Heterogeneous computing resource processing method and system, product, equipment and medium
By receiving and processing the application information of heterogeneous accelerator resources on the host side, the problem of low heterogeneous computing performance is solved, efficient utilization and performance improvement of heterogeneous accelerator is achieved, and the delay of resource calls is reduced.
Patent Information
- Application Number
- CN202411252003.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-15
AI Technical Summary
When applying for resources from heterogeneous computing devices at the remote end, heterogeneous computing resources cannot be fully utilized, resulting in a degradation of heterogeneous computing performance.
Receive the application information of the resource caller on the host side, start and configure the target heterogeneous accelerator based on the heterogeneous accelerator resources and information in the machine frame, and send complete information to the resource caller, allowing direct connection calls and cache consistency coordination, and optimize Resource calls to reduce latency and improve efficiency.
It improves the utilization and performance of heterogeneous accelerators, reduces the latency of resource calls, avoids the waste of heterogeneous accelerators, and realizes efficient interconnection and cache consistency between heterogeneous accelerators and hosts.
Smart Images

Figure CN118963970A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of resource calling, and in particular to a method, system, product, device and medium for processing heterogeneous computing resources. Background Art
[0002] With the rapid development of information technology, heterogeneous computing has become an important means to solve complex computing tasks. This computing model achieves higher computing efficiency and flexibility by distributing tasks to different computing nodes and utilizing their respective computing capabilities.
[0003] In practice, the remote end needs to apply for resources from heterogeneous computing devices. In this process, how to make full use of heterogeneous computing resources and improve the performance of heterogeneous computing is a technical problem that people in this field urgently need to solve. Summary of the invention
[0004] The purpose of the present invention is to provide a method, system, product, device and medium for processing heterogeneous computing resources to solve the technical problem that when remotely applying for resources from heterogeneous computing devices, heterogeneous computing resources cannot be fully utilized, thereby reducing the performance of heterogeneous computing.
[0005] In order to solve the above technical problems, the present invention provides a method for processing heterogeneous computing resources, which is applied to a host side; the method comprises: Receive information sent by a resource caller to characterize the application for heterogeneous accelerator resources; Starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for heterogeneous accelerator resources; wherein the frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The information indicating that the application for heterogeneous accelerator resources is completed is sent to the resource caller, so that the resource caller can directly call the resources of the started and configured target heterogeneous accelerator when receiving the information indicating that the application for heterogeneous accelerator resources is completed.
[0006] On the one hand, after sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: When the resource caller receives prompt information indicating that there are multiple target heterogeneous accelerators operating the central processing unit of the host at the same time, the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through the computing acceleration link.
[0007] On the other hand, after sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: Receiving information sent by the resource caller for characterizing accelerated computing; Invoking resources from the target heterogeneous accelerator according to the information used to characterize the accelerated computing; The resources called from the target heterogeneous accelerator are sent to the resource caller.
[0008] On the other hand, before the resource caller directly calls the resources of the enabled and configured target heterogeneous accelerator, or before the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through a computing acceleration link, or before receiving information sent by the resource caller for characterizing accelerated computing, it also includes: The resource caller receives the current stage of the target heterogeneous accelerator detected by the cloud platform; wherein the stage of the heterogeneous accelerator includes a startup and configuration stage, an accelerated computing stage, or a collaborative computing stage with the resource caller; If the information that the target heterogeneous accelerator is currently in the startup and configuration stage is received, the resource caller directly calls the resources of the started and configured target heterogeneous accelerator; If the information that the target heterogeneous accelerator is currently in the accelerated computing phase is received, the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through the computing acceleration link; If the information that the target heterogeneous accelerator is currently in the collaborative computing stage with the resource caller is received, the step of receiving information sent by the resource caller for characterizing the accelerated computing is entered.
[0009] On the other hand, the current stage of the target heterogeneous accelerator detected by the cloud platform includes: Tracking resource allocation and usage of each of the target heterogeneous accelerators; The current stage of each target heterogeneous accelerator is determined according to resource allocation and usage of each target heterogeneous accelerator.
[0010] On the other hand, the frame includes a plurality of hosts, each of which is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link; Before receiving the information sent by the resource caller and used to indicate the application for heterogeneous accelerator resources, the method further includes: Selecting a target host from all hosts in the chassis; The information sent by the resource caller and used to represent the application for heterogeneous accelerator resources includes: The target host is used to receive information sent by the resource caller and used to represent the application for heterogeneous accelerator resources.
[0011] On the other hand, determining the target host includes: Obtaining the current resource usage of all hosts in the chassis; Get the hosts whose current resource usage is less than the preset value; A host is randomly selected from all hosts whose current resource occupancy is less than a preset value as the target host.
[0012] On the other hand, starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources includes: Determine the target frame where the target heterogeneous accelerator is located according to the heterogeneous accelerator resources in all frames and the information used to characterize the application for heterogeneous accelerator resources; Determine whether the current frame and the target frame are the same frame; wherein the current frame is a frame that receives the information sent by the resource caller for characterizing the application for heterogeneous accelerator resources; If yes, the target heterogeneous accelerator is started and configured; If not, information for indicating starting and configuring the target heterogeneous accelerator is sent to the host in the target chassis, so that the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis.
[0013] On the other hand, after the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis, before sending information indicating completion of applying for heterogeneous accelerator resources to the resource caller, the method further includes: Receiving information sent by the target chassis for indicating completion of applying for heterogeneous accelerator resources; Sending information representing the resource being called from the target chassis to the resource caller; After sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: The resource caller calls resources from the target heterogeneous accelerator in the target machine frame by remote direct memory access.
[0014] On the other hand, before receiving the information sent by the resource caller and used to indicate the application for heterogeneous accelerator resources, the method further includes: The resource caller determines the target heterogeneous accelerator according to the performance parameters of each heterogeneous accelerator on the host side, and sends information containing the target heterogeneous accelerator to the host side.
