, the distributed simulation of multiple CPU cores is used to reduce the number of transactions of the entire simulation system, thereby improving the simulation efficiency.

[0076] Next, the data center intelligent training simulation transaction acceleration device according to the embodiment of the present application is described with reference to the accompanying drawings.

[0077] Figure 4 is a block schematic diagram of a data center intelligent training simulation transaction acceleration device according to an embodiment of the present application.

[0078] As shown in Figure 4, the data center intelligent training simulation transaction acceleration device 10 includes an acquisition module 100, a judgment module 200, and a simulation module 300.

[0079] The acquisition module 100 is configured to acquire a current CPU computing power and a current GPU computing power to be simulated.

[0080] The judgment module 200 is configured to determine a current distributed simulation strategy according to the current CPU computing power and the current GPU computing power to be simulated, and to determine whether there is a deterministic simulation transaction in a transaction queue in the current distributed simulation strategy when simulating based on the current distributed simulation strategy.

[0081] The simulation module 300 is configured to delete the deterministic simulation transaction if there is a deterministic simulation transaction, and to perform accelerated simulation based on the remaining simulation transactions.

[0082] According to an embodiment of the present application, the judgment module 200 is specifically configured to:

[0083] determine whether the current CPU computing power is greater than the current GPU computing power to be simulated.

[0084] If the current CPU computing power is greater than the current GPU computing power to be simulated, calculate a first inter-core communication amount using all CPU cores on average and a second inter-core communication amount using CPU cores meeting a preset quantity requirement.

[0085] If the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed simulation strategy is to use the CPU cores meeting the preset quantity requirement for simulation, otherwise, the current distributed simulation strategy is to use all CPU cores on average for simulation.

[0086] According to an embodiment of the present application, after determining whether the current CPU computing power is greater than the current GPU computing power to be simulated, the judgment module 200 is further configured to:

[0087] If the current CPU computing power is less than or equal to the current GPU computing power to be simulated, calculate a third inter-core communication amount for a single CPU core to simulate all GPUs and a fourth inter-core communication amount for each CPU core to simulate the same type of GPU.

[0088] If the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed simulation strategy is to use a single CPU core to simulate all the current GPU to be simulated, otherwise, the current distributed simulation strategy is to use each CPU core to simulate the same type of GPU.

[0089] According to one embodiment of the present application, the judging module 200 is specifically used for:

[0090] obtaining a plurality of simulation transactions of a transaction queue in a current distributed simulation strategy;

[0091] judging whether there is an association between at least two simulation transactions in the plurality of simulation transactions;

[0092] if there is no association between the at least two simulation transactions, determining that there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy.

[0093] According to one embodiment of the present application, after determining that there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy, the judging module 200 is further used for:

[0094] judging whether other target transactions are inserted into the transaction queue;

[0095] if no other target transaction is inserted, the last simulation transaction in the simulation transaction queue.

[0096] According to the data center intelligent training simulation transaction acceleration device provided by the embodiment of the present application, the current CPU computing power and the current to-be-simulated GPU computing power are obtained, the current distributed simulation strategy is determined according to the current CPU computing power and the current to-be-simulated GPU computing power, and when simulation is performed based on the current distributed simulation strategy, it is judged whether there is a deterministic simulation transaction in the transaction queue in the current distributed simulation strategy. If there is a deterministic simulation transaction, the deterministic simulation transaction is deleted, and accelerated simulation is performed based on the remaining simulation transactions. Thus, the problems of low simulation system efficiency and long time consumption in the related art are solved. Through the structure of the large model cluster and the characteristics of the model training, the distributed simulation of the multi-CPU core is used to reduce the number of transactions of the entire simulation system, thereby improving the simulation efficiency.

[0097] FIG. 5 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. The electronic device can include:

[0098] a memory 501, a processor 502, and a computer program stored in the memory 501 and executable on the processor 502.

[0099] The processor 502 implements the data center intelligent training simulation transaction acceleration method provided in the above embodiments when executing the program.

[0100] Further, the electronic device further includes:

[0101] a communication interface 503 for communication between the memory 501 and the processor 502.

[0102] The memory 501 is used to store a computer program executable on the processor 502.

[0103] The memory 501 can include a high-speed RAM memory and can also include a non-volatile memory, such as at least one disk memory.

[0104] If the memory 501, the processor 502 and the communication interface 503 are implemented independently, the communication interface 503, the memory 501 and the processor 502 can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, only one thick line is used in FIG. 5, but it does not mean that there is only one bus or only one type of bus.

[0105] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can complete communication between each other through an internal interface.

[0106] The processor 502 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present application.

[0107] The embodiments of the present application also provide a computer readable storage medium, having stored thereon a computer program, which is executed by a processor to implement the data center intelligent training simulation transaction acceleration method as above.

[0108] The embodiments of the present application also provide a computer program product, comprising computer programs / instructions, which are executed by a processor to implement the steps of the method of the above embodiments.

