Accelerator resource scheduling method, device and equipment and readable storage medium

Abstract

The invention provides an accelerator resource scheduling method, device and equipment and a readable storage medium, and the method comprises the steps: dividing a server equipped with a first type of instance into a first type of server and a second type of server; the first class instance is an instance obtained based on virtualization of a target class physical accelerator of the server; only first-class instances supporting thermal migration are deployed on the first-class servers; first-class instances which do not support thermal migration are deployed on the second-class servers; executing a live migration operation of the first type of instances for at least one server in the first type of servers and at least one server in the second type of servers; the thermal migration operation comprises the step of thermally migrating all the first-class instances deployed on one server in the first-class servers to one server in the second-class servers. According to the scheme, the problem that a large amount of invalid power consumption exists in the use scheme of the virtual accelerator in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of wireless technology, and in particular to a method, apparatus, device, and readable storage medium for scheduling accelerator resources. Background Technology

[0002] Currently, multiple physical accelerators in the wireless cloud, after being enabled by virtualization, can each generate multiple virtual accelerators for service use. However, in actual operation, the virtual accelerators used by services may be distributed across multiple physical accelerators. In this case, only a small portion of the virtual accelerators generated by a large number of physical accelerators are actually used, but these physical accelerators still require virtualization to be enabled, resulting in a significant amount of wasted power consumption. Moreover, the power consumption of physical accelerators configured with virtualization but without services loaded is currently much higher than that of those without virtualization and without services loaded. Even if only a portion of the virtualized physical accelerators are loaded with services, the overall power consumption of the physical accelerators will still be relatively high.

[0003] As shown above, existing technologies for using virtual accelerators have drawbacks such as significant wasted power consumption. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, device and readable storage medium for scheduling accelerator resources, so as to solve the problem of a large amount of ineffective power consumption in the existing technology for the use of virtual accelerators.

[0005] To address the aforementioned technical problems, embodiments of this application provide a method for scheduling accelerator resources, comprising:

[0006] Servers with first-category instances deployed are divided into first-category servers and second-category servers; the first-category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first-category servers only have first-category instances that support hot migration deployed; the second-category servers have first-category instances that do not support hot migration deployed.

[0007] Perform a hot migration operation on at least one server of the first type of servers and at least one server of the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server of the first type of servers to one server of the second type of servers.

[0008] Optionally, performing the hot migration operation for at least one server in the first type of servers and at least one server in the second type of servers includes:

[0009] Based on the first parameter information, determine the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration;

[0010] Based on the first and second information, perform a hot migration operation for the first category of instances;

[0011] The first parameter information includes at least one of the following:

[0012] The resource information corresponding to each first server in the first type of server refers to the information on the resources required to deploy all instances of the first type on the first server.

[0013] The third information of the first idle resource corresponding to each second server in the second type of server, wherein the first idle resource refers to the idle resource of the second server;

[0014] The fourth information of the second idle resources corresponding to each second server in the second type of server, wherein the second idle resources refer to the idle resources of the target type physical accelerator on the second server;

[0015] The remaining running time required for each of the first servers in the first type of servers refers to the remaining time required for all instances of the first type on the first server to complete their operation.

[0016] Optional, also includes:

[0017] Based on the virtualization accelerator resource utilization rate of the server with the first type of instance deployed, the first type of server and the second type of server are sorted in ascending order to obtain the first sort of each first server in the first type of server and the second sort of each second server in the second type of server.

[0018] The step of determining, based on the first parameter information, the first information of the first target server in the first type of servers capable of performing hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server capable of supporting hot migration, includes:

[0019] Based on the first sorting, the second sorting, and the first parameter information, the first information of the first target server in the first type of server that can perform hot migration and the second information of the second target server in the second type of server that can support hot migration are determined in sequence.

[0020] Optionally, performing the hot migration operation for the first category of instances based on the first information and the second information includes:

[0021] If the first total power consumption is less than the second total power consumption, a hot migration operation is performed on the first category of instances based on the first information and the second information; the first total power consumption refers to the target total power consumption after the hot migration operation is performed, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before the hot migration operation is performed.

[0022] Alternatively, upon receiving an execution instruction for a target situation, a hot migration operation for the first category instance is performed based on the first and second information; the target situation includes: the first total power consumption being greater than or equal to the second total power consumption, and the third total power consumption being less than the fourth total power consumption; the third total power consumption refers to the total power consumption corresponding to all the target category physical accelerators associated with the hot migration operation after the hot migration operation is performed; the fourth total power consumption refers to the total power consumption corresponding to the target category physical accelerators used for virtualization to obtain the first category instance before the hot migration operation is performed.

[0023] Optional, also includes:

[0024] The first method is used to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices;

[0025] And / or, record fifth information about the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0026] This application also provides a scheduling device for accelerator resources, including:

[0027] The first processing module is used to divide the servers with first category instances deployed into first category servers and second category servers; the first category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first category servers only have first category instances that support hot migration deployed; the second category servers have first category instances that do not support hot migration deployed.

[0028] The second processing module is configured to perform a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server in the first type of servers to one server in the second type of servers.

[0029] Optionally, performing the hot migration operation for at least one server in the first type of servers and at least one server in the second type of servers includes:

[0030] Based on the first parameter information, determine the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration;

[0031] Based on the first and second information, perform a hot migration operation for the first category of instances;

[0032] The first parameter information includes at least one of the following:

[0033] The resource information corresponding to each first server in the first type of server refers to the information on the resources required to deploy all instances of the first type on the first server.

[0034] The third information of the first idle resource corresponding to each second server in the second type of server, wherein the first idle resource refers to the idle resource of the second server;

[0035] The fourth information of the second idle resources corresponding to each second server in the second type of server, wherein the second idle resources refer to the idle resources of the target type physical accelerator on the second server;

[0036] The remaining running time required for each of the first servers in the first type of servers refers to the remaining time required for all instances of the first type on the first server to complete their operation.