[0015] On the other hand, before sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: Starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, determining whether prompt information indicating that the startup and configuration of the target heterogeneous accelerator is completed is received within a preset time period; If yes, the step of sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller is entered; If not, return to the step of starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources.
[0016] In order to solve the above technical problems, the present invention also provides a processing system for heterogeneous computing resources, including: a resource caller, a heterogeneous accelerator and a host; a machine frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The resource caller is used to send information indicating the application for heterogeneous accelerator resources to the host, and receive information sent by the resource caller indicating the completion of the application for heterogeneous accelerator resources. Upon receiving the information indicating the completion of the application for heterogeneous accelerator resources, directly call the resources of the started and configured target heterogeneous accelerator; The host is used to receive information sent by a resource caller for characterizing the application for heterogeneous accelerator resources; start and configure the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information for characterizing the application for heterogeneous accelerator resources; and send information to the resource caller for characterizing the completion of the application for heterogeneous accelerator resources.
[0017] In order to solve the above technical problem, the present invention also provides a computer program product, including a computer program / instruction, which implements the steps of the above-mentioned processing method of heterogeneous computing resources when executed by a processor.
[0018] In order to solve the above technical problems, the present invention further provides a processing device for heterogeneous computing resources, including: Memory for storing computer programs; A processor is used to implement the steps of the above-mentioned method for processing heterogeneous computing resources when executing the computer program.
[0019] In order to solve the above technical problem, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for processing heterogeneous computing resources are implemented.
[0020] The processing method of heterogeneous computing resources applied to the host side provided by the present invention first receives information sent by the resource caller for characterizing the application for heterogeneous accelerator resources; then starts and configures the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information for characterizing the application for heterogeneous accelerator resources; and finally sends information for characterizing the completion of the application for heterogeneous accelerator resources to the resource caller.
[0021] The beneficial effects of the present invention are as follows: firstly, since the target heterogeneous accelerator is started and configured in the host according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, there is no need to perform general operations such as starting and configuring in the target heterogeneous accelerator, and only accelerated computing needs to be performed in the target heterogeneous accelerator, that is, general computing is removed from the heterogeneous accelerator, but the ability of the heterogeneous accelerator to be directly called remotely is still retained. Compared with performing general computing such as starting and configuring on the heterogeneous accelerator, the waste of chip resources can be reduced, the utilization rate of the heterogeneous accelerator resources can be improved, and the performance of the heterogeneous accelerator can be improved; secondly, the heterogeneous accelerator resources are pooled, so that multiple hosts can share the resources in the heterogeneous accelerator, thereby improving the resource utilization rate. At the same time, the heterogeneous accelerator resource pool and the host are connected through a computing acceleration link (Compute Express The invention discloses a method for realizing efficient interconnection between the heterogeneous accelerator and the host through a Host-Connected Link (CXL), which satisfies the requirements of high-performance heterogeneous computing as much as possible and ensures the cache consistency between the host and the heterogeneous accelerator; secondly, when the resource caller receives the information used to indicate that the application for the heterogeneous accelerator resources is completed, the resource caller directly calls the resources of the started and configured target heterogeneous accelerator. Compared with the host calling method, the method provided by the present invention reduces the resource calling delay and improves the resource calling efficiency; in addition, when the resource caller receives the information used to indicate that the application for the heterogeneous accelerator resources is completed, the resource caller calls the resources in the target heterogeneous accelerator, which ensures that the target heterogeneous accelerator does not perform startup and configuration operations as much as possible, thereby avoiding the waste of heterogeneous accelerator resources.
[0022] In addition, when there are multiple target heterogeneous accelerators operating the host-side central processing unit at the same time, the resource caller in the method provided by the present invention shares data with each target heterogeneous accelerator and completes cache consistency coordination through a computing acceleration link, thereby realizing computing acceleration link collaborative computing, so that higher performance computing can be achieved when both the accelerator and the remote end need to calculate part of the calculation.
[0023] After sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the host calls resources from the target heterogeneous accelerator and sends the resources called from the target heterogeneous accelerator to the resource caller, that is, the host calls the resource, which ensures that the resource caller can use the resources in the heterogeneous accelerator.
[0024] The resource caller receives the current stage of the target heterogeneous accelerator detected by the cloud platform (such as the startup and configuration stage, the accelerated computing stage, or the collaborative computing stage with the resource caller), and adopts different resource calling methods according to the current stage of the target heterogeneous accelerator (such as the resource caller directly calling the resources of the target heterogeneous accelerator after startup and configuration, the resource caller shares data with each target heterogeneous accelerator and completes cache consistency coordination through computing acceleration links, or adopts the host calling method), making the resource caller more flexible in calling heterogeneous accelerator resources.
[0025] The machine frame includes multiple hosts, and each host is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link, so that the heterogeneous accelerator can be freely switched between multiple hosts.
[0026] When the resource caller makes a resource call to a heterogeneous accelerator in another chassis, the resource caller uses remote direct memory access to call resources from the target heterogeneous accelerator in the target chassis, ensuring that the resource caller can call the resources of the heterogeneous accelerator.
[0027] Starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, it is judged whether the prompt information of the completion of starting and configuring the target heterogeneous accelerator is received within the preset time. If not, it returns to starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, thereby ensuring that the host can complete the operation of starting and configuring the heterogeneous accelerator and avoiding the waste of heterogeneous accelerator resources caused by starting and configuring the heterogeneous accelerator.