[0109] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or N embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0110] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0111] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A data center intelligent training simulation transaction acceleration method, characterized in that, The method comprises the following steps: obtaining current central processing unit (CPU) computing power and current graphics processing unit (GPU) computing power to be emulated; determining a current distributed emulation strategy according to the current CPU computing power and the current GPU computing power to be emulated, and determining whether there is a deterministic emulation transaction in a transaction queue in the current distributed emulation strategy when emulation is performed based on the current distributed emulation strategy; if the deterministic emulation transaction exists, deleting the deterministic emulation transaction and performing accelerated emulation based on remaining emulation transactions.

2. The data center intelligent training emulation transaction acceleration method of claim 1, wherein, The determination of the current distributed emulation strategy according to the current CPU computing power and the current GPU computing power to be emulated comprises: determining whether the current CPU computing power is greater than the current GPU computing power to be emulated; if the current CPU computing power is greater than the current GPU computing power to be emulated, calculating a first inter-core communication amount of an average use of all CPU cores and a second inter-core communication amount of a use of a CPU core satisfying a preset quantity requirement; if the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed emulation strategy is to perform emulation by using the CPU core satisfying the preset quantity requirement, otherwise, the current distributed emulation strategy is to perform emulation by using the average use of all CPU cores.

3. The data center intelligent training emulation transaction acceleration method of claim 2, wherein, After determining whether the current CPU computing power is greater than the current GPU computing power to be emulated, the method further comprises: if the current CPU computing power is less than or equal to the current GPU computing power to be emulated, calculating a third inter-core communication amount of a single CPU core emulating all GPUs and a fourth inter-core communication amount of each CPU core emulating a same type of GPU; if the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed emulation strategy is to perform emulation by using a single CPU core to emulate all current GPU to be emulated, otherwise, the current distributed emulation strategy is to perform emulation by using each CPU core to emulate a same type of GPU.

4. The data center intelligent training emulation transaction acceleration method of claim 1, wherein, The determination of whether there is a deterministic emulation transaction in the transaction queue in the current distributed emulation strategy comprises: obtaining a plurality of emulation transactions in the transaction queue in the current distributed emulation strategy; determining whether there is an association between at least two emulation transactions in the plurality of emulation transactions; if there is no association between the at least two emulation transactions, it is determined that the transaction queue in the current distributed emulation strategy has the deterministic emulation transaction.

5. The data center intelligent training emulation transaction acceleration method of claim 4, wherein, After it is determined that the transaction queue in the current distributed emulation strategy has the deterministic emulation transaction, the method further comprises: determining whether other target transactions are inserted into the transaction queue; if the other target transactions are not inserted, emulating a last emulation transaction in the transaction queue.

6. A data center intelligent training emulation transaction acceleration device, comprising: The method comprises: an obtaining module, configured to obtain current CPU computing power and current GPU computing power to be emulated; a determining module, configured to determine a current distributed emulation strategy according to the current CPU computing power and the current GPU computing power to be emulated, and determine whether there is a deterministic emulation transaction in a transaction queue in the current distributed emulation strategy when emulation is performed based on the current distributed emulation strategy. The simulation module is configured to delete the deterministic simulation transaction and perform accelerated simulation based on the remaining simulation transactions if the deterministic simulation transaction exists.

7. The data center intelligent training emulation transaction acceleration apparatus of claim 6, wherein, The determining module is specifically configured to: determine whether the current CPU computing power is greater than the current GPU computing power to be simulated; if the current CPU computing power is greater than the current GPU computing power to be simulated, calculate a first inter-core communication amount of all CPU cores and a second inter-core communication amount of CPU cores satisfying a preset quantity requirement; if the first inter-core communication amount is greater than the second inter-core communication amount, the current distributed simulation strategy is to use the CPU cores satisfying the preset quantity requirement for simulation, otherwise, the current distributed simulation strategy is to use all the CPU cores for simulation.

8. The data center intelligent training emulation transaction acceleration apparatus of claim 7, wherein, After determining whether the current CPU computing power is greater than the current GPU computing power to be simulated, the determining module is further configured to: if the current CPU computing power is less than or equal to the current GPU computing power to be simulated, calculate a third inter-core communication amount of a single CPU core simulating all GPUs and a fourth inter-core communication amount of each CPU core simulating the same type of GPU; if the third inter-core communication amount is less than or equal to the fourth inter-core communication amount, the current distributed simulation strategy is to use a single CPU core to simulate all the current GPU to be simulated, otherwise, the current distributed simulation strategy is to use each CPU core to simulate the same type of GPU.

9. The data center intelligent training emulation transaction acceleration apparatus of claim 6, wherein, The determining module is specifically configured to: obtain a plurality of simulation transactions of a transaction queue in the current distributed simulation strategy; determine whether there is an association between at least two simulation transactions in the plurality of simulation transactions; if there is no association between the at least two simulation transactions, it is determined that the transaction queue in the current distributed simulation strategy has the deterministic simulation transaction.

10. The data center intelligent training emulation transaction acceleration device of claim 9, wherein, After determining that the transaction queue in the current distributed simulation strategy has the deterministic simulation transaction, the determining module is further configured to: determine whether other target transactions are inserted into the transaction queue; if the other target transactions are not inserted, simulate the last simulation transaction in the transaction queue.

11. An electronic device, comprising: comprise: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the data center intelligent training simulation transaction acceleration method according to any one of claims 1-5.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the data center intelligent training simulation transaction acceleration method according to any one of claims 1-5.

13. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of claim 1.

Images