[0037] Optional, also includes:

[0038] The third processing module is used to sort the first type of servers and the second type of servers in ascending order according to the virtualization accelerator resource utilization rate of the servers with the first type of instances deployed, so as to obtain the first sort of each first server in the first type of servers and the second sort of each second server in the second type of servers.

[0039] The step of determining, based on the first parameter information, the first information of the first target server in the first type of servers capable of performing hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server capable of supporting hot migration, includes:

[0040] Based on the first sorting, the second sorting, and the first parameter information, the first information of the first target server in the first type of server that can perform hot migration and the second information of the second target server in the second type of server that can support hot migration are determined in sequence.

[0041] Optionally, performing the hot migration operation for the first category of instances based on the first information and the second information includes:

[0042] If the first total power consumption is less than the second total power consumption, a hot migration operation is performed on the first category of instances based on the first information and the second information; the first total power consumption refers to the target total power consumption after the hot migration operation is performed, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before the hot migration operation is performed.

[0043] Alternatively, upon receiving an execution instruction for a target situation, a hot migration operation for the first category instance is performed based on the first and second information; the target situation includes: the first total power consumption being greater than or equal to the second total power consumption, and the third total power consumption being less than the fourth total power consumption; the third total power consumption refers to the total power consumption corresponding to all the target category physical accelerators associated with the hot migration operation after the hot migration operation is performed; the fourth total power consumption refers to the total power consumption corresponding to the target category physical accelerators used for virtualization to obtain the first category instance before the hot migration operation is performed.

[0044] Optional, also includes:

[0045] The fourth processing module is configured to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption using a first method; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices;

[0046] And / or, record fifth information about the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0047] This application also provides an accelerator resource scheduling device, including: a processor;

[0048] The processor is configured to divide servers with first-category instances deployed into first-category servers and second-category servers; the first-category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first-category servers only have first-category instances that support hot migration deployed; the second-category servers have first-category instances that do not support hot migration deployed.

[0049] Perform a hot migration operation on at least one server of the first type of servers and at least one server of the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server of the first type of servers to one server of the second type of servers.

[0050] Optionally, performing the hot migration operation for at least one server in the first type of servers and at least one server in the second type of servers includes:

[0051] Based on the first parameter information, determine the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration;

[0052] Based on the first and second information, perform a hot migration operation for the first category of instances;

[0053] The first parameter information includes at least one of the following:

[0054] The resource information corresponding to each first server in the first type of server refers to the information on the resources required to deploy all instances of the first type on the first server.

[0055] The third information of the first idle resource corresponding to each second server in the second type of server, wherein the first idle resource refers to the idle resource of the second server;

[0056] The fourth information of the second idle resources corresponding to each second server in the second type of server, wherein the second idle resources refer to the idle resources of the target type physical accelerator on the second server;

[0057] The remaining running time required for each of the first servers in the first type of servers refers to the remaining time required for all instances of the first type on the first server to complete their operation.

[0058] Optionally, the processor is further configured to:

[0059] Based on the virtualization accelerator resource utilization rate of the server with the first type of instance deployed, the first type of server and the second type of server are sorted in ascending order to obtain the first sort of each first server in the first type of server and the second sort of each second server in the second type of server.

[0060] The step of determining, based on the first parameter information, the first information of the first target server in the first type of servers capable of performing hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server capable of supporting hot migration, includes:

[0061] Based on the first sorting, the second sorting, and the first parameter information, the first information of the first target server in the first type of server that can perform hot migration and the second information of the second target server in the second type of server that can support hot migration are determined in sequence.

[0062] Optionally, performing the hot migration operation for the first category of instances based on the first information and the second information includes:

[0063] If the first total power consumption is less than the second total power consumption, a hot migration operation is performed on the first category of instances based on the first information and the second information; the first total power consumption refers to the target total power consumption after the hot migration operation is performed, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before the hot migration operation is performed.

[0064] Alternatively, upon receiving an execution instruction for a target situation, a hot migration operation for the first category instance is performed based on the first and second information; the target situation includes: the first total power consumption being greater than or equal to the second total power consumption, and the third total power consumption being less than the fourth total power consumption; the third total power consumption refers to the total power consumption corresponding to all the target category physical accelerators associated with the hot migration operation after the hot migration operation is performed; the fourth total power consumption refers to the total power consumption corresponding to the target category physical accelerators used for virtualization to obtain the first category instance before the hot migration operation is performed.

[0065] Optionally, the processor is further configured to:

[0066] The first method is used to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices through a transceiver;

[0067] And / or, record fifth information about the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0068] This application also provides an accelerator resource scheduling device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described accelerator resource scheduling method.

[0069] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described accelerator resource scheduling method.

[0070] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-described accelerator resource scheduling method.

[0071] The beneficial effects of the above technical solution in this application are as follows:

[0072] In the above scheme, the accelerator resource scheduling method divides servers with first-category instances into first-category servers and second-category servers. The first-category instances are instances obtained by virtualization based on the target category physical accelerators of the servers. Only first-category instances that support hot migration are deployed on the first-category servers. The second-category servers deploy first-category instances that do not support hot migration. Hot migration operations of the first-category instances are performed on at least one server in the first-category servers and at least one server in the second-category servers. The hot migration operation includes: hot migrating all first-category instances deployed on one server in the first-category servers to one server in the second-category servers. After multiple physical accelerators (i.e., physical accelerators used to obtain first-category instances by virtualization on servers with first-category instances) are virtualized and enabled, the virtual accelerators (i.e., first-category instances) that are actually occupied are hot migrated, so that the virtual accelerators occupy the physical accelerators of a few servers as much as possible. This allows the virtualization of idle physical accelerators to be deactivated, thereby reducing ineffective energy consumption, achieving energy saving, and solving the problem of a large amount of ineffective power consumption in the existing technology for the use of virtual accelerators. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the scheduling method for accelerator resources according to an embodiment of this application;

[0074] Figure 2 This is a schematic diagram illustrating the specific implementation process of the accelerator resource scheduling method in an embodiment of this application;

[0075] Figure 3 This is a schematic diagram of the accelerator resource scheduling device according to an embodiment of this application;

[0076] Figure 4 This is a schematic diagram of the accelerator resource scheduling device structure according to an embodiment of this application. Detailed Implementation

[0077] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0078] The following is a brief introduction to the relevant content of this plan.