[0028] In addition, the present invention also provides a processing system for heterogeneous computing resources, a computer program product, a processing device for heterogeneous computing resources, and a computer-readable storage medium, which have the same or corresponding technical features as the above-mentioned processing method for heterogeneous computing resources and have the same effects as above. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In order to more clearly illustrate the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0030] Figure 1 A flowchart of a method for processing heterogeneous computing resources provided by an embodiment of the present invention; Figure 2A schematic diagram of a first heterogeneous resource pooling provided in an embodiment of the present invention; Figure 3 A schematic diagram of a second heterogeneous resource pooling provided by an embodiment of the present invention; Figure 4 A schematic diagram of a heterogeneous computing system architecture provided by an embodiment of the present invention; Figure 5 A schematic diagram of a heterogeneous accelerator pooling architecture based on CXL provided in an embodiment of the present invention; Figure 6 A schematic diagram of remotely calling heterogeneous accelerator resources provided by an embodiment of the present invention; Figure 7 A structural diagram of a processing device for heterogeneous computing resources provided in another embodiment of the present invention. DETAILED DESCRIPTION
[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0032] The core of the present invention is to provide a method, system, product, device and medium for processing heterogeneous computing resources to solve the technical problem that when remotely applying for resources from heterogeneous computing devices, heterogeneous computing resources cannot be fully utilized, thereby reducing the performance of heterogeneous computing.
[0033] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Figure 1 A flowchart of a method for processing heterogeneous computing resources provided by an embodiment of the present invention, the method is applied to a host side; the method comprises: S10: receiving information sent by the resource caller for representing the application for heterogeneous accelerator resources; S11: starting and configuring a target heterogeneous accelerator according to the heterogeneous accelerator resources in the chassis and the information used to characterize the application for the heterogeneous accelerator resources; The frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; S12: Sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, so that the resource caller can directly call the resources of the started and configured target heterogeneous accelerator when receiving the information indicating that the application for heterogeneous accelerator resources is completed.
[0034] Heterogeneous computing mainly refers to the computing method of using computing units with different types of instruction sets and architectures to form a system. Heterogeneous accelerators refer to computing devices other than the central processing unit (CPU), such as graphics processing units (GPU), field-programmable gate arrays (FPGA), application-specific integrated circuits (ASIC), etc.
[0035] As an interconnection technology standard, CXL enables high-speed and efficient interconnection between CPU and GPU, FPGA or other accelerators, thereby meeting the requirements of high-performance heterogeneous computing, and it maintains cache consistency between CPU and heterogeneous devices. In general, its advantages lie in data transmission and cache consistency between devices. Since CXL has a cache consistency function, the computing work can be split onto different heterogeneous devices for computing. In this way, CXL can be used to connect devices (CPU, memory, solid state drive (SSD), GPU, FPGA, etc.) to form bare metal servers and virtual machine servers.
[0036] Heterogeneous accelerator pooling, like storage pooling, can bring many benefits: (1) More flexible heterogeneous accelerator resource configuration. For example, in a physical machine, whether one CPU is connected to one GPU or 16 GPUs, or whether 64GB of memory or 512GB of memory is used, the requirements for different workloads are usually inconsistent. For cloud vendors, facing the different preferences of different customers, non-pooled solutions may require frequent installation to implement.
[0037] (2) Higher heterogeneous accelerator resource utilization. For example, machine learning inference tasks do not require the entire GPU card. Different time periods and different usage methods will cause the GPU card to be idle in time and space. Through GPU virtualization / pooling, it is possible to oversell GPUs.
[0038] Heterogeneous resource pooling is to group the devices in the pooled chassis into units that can be called remotely. Figure 2 A schematic diagram of a first heterogeneous resource pooling provided by an embodiment of the present invention, such as Figure 2As shown in the figure, in the chassis, the central processing unit, memory, solid-state drive, graphics processor, field programmable gate array, etc. together form the bare metal server and virtual machine server. Through configuration, the resources in the pooled chassis are divided into various hosts, and the divided hosts provide the basis for remote calls of heterogeneous resources. Host calls are made using methods such as remote procedure calls (RPC), which require a large amount of interaction between the CPU and the heterogeneous acceleration card. During this interaction, a large processing delay will be generated, which is usually 100 to 1000 times slower than the direct connection method of remotely calling resources directly from heterogeneous accelerators, which will seriously affect the performance of distributed heterogeneous computing.
[0039] Figure 3 A schematic diagram of a second heterogeneous resource pooling provided in an embodiment of the present invention, such as Figure 3 As shown in FIG. 1 , the chassis only contains multiple heterogeneous accelerators. The heterogeneous accelerators themselves have the ability to communicate remotely and can be directly called remotely.
[0040] However, for the first method, all remote calls are completed through the cooperation of different devices between hosts. The CPU is heavily involved in communication and other calculations, and there will be a lot of interactions between the CPU, memory, GPU, and FPGA, which reduces the performance of pooling.
[0041] In the second method, heterogeneous accelerators can be directly connected to the caller remotely without the need for interaction between the CPU, memory, GPU, and FPGA. They have high performance, but heterogeneous accelerators are not suitable for general computing such as configuration and startup, resulting in a lot of chip resources being used to process tasks that are more suitable for CPU computing. If the chip is used entirely for general computing rather than accelerated computing, more than half of the chip resources will be wasted. General computing is different from accelerated computing. Using only accelerator chips will waste a lot of chip resources and should be implemented in software. Accelerated computing should be responsible for some computing tasks that are simple in function but extremely computationally intensive, such as artificial intelligence (AI), encryption and decryption, and video encoding and decoding.
[0042] It can be seen that both the first and second methods provide remote calls of pooled heterogeneous accelerator resources in a fixed manner, and both methods have problems. The first method is that the host provides the overall function, so a lot of interaction is required between the various components of the host, which reduces performance and has performance shortcomings. Although the second method has performance advantages, it cannot handle complex general-purpose calculations, which means that the solution can only handle particularly simple computing problems or waste a lot of chip resources to handle things that are more suitable for the CPU, and there are functional shortcomings.
[0043] Therefore, in the embodiment of the present invention, the heterogeneous accelerators are further reconstructed in a pooling manner based on the CXL technology. Figure 4 A schematic diagram of a heterogeneous computing system architecture provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown, the heterogeneous computing system architecture is divided into four layers, including heterogeneous device connection methods, heterogeneous computing hardware resources, heterogeneous resource organization methods, and pooled heterogeneous accelerator remote calls.
[0044] The heterogeneous device connection method provided in the embodiment of the present invention adopts computing acceleration link, direct memory access (Direct Memory Access, DMA), remote direct memory access (Remote Direct Memory Access, RDMA) and others. CXL is responsible for high-performance interaction between CPU, memory, GPU, and FPGA, and DMA / RDMA is responsible for the transmission of large blocks of data. CXL connection can avoid the need to divide heterogeneous accelerators into hosts, which increases flexibility. Heterogeneous computing hardware resources include graphics processors, field programmable gate arrays, application-specific integrated circuits and others. The first method adopts the Peripheral Component Interconnect Express (PCIe) / DMA / RDMA method, in which heterogeneous devices are divided into hosts, which has poor flexibility, and the business process requires frequent interaction between CPU and heterogeneous accelerators. The second method uses RDMA and other methods to directly connect to the remote caller, and must rely on heterogeneous accelerators to complete transactional general computing work, which is more suitable for CPU.
[0045] The heterogeneous resource organization method consists of pooling management based on computing acceleration links and direct connection configuration of heterogeneous accelerators. Pooling management based on CXL is responsible for organizing resources. Heterogeneous accelerators obtain platform information and perform resource scheduling. Heterogeneous accelerator direct connection configuration provides a configuration basis for high-speed calling of heterogeneous accelerators. The advantage of this is that it can provide a flexible and dynamic heterogeneous accelerator calling method. The first method is organized and managed through openstack, k8s, etc. The second method has no organizational management in the pooled machine frame, and all management is completed by the remote caller. In an embodiment of the present invention, general computing is completed by the CPU, but the host with the CPU as the core and the heterogeneous accelerator are in a weakly coupled relationship. After the general computing is completed and configured, the heterogeneous accelerator will be in a direct call or CXL collaborative computing state.
[0046] In the embodiment of the present invention, the pooled heterogeneous accelerator remote call provides three modes: computing acceleration link collaborative computing, direct call and host call. The three modes can be switched dynamically, and not only one mode can be used for calling. The first mode only provides host call, and the second mode only provides direct call. The remote call provided by the embodiment of the present invention has the versatility of host call and the performance of direct call.
[0047] In order to implement the method for remote calling of pooled heterogeneous accelerators provided by the present invention, Figure 5 A schematic diagram of a heterogeneous accelerator pooling architecture based on CXL provided in an embodiment of the present invention is shown in FIG. Figure 5 As shown in the figure, the chassis contains multiple heterogeneous accelerators, and all heterogeneous accelerators form a heterogeneous accelerator resource pool. CXL-based pool management no longer assigns heterogeneous accelerators to a host, but connects to the management host through a computing acceleration link. The management host is in the same chassis and consists of a central processing unit, memory, solid-state drive, etc. The management host runs an operating system and management software, which is responsible for general computing processing such as configuration and startup. After the heterogeneous accelerator is configured and started, it can provide direct remote calls.
[0048] Based on the heterogeneous accelerator pooling architecture of CXL provided by an embodiment of the present invention, in the method for processing heterogeneous computing resources, the method is applied to the host side, and the host side receives the information sent by the resource caller for characterizing the application for heterogeneous accelerator resources, and then starts and configures the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information for characterizing the application for heterogeneous accelerator resources; finally, the information for characterizing the completion of the application for heterogeneous accelerator resources is sent to the resource caller. When the resource caller receives the information for characterizing the completion of the application for heterogeneous accelerator resources, it directly calls the resources of the started and configured target heterogeneous accelerator. This method is also called a direct call method.
[0049] In addition, after sending information indicating the completion of the application for heterogeneous accelerator resources to the resource caller, it also includes: When the resource caller receives prompt information indicating that there are multiple target heterogeneous accelerators operating the host CPU at the same time, it shares data with each target heterogeneous accelerator and completes cache consistency coordination through the computing acceleration link. That is, it supports host calls and CXL collaborative computing. CXL collaborative computing is similar to direct calls, but not all data is transmitted to the remote end. The heterogeneous accelerator shares data with the caller, completes cache consistency coordination through CXL, and realizes collaborative computing.
[0050] In addition, after sending information indicating the completion of applying for heterogeneous accelerator resources to the resource caller, it also includes: Receiving information sent by a resource caller for characterizing accelerated computing; Invoke resources from a target heterogeneous accelerator according to information used to characterize accelerated computing; Send the resources called from the target heterogeneous accelerator to the resource caller.
[0051] In the embodiment of the present invention, this method is also referred to as a host call method. In the host call method, the heterogeneous accelerator is still connected to the management host through CXL, and is not directly assigned to a host as in the first method described above. Even the host call method has greater flexibility than the first method. The heterogeneous accelerator can switch freely between multiple hosts, or dynamically switch from host call to direct call.
[0052] It can be seen that for the three ways of remote calling of the pooled heterogeneous accelerator provided by the embodiment of the present invention, CXL collaborative computing, direct connection calling and host calling, the three ways can be switched dynamically, not just one way of calling, making the resource calling method more flexible. At the same time, the CXL collaborative computing method is added, which can achieve higher performance computing when both the accelerator and the remote end need to calculate part of it. The embodiment of the present invention uses host calling when starting and configuring processing; direct connection calling is used for simple accelerated computing; CXL collaborative computing is used for collaborative computing between the remote end and the pooled accelerator.