[0079] In the computer field, MIG (Multi-Instance GPU) can refer to a GPU virtualization technology that allows a physical GPU (graphics processor) to be divided into multiple independent GPU instances. These GPU instances can be assigned to different virtual machines, containers, or users, thereby improving the resource utilization of the physical GPU and its multi-tenancy support capabilities.

[0080] Similarly, wireless cloudification utilizes accelerators such as FPGAs (Field-Programmable Arrays), GPUs, and DPUs (Data Processors), most of which support virtualization functions similar to MIG mentioned above. Generally, assuming a server contains type A accelerators with virtualization capabilities, one type A physical accelerator can be virtualized into several type A virtual accelerators, effectively improving the resource utilization and multi-tenant support capabilities of the type A physical accelerators. Specifically, taking GPUs as an example, a type A accelerator is a GPU. Based on MIG technology, one GPU can be virtualized into multiple virtual GPUs, called MIG; or, based on virtualization technology, one GPU can be virtualized into several virtual GPUs, called vGPU (virtual GPU).

[0081] Among them, physical accelerators such as GPUs (Graphics Processing Units) and FPGAs (Field Programmable Arrays) experience an increase in power consumption of about 5% after virtualization is enabled.

[0082] Based on the above, this application addresses the problem of significant inefficient power consumption in existing virtual accelerator usage schemes by providing a method for scheduling accelerator resources, such as... Figure 1 As shown, it includes:

[0083] Step 11: Divide the servers with first-category instances into first-category servers and second-category servers; the first-category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first-category servers only have first-category instances that support hot migration deployed; the second-category servers have first-category instances that do not support hot migration deployed.

[0084] Step 12: Perform a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers; the hot migration operation includes: hot migrating all first type of instances deployed on one server in the first type of servers to one server in the second type of servers.

[0085] The target category of physical accelerators may include a certain type of heterogeneous accelerator, such as GPU, FPGA or DPU, and / or, this solution can be applied to cloud platforms. Step 11 or step 12 may be triggered by the cloud platform itself or a third network element through external events (such as user instructions), internal events (such as low resource utilization), or periodic triggering, and is not limited here.

[0086] In addition, this solution can be understood as deploying a first-category instance using the target-category physical accelerator on the second server in the second-category server, thereby calling the resources of the target-category physical accelerator (on the second server in the second-category server) and freeing up the target-category physical accelerator on the first server in the first-category server. This frees up the virtualization of the target-category physical accelerator, thereby achieving energy saving.

[0087] The accelerator resource scheduling method provided in this application divides servers with first-category instances into first-category servers and second-category servers. The first-category instances are instances obtained by virtualization based on the target category physical accelerators of the servers. Only first-category instances that support hot migration are deployed on the first-category servers. The second-category servers deploy first-category instances that do not support hot migration. Hot migration operations of the first-category instances are performed on at least one server in the first-category servers and at least one server in the second-category servers. The hot migration operation includes: hot migrating all first-category instances deployed on one server in the first-category servers to one server in the second-category servers. After multiple physical accelerators (i.e., physical accelerators used to obtain first-category instances by virtualization on the servers with first-category instances) are virtualized and enabled, the virtual accelerators (i.e., first-category instances) that are actually occupied are hot migrated, so that the virtual accelerators occupy the physical accelerators of a few servers as much as possible. This allows the virtualization of idle physical accelerators to be deactivated, thereby reducing ineffective energy consumption, achieving energy saving, and solving the problem of a large amount of ineffective power consumption in the existing technology for the use of virtual accelerators.

[0088] The step of performing a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers includes: determining, based on first parameter information, first information of a first target server in the first type of servers capable of performing hot migration, and second information of a second target server in the second type of servers corresponding to the first target server capable of supporting hot migration; and performing a hot migration operation on the first type of instances based on the first and second information. The first parameter information includes at least one of the following: resource information corresponding to each first server in the first type of servers, where the resource information refers to information about the resources required to deploy all first type instances on the first server; third information of first idle resources corresponding to each second server in the second type of servers, where the first idle resources refer to the idle resources of the second server; fourth information of second idle resources corresponding to each second server in the second type of servers, where the second idle resources refer to the idle resources of the target type physics accelerator on the second server; and the remaining running time required for each first server in the first type of servers, where the remaining running time requires the remaining time required for all first type instances on the first server to complete their operation.

[0089] This allows for accurate execution of the hot migration operation. Specifically, instances of the first category on all physical accelerators of the first target service can be hot-migrated to the corresponding second target server; and / or, "remaining time required for completion" can be understood as the remaining time required for the operation to end, but is not limited to this.

[0090] Furthermore, the accelerator resource scheduling method further includes: sorting the first type of servers and the second type of servers in ascending order according to the virtualization accelerator resource utilization rate corresponding to the servers deploying the first type of instances, to obtain a first sorting of each first server in the first type of servers and a second sorting of each second server in the second type of servers; determining the first information of the first target server in the first type of servers capable of hot migration and the second information of the second target server in the second type of servers corresponding to the first target server capable of hot migration according to the first parameter information includes: determining the first information of the first target server in the first type of servers capable of hot migration and the second information of the second target server in the second type of servers corresponding to the first target server capable of hot migration in sequence according to the first sorting, the second sorting, and the first parameter information.