[0053] In order to determine the calling mode to be adopted, before the resource caller directly calls the resources of the enabled and configured target heterogeneous accelerator, or before the resource caller shares data with each target heterogeneous accelerator and completes cache consistency coordination through the computing acceleration link, or before receiving the information sent by the resource caller for characterizing the accelerated computing, it also includes: The resource caller receives the current stage of the target heterogeneous accelerator detected by the cloud platform; wherein the stage of the heterogeneous accelerator includes the startup and configuration stage, the accelerated computing stage, or the collaborative computing stage with the resource caller; If the information is received that the target heterogeneous accelerator is currently in the startup and configuration stage, the resource caller directly calls the resources of the started and configured target heterogeneous accelerator.
[0054] If the resource caller receives information that the target heterogeneous accelerator is currently in the accelerated computing stage, the resource caller shares data with each target heterogeneous accelerator and completes cache consistency coordination through the computing acceleration link.
[0055] If information is received that the target heterogeneous accelerator is currently in a collaborative computing phase with the resource caller, the process proceeds to a step of receiving information sent by the resource caller for characterizing the accelerated computing.
[0056] The current stage of the target heterogeneous accelerator detected by the cloud platform includes: Track resource allocation and usage of each target heterogeneous accelerator; The current stage of each target heterogeneous accelerator is determined according to the resource allocation and usage of each target heterogeneous accelerator.
[0057] In the method provided by the embodiment of the present invention, the resource caller receives the current stage of the target heterogeneous accelerator detected by the cloud platform (such as the startup and configuration stage, the accelerated computing stage, or the collaborative computing stage with the resource caller), thereby adopting different resource calling methods according to the current stage of the target heterogeneous accelerator (such as the resource caller directly calling the resources of the target heterogeneous accelerator after startup and configuration, the resource caller sharing data with each target heterogeneous accelerator and completing cache consistency coordination through computing acceleration links, or adopting a host calling method), so that the resource caller can call the heterogeneous accelerator resources in a more flexible way.
[0058] When there are multiple hosts in the chassis, each host is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link; Before receiving the information sent by the resource caller for representing the application for heterogeneous accelerator resources, the method further includes: Select the target host from all hosts in the chassis; The information sent by the resource caller to characterize the application for heterogeneous accelerator resources includes: The target host receives information sent by the resource caller for representing the application for heterogeneous accelerator resources.
[0059] In order to improve the efficiency of resource calling, in implementation, determining the target host includes: Get the current resource usage of all hosts in the chassis; Get the hosts whose current resource usage is less than the preset value; A host is randomly selected as the target host from all hosts whose current resource usage is less than the preset value.
[0060] There is no limitation on the preset value, which is determined according to the actual situation. In the method provided in this embodiment, the machine frame includes multiple hosts, and each host is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link, so that the heterogeneous accelerator can be freely switched between multiple hosts, and because the host whose current resource occupancy is less than the preset value is selected as the target host, the efficiency of resource calling is improved.
[0061] In practice, the heterogeneous accelerator called by the resource caller may not be located in the chassis where the resource caller initiates the resource application. In this case, starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the chassis and the information used to characterize the application for the heterogeneous accelerator resources includes: Determine the target chassis where the target heterogeneous accelerator is located according to the heterogeneous accelerator resources in all chassis and the information used to characterize the application for the heterogeneous accelerator resources; Determine whether the current chassis and the target chassis are the same chassis; wherein the current chassis is the chassis that receives the information sent by the resource caller for representing the application for heterogeneous accelerator resources; If yes, the target heterogeneous accelerator is started and configured; If not, information for indicating starting and configuring the target heterogeneous accelerator is sent to the host in the target chassis, so that the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis.
[0062] After the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis, before sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the following further includes: Receiving information sent by the target chassis to indicate that the application for heterogeneous accelerator resources is completed; Sending information for representing the resource call from the target chassis to the resource caller; After sending information indicating the completion of applying for heterogeneous accelerator resources to the resource caller, it also includes: The resource caller uses remote direct memory access to call resources from the target heterogeneous accelerator in the target chassis.
[0063] In the method provided in this embodiment, when the resource caller makes a resource call to a heterogeneous accelerator in another machine frame, the resource caller uses remote direct memory access to call resources from the target heterogeneous accelerator in the target machine frame, thereby ensuring that the resource caller can call the resources of the heterogeneous accelerator.
[0064] In order to ensure that the resource caller can apply for the resources of the heterogeneous accelerator, in implementation, before receiving the information sent by the resource caller for representing the application for the heterogeneous accelerator resources, the following is also included: The resource caller determines the target heterogeneous accelerator according to the performance parameters of each heterogeneous accelerator on the host side, and sends information including the target heterogeneous accelerator to the host side.
[0065] The performance parameters of the heterogeneous accelerators, such as the type and computing power of the heterogeneous accelerators, are not limited. By determining the target heterogeneous accelerator according to the performance parameters of each heterogeneous accelerator, the selected target heterogeneous accelerator is more suitable.
[0066] In practice, after the host starts and configures the target heterogeneous accelerator according to the heterogeneous accelerator resources in the chassis and the information used to characterize the application for the heterogeneous accelerator resources, the target heterogeneous accelerator may not be started and configured successfully. Therefore, in some embodiments, before sending the information used to characterize the completion of the application for the heterogeneous accelerator resources to the resource caller, it also includes: Starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the chassis and the information used to characterize the application for the heterogeneous accelerator resources, determining whether a prompt message indicating that the startup and configuration of the target heterogeneous accelerator is completed is received within a preset time period; If yes, the step of sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller is entered; If not, the process returns to the step of starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the chassis and the information used to characterize the application for the heterogeneous accelerator resources.
[0067] There is no limit on the preset duration, which will be determined based on actual conditions.
[0068] In the method provided in this embodiment, starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, it is judged whether the prompt information of the completion of starting and configuring the target heterogeneous accelerator is received within the preset time. If not, it returns to starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, thereby ensuring that the host can complete the operation of starting and configuring the heterogeneous accelerator and avoiding the waste of heterogeneous accelerator resources caused by starting and configuring the heterogeneous accelerator.
[0069] In order to make those skilled in the art better understand the present invention, Figure 6 The present invention is further described in detail with specific implementation modes. Figure 6 A schematic diagram of remotely calling heterogeneous accelerator resources provided in an embodiment of the present invention. The process of remotely calling heterogeneous accelerator resources is an interaction between the remote end, the heterogeneous accelerator, and the management host. The process includes: the remote end applies for heterogeneous acceleration resources from the management host; the management host starts / configures the operation of the heterogeneous accelerator; the heterogeneous accelerator sends completion information to the management host; the management host sends information about the completion of the application for heterogeneous acceleration resources to the remote end (the above process can be called the management configuration stage of the host). The remote end sends a remote direct call message to the heterogeneous accelerator; the heterogeneous accelerator responds to the remote end with a remote direct call (this process can be called the stage where the heterogeneous accelerator performs accelerated computing).
[0070] That is, the remote end first applies for heterogeneous accelerator resources from the management host. The management host configures relevant information to the heterogeneous accelerator based on the heterogeneous accelerator resources in the frame and the information of the remote caller, thereby completing the heterogeneous resource application. After that, the remote caller and the heterogeneous accelerator perform high-performance remote calls through direct connection. In addition to direct calls, the embodiments of the present invention also support host calls and CXL collaborative computing, and the running process and Figure 6 It is the same and will not be repeated here.
[0071] In implementation, there may be multiple remote ends applying for heterogeneous accelerator resources at the same time. In this case, in order to enable each remote end to call the corresponding heterogeneous accelerator resources, the method includes obtaining the attributes of the heterogeneous resources to be applied for by each remote end; preliminarily determining the heterogeneous accelerator to be applied for according to the attributes of the heterogeneous resources to be applied for by each remote end; obtaining the remaining computing power of the heterogeneous accelerator to be applied for; if the remaining computing power of the heterogeneous accelerator resources to be applied for is greater than or equal to the preset computing power, the preliminarily determined heterogeneous accelerator to be applied for is used as the final heterogeneous accelerator to be applied for; if the remaining computing power of the heterogeneous accelerator resources to be applied for is less than the preset computing power, an idle heterogeneous accelerator is obtained; and the idle heterogeneous accelerator is used as the final heterogeneous accelerator to be applied for. In the method, in the case where multiple remote ends apply for heterogeneous accelerator resources at the same time, the final heterogeneous accelerator to be applied for is determined according to the attributes of the heterogeneous resources requested by each remote end and the preliminarily determined remaining computing power of the heterogeneous accelerator resources to be applied for, so that a suitable heterogeneous accelerator can be selected to process the resource requests of each remote end.
[0072] In the method provided by the embodiment of the present invention, the organization form of bare metal servers and virtual machine servers is weakened, and a more flexible heterogeneous accelerator interconnection and interoperation mode is realized through CXL. Flexible operation can realize dynamic remote calling mode and weak coupling between heterogeneous accelerators and CPUs; general computing is removed from heterogeneous accelerators, but the ability of heterogeneous accelerators to be directly called remotely is still retained.
[0073] A method for processing heterogeneous computing resources is described above. This embodiment further provides a system for processing heterogeneous computing resources, including: a resource caller, a heterogeneous accelerator, and a host; a machine frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The resource caller is used to send information indicating the application for heterogeneous accelerator resources to the host, and receive information sent by the resource caller indicating the completion of the application for heterogeneous accelerator resources. When receiving the information indicating the completion of the application for heterogeneous accelerator resources, directly call the resources of the started and configured target heterogeneous accelerator; The host is used to receive information sent by the resource caller for representing the application for heterogeneous accelerator resources; start and configure the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information for representing the application for heterogeneous accelerator resources; and send information to the resource caller for representing the completion of the application for heterogeneous accelerator resources.
[0074] The processing system for heterogeneous computing resources provided in this embodiment has the same or corresponding technical features as the processing method for heterogeneous computing resources described above. The embodiment of the processing method for heterogeneous computing resources has been described in detail above, and the embodiment of the processing system for heterogeneous computing resources will not be repeated here, and the effect is the same as above.
[0075] In the above embodiments, the processing method of heterogeneous computing resources is described in detail, and the present invention also provides a processing device for heterogeneous computing resources and a corresponding embodiment of the processing equipment for heterogeneous computing resources. It should be noted that the present invention describes the embodiments of the device part from two perspectives, one is based on the perspective of functional modules, and the other is based on the perspective of hardware.
[0076] An embodiment of the present invention provides a processing device for heterogeneous computing resources, which includes, based on the perspective of functional modules: A first receiving module is used to receive information sent by a resource caller and used to represent the application for heterogeneous accelerator resources; A startup and configuration module, used to start and configure a target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources; wherein the frame includes a host and a heterogeneous accelerator resource pool composed of a plurality of heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The first sending module is used to send information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, so that the resource caller can directly call the resources of the started and configured target heterogeneous accelerator when receiving the information indicating that the application for heterogeneous accelerator resources is completed.
[0077] In some embodiments, in a processing device of heterogeneous computing resources, when a resource caller receives prompt information indicating that there are multiple target heterogeneous accelerators operating a central processing unit on a host side at the same time, it shares data with each target heterogeneous accelerator and completes cache consistency coordination through a computing acceleration link.
[0078] In some embodiments, the processing device of the heterogeneous computing resources further includes: A second receiving module is used to receive information sent by the resource caller for characterizing the accelerated computing; A calling module, used for calling resources from a target heterogeneous accelerator according to information used for characterizing accelerated computing; The second sending module is used to send the resources called from the target heterogeneous accelerator to the resource caller.