[0091] This allows for accurate acquisition of the first information of the first target server and the second information of the second target server. If multiple instances are bound together, meaning they must be deployed on the same server, these instances can be grouped together and sorted as an element, but this is not a limitation. Furthermore, in this solution, when determining the second target server corresponding to the first target server, it can be determined based on the aforementioned resource information and the idle resources of the second server in the second type of server. For example, the difference between the required resources of the determined first target server and the idle resources of the second target server is less than a resource threshold. In this approach, whether or not the aforementioned sorting operation is performed is optional and not limited here.

[0092] The step of determining, in sequence, the first information of the first target server capable of hot migration in the first type of servers, and the second information of the second target server in the second type of servers corresponding to the first target server, capable of supporting hot migration, based on the first sorting, the second sorting, and the first parameter information, includes: performing at least one first operation based on the first sorting to determine the first information and the second information; wherein the first operation includes: performing a hot migration judgment operation on the currently to-be-judged server in the first type of servers to obtain a migration judgment result; the currently to-be-judged server refers to the first server in the first sorting that has not yet performed the hot migration judgment operation; the hot migration judgment operation includes: traversing and searching for the target server in the second type of servers according to the second sorting based on the resource information, the third information, and the fourth information. The first second server whose idle resources meet the resource requirements of the current server to be judged; if the first second server is found, it is determined whether the remaining running time required for the current server to be judged is greater than a first threshold; if the remaining running time required is greater than the first threshold, the current server to be judged is designated as the first target server, and the first second server is designated as the second target server; if the first second server is not found, the execution of the first operation is terminated; (if the remaining running time required is greater than the first threshold) the migration judgment result includes: information of the first target server, and information of the second target server; or (if the first second server is not found) the migration judgment result indicates that the current server to be judged should not perform hot migration.

[0093] This allows for a more specific acquisition of the aforementioned first and second information. The target idle resources can include, but are not limited to, server idle resources and accelerator idle resources. The phrase "terminating the execution of the first operation if the first second server is not found" can be understood as follows: if no corresponding second server is found for a given first server (i.e., no corresponding second target server is found), all subsequent search operations are terminated. This avoids useless searches; because the later the first server is ranked, the more deployment resources it requires. If the earlier first servers cannot meet the requirements, the later ones will be even less able to. Therefore, all subsequent search operations can be stopped directly. Based on this, the operation to determine the first and second information can be understood as follows: according to the first ranking, for each first server, perform: traverse each second server in the second ranking to find a second server that meets the resource requirements. If found, stop the search and obtain the aforementioned first and second information (continue searching for the next first server, i.e., update the current server to be judged, and return "traverse each second server in the second ranking to find a second server that meets the resource requirements"); if not found, stop all subsequent search operations, but this is not a limitation. Specifically, in the case where "the migration judgment result indicates that the current server to be judged will not perform hot migration", the first server ranked after the current server to be judged will also not perform hot migration, while the first server ranked before the current server to be judged may perform hot migration, but this is not a limitation.

[0094] The step of performing a hot migration operation on the first category instance based on the first information and the second information includes: (1) performing a hot migration operation on the first category instance based on the first information and the second information when the first total power consumption is less than the second total power consumption; the first total power consumption refers to the target total power consumption after performing the hot migration operation, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category instance is deployed before performing the hot migration operation; or, (2) performing a hot migration operation on the first category instance based on the first information and the second information when an execution instruction for a target situation is obtained; the target situation includes: the first total power consumption is greater than or equal to the second total power consumption, and the third total power consumption is less than the fourth total power consumption; the third total power consumption refers to the total power consumption of all target category physical accelerators related to the hot migration operation after performing the hot migration operation; the fourth total power consumption refers to the total power consumption of the target category physical accelerators used for virtualization to obtain the first category instance before performing the hot migration operation.

[0095] This further ensures energy-saving performance after hot migration. The target total power consumption can be understood as: the total power consumption of servers with first-category instances deployed before the hot migration operation, after the migration operation. Further, this total power consumption can be the total power consumption under a first scenario, which includes: target-category physical accelerators no longer having first-category instances deployed after the hot migration operation entering a low-power mode (e.g., enabling hibernation or sleep, or disabling virtualization configuration), and / or servers no longer having first-category instances deployed entering a low-power mode (e.g., enabling hibernation or sleep, or disabling virtualization configuration of target-category physical accelerators), etc.; and / or, the target scenario can be understood as: the total server power consumption increases compared to before the resource adjustment (i.e., hot migration), but the total accelerator resource power consumption decreases; and / or, the third total power consumption can be the total power consumption under a second scenario, which includes: target-category physical accelerators no longer having first-category instances deployed after the hot migration operation entering a low-power mode (e.g., enabling hibernation or sleep, or disabling virtualization configuration), etc., but is not limited to these scenarios.

[0096] Furthermore, the accelerator resource scheduling method further includes: using a first method to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices; and / or recording the fifth information of the first category instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0097] This allows for the acquisition of power consumption-related data and / or information about instances undergoing hot migration in multiple ways. "Obtaining data by receiving information from other devices" includes, for example, directly obtaining data from a fourth network element, but is not limited to this.

[0098] The following is an example illustrating the accelerator resource scheduling method provided in the embodiments of this application.

[0099] To address the aforementioned technical issues, this application provides a method for scheduling accelerator resources. Specifically, it can be implemented as a cloud platform accelerator resource scheduling method, which mainly involves: after multiple physical accelerators are virtualized and enabled, when certain conditions are met (such as a low actual total occupancy rate of physical accelerators), the actually occupied virtual accelerators can be hot-migrated to a few physical accelerators, while the idle physical accelerators can be de-virtualized to achieve energy saving (essentially, under the premise of unchanged total occupancy rate, virtualization instances are used as centrally as possible to achieve energy saving).

[0100] Specifically, this solution can be applied to cloud platform clusters with multiple heterogeneous accelerator hardware, optimizing resource management for the same type of heterogeneous accelerators (e.g., GPUs, FPGAs, or DPUs); for example... Figure 2 As shown, the following steps are used to achieve this:

[0101] 1. The second network element (such as a cloud platform) instructs the first network element (such as a server with an accelerator) to establish a virtual machine instance or container instance (hereinafter referred to as an instance) with virtualization accelerator resources; the second network element obtains the target total runtime t of the instance. Here, a server with an accelerator can also be referred to as a node with an accelerator, which is not limited here.

[0102] 2. The second network element in the cluster categorizes servers containing accelerator resources;

[0103] Specifically, within a cloud platform cluster, servers containing Class A physical accelerators (corresponding to the aforementioned target category of physical accelerators) are categorized (corresponding to the categorization of servers with Class A instances into Class I and Class II servers): servers occupied by non-hot-migratable instances are categorized as Class S (corresponding to Class II servers with Class A Class A instances that do not support hot migration), and servers containing only hot-migratable instances are categorized as Class R (corresponding to Class I servers with only Class A Class A instances that support hot migration). If multiple instances are bound together, meaning these instances must be deployed on the same server, these instances can be grouped as an instance group.

[0104] The S-class servers are sorted in ascending order based on their virtualization accelerator resource utilization, forming S1, S2…Sn (corresponding to the second sorting mentioned above). The R-class servers are also sorted in ascending order based on their virtualization accelerator resource utilization rate, forming R1, R2…Rn (corresponding to the first sorting mentioned above). Corresponding to the virtualization accelerator resource utilization rate of the servers deployed with the first-class instances, the first-class and second-class servers are sorted in ascending order respectively, resulting in the first sorting of each first-class server and the second sorting of each second-class server. Instance groups are sorted as a group (with an overall utilization rate) as an element, grouped together with other individual accelerators. "Virtualization accelerator resource utilization rate" can be understood as the proportion of physical accelerator resources used by an instance, i.e., the ratio between the physical accelerator resources used by the instance and the total physical accelerator resources; however, it is not limited to this. Correspondingly, instances will request a portion of the physical accelerator resources, which can be managed and recorded by the cloud platform; however, it is not limited to this either.

[0105] 3. The second or third network element triggers accelerator resource adjustment through events or periodically (corresponding to the execution of accelerator resource adjustment when the above specific conditions are met).

[0106] Specifically, cloud platforms or third-party network elements may trigger accelerator resource adjustments through external events (such as user instructions) or internal events (such as low resource utilization), or periodically. If it is event-triggered, it could be: the average utilization rate of accelerators in the cloud platform is higher than a certain threshold, or the power consumption of at least one accelerator is higher than a certain threshold; or the utilization rate of at least one accelerator is lower than a certain threshold; or the actual total occupancy rate of physical accelerators is lower than a certain threshold, etc.

[0107] Specific resource adjustment methods may include: searching for migrated target servers (i.e., second servers that support migration) based on the remaining accelerator resources, such as steps 4 to 6 below (performing migration between servers according to the corresponding steps):

[0108] 4. The second network element checks whether the instance corresponding to R1 can be migrated to the server where S1 is located. It mainly checks whether the server where S1 is located has enough idle resources (such as whether the server has enough idle resources and whether the accelerator has enough idle resources) to deploy the instance corresponding to R1.

[0109] 5. If the available resources meet the requirements, assuming the time threshold is T, the current time is t1, and the instance creation time is t0, if t - (t1 - t0) > T, the second network element can determine to perform pre-migration (and record the status after migration); otherwise, no operation is performed. Here, t - (t1 - t0) corresponds to the remaining running time required. If the available resources do not meet the requirements, the second network element can continue to check whether the instance corresponding to R1 can be migrated to the server where S2 is located, and re-execute resource and time checks, similar to steps 4 and 5 (except that S1 is updated to S2). This process continues until it is determined that R1 should be pre-migrated. If the instance on R1 still cannot be pre-migrated up to Sn, no migration is performed on R1, and the migration process ends. Once pre-migration for R1 is confirmed, checks and the above-mentioned related operations can be performed sequentially on R2 to Rn (similar to R1, and will not be repeated here). The migration process ends if any R-class server is detected as unable to be pre-migrated. After this step, you can directly perform the migration (and then proceed to step 7), or you can perform steps 6 and 7.

[0110] Steps 4+5 can correspond to the above-mentioned determination of the first target server in the first type of servers that can perform hot migration according to the first sorting, the second sorting, and the first parameter information, and the second target server in the second type of servers that can support hot migration corresponding to the first target server; and the execution of the hot migration operation of the first type of instance according to the first information and the second information.

[0111] 6. (Optional) After determining the relevant information for the pre-migration, the second network element evaluates the power consumption before and after the migration;

[0112] For example, after pre-migration, if any server among R1, R2...Rn has no service, it can enable low-power modes such as sleep mode; the second network element calculates the total power consumption (corresponding to the first total power consumption) of all servers involved in the above instance migration (specifically, servers that have performed migration-related operations for instances corresponding to Class A physical accelerators) on the cloud platform based on the pre-migration situation. If the total power consumption increases compared to before resource adjustment (corresponding to the first total power consumption being greater than or equal to the second total power consumption), the change in the total power consumption of accelerator resources is obtained (specifically, the change in the total power consumption of accelerator resources of Class A physical accelerators involved in the above instance migration is obtained). If the total power consumption of accelerators decreases (corresponding to the third total power consumption being less than the fourth total power consumption), the user is consulted to determine whether to migrate, and the consultation result is obtained. If the consultation result indicates execution, the relevant hot migration operation is performed; corresponding to the above, when an execution instruction for the target situation is obtained, the hot migration operation of the first category of instances is performed based on the first and second information.

[0113] If the total power consumption of the server is lower than before the resource adjustment, the instance migration will be performed directly (corresponding to the above-mentioned hot migration operation of the first type of instance based on the first information and the second information when the first total power consumption is less than the second total power consumption).