[0079] In some embodiments, in a processing device of heterogeneous computing resources, a resource caller receives the current stage of a target heterogeneous accelerator detected by a cloud platform; wherein the stage of the heterogeneous accelerator includes a startup and configuration stage, an accelerated computing stage, or a collaborative computing stage with the resource caller; if information is received that the current stage of the target heterogeneous accelerator is the startup and configuration stage, the resource caller directly calls the resources of the started and configured target heterogeneous accelerator; if information is received that the current stage of the target heterogeneous accelerator is the accelerated computing stage, the resource caller shares data with each target heterogeneous accelerator and completes cache consistency coordination through a computing acceleration link; if information is received that the current stage of the target heterogeneous accelerator is the collaborative computing stage with the resource caller, the second receiving module is triggered.
[0080] In some embodiments, the chassis includes a plurality of hosts, each of which is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link; The processing device of the heterogeneous computing resources also includes: A selection module is used to select a target host from all hosts in the chassis; The first receiving module is specifically used to use the target host to receive information sent by the resource caller and used to represent the application for heterogeneous accelerator resources.
[0081] In some embodiments, the startup and configuration module includes: A determination module, used to determine a target chassis where a target heterogeneous accelerator is located according to the heterogeneous accelerator resources in all chassis and the information used to characterize the application for heterogeneous accelerator resources; The first judgment module is used to judge whether the current frame and the target frame are the same frame; wherein the current frame is the frame that receives the information sent by the resource caller for representing the application for heterogeneous accelerator resources; if so, the startup and configuration submodule is triggered; if not, the third sending module is triggered; The startup and configuration submodule is used to start and configure the target heterogeneous accelerator; The third sending module is used to send information used to characterize the startup and configuration of the target heterogeneous accelerator to the host in the target machine frame, so that the host in the target machine frame can start and configure the target heterogeneous accelerator located in the target machine frame.
[0082] In some embodiments, the processing device of the heterogeneous computing resources further includes: A third receiving module is used to receive information sent by the target chassis to indicate that the application for heterogeneous accelerator resources is completed; A fourth sending module, used for sending information representing the resource being called from the target chassis to the resource caller; The resource caller of the processing device of the heterogeneous computing resource uses a remote direct memory access method to call resources from the target heterogeneous accelerator in the target machine frame.
[0083] In some embodiments, a resource caller in a processing device of heterogeneous computing resources determines a target heterogeneous accelerator according to performance parameters of each heterogeneous accelerator on the host side, and sends information including the target heterogeneous accelerator to the host side.
[0084] In some embodiments, the processing device of the heterogeneous computing resources further includes: The second judgment module is used to judge whether a prompt message indicating that the startup and configuration of the target heterogeneous accelerator is completed is received within a preset time period, starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources; if so, trigger the first sending module; if not, return to trigger the startup and configuration module.
[0085] Since the embodiments of the device part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the device part, and they will not be repeated here. The effects are the same as above.
[0086] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, which implements the steps of the above-mentioned method for processing heterogeneous computing resources when executed by a processor.
[0087] Figure 7 This is a structural diagram of a processing device for heterogeneous computing resources provided by another embodiment of the present invention. This embodiment is based on the hardware perspective, such as Figure 7 As shown, the processing devices of the heterogeneous computing resources include: A memory 20, used for storing computer programs; The processor 21 is used to implement the steps of the method for processing heterogeneous computing resources mentioned in the above embodiment when executing a computer program.
[0088] Among them, the processor 21 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 can be implemented in at least one hardware form of a digital signal processor (DSP), an FPGA, and a programmable logic array (PLA). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU; the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 21 may be integrated with a GPU, which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 21 may also include an AI processor, which is used to process computing operations related to machine learning.
[0089] The memory 20 may include one or more computer-readable storage media, which may be non-transitory. The memory 20 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In this embodiment, the memory 20 is at least used to store the following computer program 201, wherein, after the computer program is loaded and executed by the processor 21, the relevant steps of the processing method of heterogeneous computing resources disclosed in any of the aforementioned embodiments can be implemented. In addition, the resources stored in the memory 20 may also include an operating system 202 and data 203, etc., and the storage method may be temporary storage or permanent storage. Among them, the operating system 202 may include Windows, Unix, Linux, etc. Data 203 may include, but is not limited to, data involved in the above-mentioned processing method of heterogeneous computing resources, etc.
[0090] In some embodiments, the processing device of the heterogeneous computing resources may also include a display screen 22 , an input and output interface 23 , a communication interface 24 , a power supply 25 , and a communication bus 26 .
[0091] Those skilled in the art will understand that Figure 7 The structure shown in the figure does not constitute a limitation on the processing device of heterogeneous computing resources, and may include more or fewer components than shown in the figure.
[0092] The processing device for heterogeneous computing resources provided by the embodiment of the present invention includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the following method: the processing method for heterogeneous computing resources has the same effect as above.
[0093] Finally, the present invention also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps recorded in the above method embodiment are implemented.
[0094] It is understandable that if the method in the above embodiment is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc. Various media that can store program codes.
[0095] The computer-readable storage medium provided by the present invention includes the above-mentioned method for processing heterogeneous computing resources, and the effect is the same as above.
[0096] The above is a detailed introduction to a processing method, system, product, device and medium for heterogeneous computing resources provided by the present invention. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referenced to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the present invention.
[0097] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.
Claims
1. A method for processing heterogeneous computing resources, characterized in that: Applied to the host side; the method comprises: Receive information sent by a resource caller to characterize the application for heterogeneous accelerator resources; Starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for heterogeneous accelerator resources; wherein the frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The information indicating that the application for heterogeneous accelerator resources is completed is sent to the resource caller, so that the resource caller can directly call the resources of the started and configured target heterogeneous accelerator when receiving the information indicating that the application for heterogeneous accelerator resources is completed.
2. The method for processing heterogeneous computing resources according to claim 1, characterized in that: After sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: When the resource caller receives prompt information indicating that there are multiple target heterogeneous accelerators operating the central processing unit of the host at the same time, the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through the computing acceleration link.
3. The method for processing heterogeneous computing resources according to claim 2, characterized in that: After sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: Receiving information sent by the resource caller for characterizing accelerated computing; Invoking resources from the target heterogeneous accelerator according to the information used to characterize the accelerated computing; The resources called from the target heterogeneous accelerator are sent to the resource caller.
4. The method for processing heterogeneous computing resources according to claim 3, characterized in that: Before the resource caller directly calls the resources of the started and configured target heterogeneous accelerator, or before the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through a computing acceleration link, or before receiving information sent by the resource caller for characterizing accelerated computing, the method further includes: The resource caller receives the current stage of the target heterogeneous accelerator detected by the cloud platform; wherein the stage of the heterogeneous accelerator includes a startup and configuration stage, an accelerated computing stage, or a collaborative computing stage with the resource caller; If the information that the target heterogeneous accelerator is currently in the startup and configuration stage is received, the resource caller directly calls the resources of the started and configured target heterogeneous accelerator; If the information that the target heterogeneous accelerator is currently in the accelerated computing phase is received, the resource caller shares data with each of the target heterogeneous accelerators and completes cache consistency coordination through the computing acceleration link; If the information that the target heterogeneous accelerator is currently in the collaborative computing stage with the resource caller is received, the step of receiving information sent by the resource caller for characterizing the accelerated computing is entered.
5. The method for processing heterogeneous computing resources according to claim 4, characterized in that: The current stage of the target heterogeneous accelerator detected by the cloud platform includes: Tracking resource allocation and usage of each of the target heterogeneous accelerators; The current stage of each target heterogeneous accelerator is determined according to resource allocation and usage of each target heterogeneous accelerator.
6. The method for processing heterogeneous computing resources according to any one of claims 1 to 5, characterized in that: The frame includes a plurality of hosts, each of which is connected to each heterogeneous accelerator in the heterogeneous accelerator resource pool through a computing acceleration link; Before receiving the information sent by the resource caller and used to indicate the application for heterogeneous accelerator resources, the method further includes: Selecting a target host from all hosts in the chassis; The information sent by the resource caller and used to represent the application for heterogeneous accelerator resources includes: The target host is used to receive information sent by the resource caller and used to represent the application for heterogeneous accelerator resources.
7. The method for processing heterogeneous computing resources according to claim 6, characterized in that: Determining the target host includes: Obtaining the current resource usage of all hosts in the chassis; Get the hosts whose current resource usage is less than the preset value; A host is randomly selected from all hosts whose current resource occupancy is less than a preset value as the target host.
8. The method for processing heterogeneous computing resources according to claim 1, characterized in that: The starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources includes: Determine the target frame where the target heterogeneous accelerator is located according to the heterogeneous accelerator resources in all frames and the information used to characterize the application for heterogeneous accelerator resources; Determine whether the current frame and the target frame are the same frame; wherein the current frame is a frame that receives the information sent by the resource caller for characterizing the application for heterogeneous accelerator resources; If yes, the target heterogeneous accelerator is started and configured; If not, information for indicating starting and configuring the target heterogeneous accelerator is sent to the host in the target chassis, so that the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis.
9. The method for processing heterogeneous computing resources according to claim 8, characterized in that: After the host in the target chassis starts and configures the target heterogeneous accelerator in the target chassis, before sending information indicating completion of applying for heterogeneous accelerator resources to the resource caller, the method further includes: Receiving information sent by the target chassis for indicating completion of applying for heterogeneous accelerator resources; Sending information representing the resource being called from the target chassis to the resource caller; After sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: The resource caller calls resources from the target heterogeneous accelerator in the target machine frame by remote direct memory access.
10. The method for processing heterogeneous computing resources according to claim 1, characterized in that: Before receiving the information sent by the resource caller and used to indicate the application for heterogeneous accelerator resources, the method further includes: The resource caller determines the target heterogeneous accelerator according to the performance parameters of each heterogeneous accelerator on the host side, and sends information containing the target heterogeneous accelerator to the host side.
11. The method for processing heterogeneous computing resources according to claim 1, characterized in that: Before sending the information indicating that the application for heterogeneous accelerator resources is completed to the resource caller, the method further includes: Starting from starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources, determining whether prompt information indicating that the startup and configuration of the target heterogeneous accelerator is completed is received within a preset time period; If yes, the step of sending information indicating that the application for heterogeneous accelerator resources is completed to the resource caller is entered; If not, return to the step of starting and configuring the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information used to characterize the application for the heterogeneous accelerator resources.
12. A processing system for heterogeneous computing resources, characterized in that: include: Resource caller, heterogeneous accelerator and host; the machine frame includes a host and a heterogeneous accelerator resource pool composed of multiple heterogeneous accelerators, and the heterogeneous accelerator resource pool is connected to the host through a computing acceleration link; The resource caller is used to send information indicating the application for heterogeneous accelerator resources to the host, and receive information sent by the resource caller indicating the completion of the application for heterogeneous accelerator resources. Upon receiving the information indicating the completion of the application for heterogeneous accelerator resources, directly call the resources of the started and configured target heterogeneous accelerator; The host is used to receive information sent by a resource caller for characterizing the application for heterogeneous accelerator resources; start and configure the target heterogeneous accelerator according to the heterogeneous accelerator resources in the frame and the information for characterizing the application for heterogeneous accelerator resources; and send information to the resource caller for characterizing the completion of the application for heterogeneous accelerator resources.
13. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for processing heterogeneous computing resources described in any one of claims 1 to 11 are implemented.
14. A processing device for heterogeneous computing resources, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method for processing heterogeneous computing resources as described in any one of claims 1 to 11 when executing the computer program.
15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method for processing heterogeneous computing resources according to any one of claims 1 to 11 are implemented.