[0114] The power consumption data in this step can be obtained from the fourth network element or determined by the second network element itself (corresponding to the above-mentioned first method, obtaining at least one of the first total power consumption, second total power consumption, third total power consumption and fourth total power consumption; wherein, the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices).

[0115] In this step, if a migration is performed, the second network element can instruct the relevant server to perform the specific instance migration operation. The current method can be used, which will not be elaborated here.

[0116] 7. The second network element records the information of the migrated instance and the resource usage of the migrated instance; the fifth information corresponding to the first category instance that is hot-migrated according to the above records, the fifth information including: identity information and / or resource usage information.

[0117] Among them, the third and fourth network elements mentioned above do not restrict the specific equipment type.

[0118] Therefore, this solution can reduce the overall power consumption of the accelerator by migrating cloud instances. Furthermore, this solution demonstrates its advantages more readily in scenarios with larger cluster sizes.

[0119] This application also provides a scheduling device for accelerator resources, such as... Figure 3 As shown, it includes:

[0120] The first processing module 31 is used to divide the servers with first category instances into first category servers and second category servers; the first category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first category servers only have first category instances that support hot migration deployed; the second category servers have first category instances that do not support hot migration deployed.

[0121] The second processing module 32 is configured to perform a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers; the hot migration operation includes: hot migrating all first type of instances deployed on one server in the first type of servers to one server in the second type of servers.

[0122] The accelerator resource scheduling device provided in this application divides servers with first-category instances into first-category servers and second-category servers. The first-category instances are instances obtained by virtualization based on the target category physical accelerators of the servers. Only first-category instances that support hot migration are deployed on the first-category servers. The second-category servers deploy first-category instances that do not support hot migration. Hot migration operation of the first-category instances is performed on at least one server in the first-category servers and at least one server in the second-category servers. The hot migration operation includes: hot migrating all first-category instances deployed on one server in the first-category servers to one server in the second-category servers. It can support multiple physical accelerators (i.e., physical accelerators used to obtain first-category instances by virtualization on servers with first-category instances) to be virtualized and enabled, and then perform hot migration operation on the virtual accelerators (i.e., first-category instances) that are actually occupied, so that the virtual accelerators occupy the physical accelerators of a few servers as much as possible, thereby canceling the virtualization of idle physical accelerators, reducing ineffective energy consumption, achieving energy saving, and solving the problem of a large amount of ineffective power consumption in the existing technology for the use of virtual accelerators.

[0123] The step of performing a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers includes: determining, based on first parameter information, first information of a first target server in the first type of servers capable of performing hot migration, and second information of a second target server in the second type of servers corresponding to the first target server capable of supporting hot migration; and performing a hot migration operation on the first type of instances based on the first and second information. The first parameter information includes at least one of the following: resource information corresponding to each first server in the first type of servers, where the resource information refers to information about the resources required to deploy all first type instances on the first server; third information of first idle resources corresponding to each second server in the second type of servers, where the first idle resources refer to the idle resources of the second server; fourth information of second idle resources corresponding to each second server in the second type of servers, where the second idle resources refer to the idle resources of the target type physics accelerator on the second server; and the remaining running time required for each first server in the first type of servers, where the remaining running time requires the remaining time required for all first type instances on the first server to complete their operation.

[0124] Furthermore, the accelerator resource scheduling device further includes: a third processing module, used to sort the first type of servers and the second type of servers in ascending order according to the virtualization accelerator resource utilization rate corresponding to the servers deployed with the first type of instances, to obtain a first sort of each first server in the first type of servers and a second sort of each second server in the second type of servers; the step of determining the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration, according to the first parameter information, includes: determining the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration, in sequence according to the first sort, the second sort, and the first parameter information.

[0125] The step of performing a hot migration operation on the first category of instances based on the first and second information includes: performing a hot migration operation on the first category of instances based on the first and second information when the first total power consumption is less than the second total power consumption; the first total power consumption refers to the target total power consumption after performing the hot migration operation, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before performing the hot migration operation; or, performing a hot migration operation on the first category of instances based on the first and second information when an execution instruction for a target situation is obtained; the target situation includes: the first total power consumption is greater than or equal to the second total power consumption, and the third total power consumption is less than the fourth total power consumption; the third total power consumption refers to the total power consumption of all target category physical accelerators related to the hot migration operation after performing the hot migration operation; the fourth total power consumption refers to the total power consumption of the target category physical accelerators used for virtualization to obtain the first category of instances before performing the hot migration operation.

[0126] Furthermore, the accelerator resource scheduling device further includes: a fourth processing module, used to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption using a first method; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices; and / or recording the fifth information of the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0127] The implementation embodiments of the above-mentioned accelerator resource scheduling method are all applicable to the embodiments of the accelerator resource scheduling device, and can achieve the same technical effect.

[0128] This application also provides a scheduling device for accelerator resources, such as... Figure 4 As shown, it includes: processor 41;

[0129] The processor 41 is used to divide the server with the first category of instances into a first category of servers and a second category of servers; the first category of instances are instances obtained by virtualization based on the target category of physical accelerators of the server; the first category of servers only have first category of instances that support hot migration deployed; the second category of servers have first category of instances that do not support hot migration deployed.

[0130] Perform a hot migration operation on at least one server of the first type of servers and at least one server of the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server of the first type of servers to one server of the second type of servers.

[0131] In this embodiment of the application, the scheduling device may further include a transceiver 42 capable of communicating with the processor 41, but is not limited thereto.

[0132] The accelerator resource scheduling device provided in this application divides servers with first-category instances into first-category servers and second-category servers. The first-category instances are instances obtained by virtualization based on the target category physical accelerators of the servers. Only first-category instances that support hot migration are deployed on the first-category servers. The second-category servers deploy first-category instances that do not support hot migration. Hot migration operations of the first-category instances are performed on at least one server in the first-category servers and at least one server in the second-category servers. The hot migration operation includes: hot migrating all first-category instances deployed on one server in the first-category servers to one server in the second-category servers. It can support multiple physical accelerators (i.e., physical accelerators used to obtain first-category instances by virtualization on servers with first-category instances) to be virtualized and enabled, and then perform hot migration operations on the virtual accelerators (i.e., first-category instances) that are actually occupied, so that the virtual accelerators occupy the physical accelerators of a few servers as much as possible, thereby deactivating the virtualization of idle physical accelerators, reducing ineffective energy consumption, achieving energy saving, and solving the problem of a large amount of ineffective power consumption in the existing technology for the use of virtual accelerators.

[0133] The step of performing a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers includes: determining, based on first parameter information, first information of a first target server in the first type of servers capable of performing hot migration, and second information of a second target server in the second type of servers corresponding to the first target server capable of supporting hot migration; and performing a hot migration operation on the first type of instances based on the first and second information. The first parameter information includes at least one of the following: resource information corresponding to each first server in the first type of servers, where the resource information refers to information about the resources required to deploy all first type instances on the first server; third information of first idle resources corresponding to each second server in the second type of servers, where the first idle resources refer to the idle resources of the second server; fourth information of second idle resources corresponding to each second server in the second type of servers, where the second idle resources refer to the idle resources of the target type physics accelerator on the second server; and the remaining running time required for each first server in the first type of servers, where the remaining running time requires the remaining time required for all first type instances on the first server to complete their operation.

[0134] Furthermore, the processor is also configured to: sort the first type of servers and the second type of servers in ascending order according to the virtualization accelerator resource utilization rate corresponding to the servers deployed with the first type of instances, to obtain a first sorting of each first server in the first type of servers and a second sorting of each second server in the second type of servers; the step of determining the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration, according to the first parameter information, includes: determining the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration, in sequence according to the first sorting, the second sorting, and the first parameter information.

[0135] The step of performing a hot migration operation on the first category of instances based on the first and second information includes: performing a hot migration operation on the first category of instances based on the first and second information when the first total power consumption is less than the second total power consumption; the first total power consumption refers to the target total power consumption after performing the hot migration operation, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before performing the hot migration operation; or, performing a hot migration operation on the first category of instances based on the first and second information when an execution instruction for a target situation is obtained; the target situation includes: the first total power consumption is greater than or equal to the second total power consumption, and the third total power consumption is less than the fourth total power consumption; the third total power consumption refers to the total power consumption of all target category physical accelerators related to the hot migration operation after performing the hot migration operation; the fourth total power consumption refers to the total power consumption of the target category physical accelerators used for virtualization to obtain the first category of instances before performing the hot migration operation.

[0136] Furthermore, the processor is also configured to: acquire at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption using a first method; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices through a transceiver; and / or recording fifth information of the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

[0137] The implementation embodiments of the above-mentioned accelerator resource scheduling method are all applicable to the embodiments of the accelerator resource scheduling device, and can achieve the same technical effect.

[0138] This application also provides an accelerator resource scheduling device, including a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described accelerator resource scheduling method.

[0139] The implementation embodiments of the above-mentioned accelerator resource scheduling method are all applicable to the embodiments of the accelerator resource scheduling device, and can achieve the same technical effect.

[0140] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described accelerator resource scheduling method.

[0141] The implementation embodiments of the above-described accelerator resource scheduling method are all applicable to the embodiments of the readable storage medium and can achieve the same technical effect.

[0142] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described method embodiment for scheduling accelerator resources and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0143] It should be noted that many of the functional components described in this specification are referred to as modules in order to more specifically emphasize the independence of their implementation.

[0144] In this embodiment, the module can be implemented in software so that it can be executed by various types of processors. For example, an identified executable code module may include one or more physical or logical blocks of computer instructions, which may be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but may include different instructions stored in different bits, which, when logically combined, constitute the module and achieve the module's intended purpose.

[0145] In practice, an executable code module can be a single instruction or many instructions, and can even be distributed across multiple different code segments, different programs, and across multiple memory devices. Similarly, operational data can be identified within the module and can be implemented in any suitable form and organized within any suitable type of data structure. This operational data can be collected as a single dataset or distributed across different locations (including different storage devices), and can exist, at least in part, solely as electronic signals within the system or network.

[0146] When a module can be implemented using software, considering the current level of hardware technology, modules that can be implemented in software can be implemented using hardware circuits by those skilled in the art to achieve the corresponding functions, without considering cost. These hardware circuits include conventional very-large-scale integrated circuits (VLSI) or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules can also be implemented using programmable hardware devices, such as field-programmable gate arrays, programmable array logic, and programmable logic devices.

[0147] The above describes the preferred embodiments of this application. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for scheduling accelerator resources, characterized in that, include: Servers with first-category instances deployed are divided into first-category servers and second-category servers; the first-category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first-category servers only have first-category instances that support hot migration deployed; the second-category servers have first-category instances that do not support hot migration deployed. Perform a hot migration operation on at least one server of the first type of servers and at least one server of the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server of the first type of servers to one server of the second type of servers.

2. The accelerator resource scheduling method according to claim 1, characterized in that, The hot migration operation of the first category instance for at least one server in the first category and at least one server in the second category includes: Based on the first parameter information, determine the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration; Based on the first and second information, perform a hot migration operation for the first category of instances; The first parameter information includes at least one of the following: The resource information corresponding to each first server in the first type of server refers to the information on the resources required to deploy all instances of the first type on the first server. The third information of the first idle resource corresponding to each second server in the second type of server, wherein the first idle resource refers to the idle resource of the second server; The fourth information of the second idle resources corresponding to each second server in the second type of server, wherein the second idle resources refer to the idle resources of the target type physical accelerator on the second server; The remaining running time required for each of the first servers in the first type of servers refers to the remaining time required for all instances of the first type on the first server to complete their operation.

3. The accelerator resource scheduling method according to claim 2, characterized in that, Also includes: Based on the virtualization accelerator resource utilization rate of the server with the first type of instance deployed, the first type of server and the second type of server are sorted in ascending order to obtain the first sort of each first server in the first type of server and the second sort of each second server in the second type of server. The step of determining, based on the first parameter information, the first information of the first target server in the first type of servers capable of performing hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server capable of supporting hot migration, includes: Based on the first sorting, the second sorting, and the first parameter information, the first information of the first target server in the first type of server that can perform hot migration and the second information of the second target server in the second type of server that can support hot migration are determined in sequence.

4. The accelerator resource scheduling method according to claim 2, characterized in that, The step of performing a hot migration operation on the first category of instances based on the first information and the second information includes: If the first total power consumption is less than the second total power consumption, a hot migration operation is performed on the first category of instances based on the first information and the second information; the first total power consumption refers to the target total power consumption after the hot migration operation is performed, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before the hot migration operation is performed. Alternatively, upon receiving an execution instruction for a target situation, a hot migration operation for the first category instance is performed based on the first and second information; the target situation includes: the first total power consumption being greater than or equal to the second total power consumption, and the third total power consumption being less than the fourth total power consumption; the third total power consumption refers to the total power consumption corresponding to all the target category physical accelerators associated with the hot migration operation after the hot migration operation is performed; the fourth total power consumption refers to the total power consumption corresponding to the target category physical accelerators used for virtualization to obtain the first category instance before the hot migration operation is performed.

5. The accelerator resource scheduling method according to claim 4, characterized in that, Also includes: The first method is used to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices; And / or, record fifth information about the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

6. A scheduling device for accelerator resources, characterized in that, include: The first processing module is used to divide the servers with first category instances deployed into first category servers and second category servers; the first category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first category servers only have first category instances that support hot migration deployed; the second category servers have first category instances that do not support hot migration deployed. The second processing module is configured to perform a hot migration operation on at least one server in the first type of servers and at least one server in the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server in the first type of servers to one server in the second type of servers.

7. The accelerator resource scheduling device according to claim 6, characterized in that, The hot migration operation of the first category instance for at least one server in the first category and at least one server in the second category includes: Based on the first parameter information, determine the first information of the first target server in the first type of servers that can perform hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server that can support hot migration; Based on the first and second information, perform a hot migration operation for the first category of instances; The first parameter information includes at least one of the following: The resource information corresponding to each first server in the first type of server refers to the information on the resources required to deploy all instances of the first type on the first server. The third information of the first idle resource corresponding to each second server in the second type of server, wherein the first idle resource refers to the idle resource of the second server; The fourth information of the second idle resources corresponding to each second server in the second type of server, wherein the second idle resources refer to the idle resources of the target type physical accelerator on the second server; The remaining running time required for each of the first servers in the first type of servers refers to the remaining time required for all instances of the first type on the first server to complete their operation.

8. The accelerator resource scheduling device according to claim 7, characterized in that, Also includes: The third processing module is used to sort the first type of servers and the second type of servers in ascending order according to the virtualization accelerator resource utilization rate of the servers with the first type of instances deployed, so as to obtain the first sort of each first server in the first type of servers and the second sort of each second server in the second type of servers. The step of determining, based on the first parameter information, the first information of the first target server in the first type of servers capable of performing hot migration, and the second information of the second target server in the second type of servers corresponding to the first target server capable of supporting hot migration, includes: Based on the first sorting, the second sorting, and the first parameter information, the first information of the first target server in the first type of server that can perform hot migration and the second information of the second target server in the second type of server that can support hot migration are determined in sequence.

9. The accelerator resource scheduling device according to claim 7, characterized in that, The step of performing a hot migration operation on the first category of instances based on the first information and the second information includes: If the first total power consumption is less than the second total power consumption, a hot migration operation is performed on the first category of instances based on the first information and the second information; the first total power consumption refers to the target total power consumption after the hot migration operation is performed, and the target total power consumption refers to the total power consumption of all servers related to the hot migration operation; the second total power consumption refers to the total power consumption of the servers on which the first category of instances are deployed before the hot migration operation is performed. Alternatively, upon receiving an execution instruction for a target situation, a hot migration operation for the first category instance is performed based on the first and second information; the target situation includes: the first total power consumption being greater than or equal to the second total power consumption, and the third total power consumption being less than the fourth total power consumption; the third total power consumption refers to the total power consumption corresponding to all the target category physical accelerators associated with the hot migration operation after the hot migration operation is performed; the fourth total power consumption refers to the total power consumption corresponding to the target category physical accelerators used for virtualization to obtain the first category instance before the hot migration operation is performed.

10. The accelerator resource scheduling device according to claim 9, characterized in that, Also includes: The fourth processing module is configured to obtain at least one of the first total power consumption, the second total power consumption, the third total power consumption, and the fourth total power consumption using a first method; wherein the first method includes: obtaining it based on the first information and the second information; and / or obtaining it by receiving information sent by other devices; And / or, record fifth information about the first category of instances undergoing hot migration, the fifth information including: identity information and / or resource usage information.

11. A scheduling device for accelerator resources, characterized in that, include: processor; The processor is configured to divide servers with first-category instances deployed into first-category servers and second-category servers; the first-category instances are instances obtained by virtualization based on the target category physical accelerator of the server; the first-category servers only have first-category instances that support hot migration deployed; the second-category servers have first-category instances that do not support hot migration deployed. Perform a hot migration operation on at least one server of the first type of servers and at least one server of the second type of servers; the hot migration operation includes: hot migrating all first-category instances deployed on one server of the first type of servers to one server of the second type of servers.

12. A scheduling device for accelerator resources, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the accelerator resource scheduling method as described in any one of claims 1 to 5.

13. A readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the accelerator resource scheduling method as described in any one of claims 1 to 5.

14. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the accelerator resource scheduling method as described in any one of claims 1 to 5.

Images