A method for allocating consensus of multi-copy computing power slices based on shared resource identification inheritance

Abstract

This invention belongs to the field of divisible computing resource management technology, and particularly relates to a method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance. The method includes the following steps: When allocating computing power slices for the first time for a multi-replica training task, a shared resource identifier is generated for each replica participating in resource sharing by the shared identifier management system. During replica failure reconstruction, this invention prioritizes inheriting the original shared resource identifier, without modifying the normal replica slice mapping, effectively reducing fluctuations in sharing relationships and training performance disturbances. Only local rebinding adjustments are made to the failed replicas, significantly reducing the number of times the entire computing power slice is re-allocated, reducing cluster scheduling overhead. Resource reclamation scheduling takes into account device topology locality, fairness in multi-replica resource allocation, and computing power resource utilization. Simultaneously, through identifier inheritance, local rebinding, and a controlled grouping mechanism, the growth of GPU memory fragmentation is slowed down, improving the long-term stability and reliability of the divisible computing power resource pool.

Description

Technical Field

[0001] This invention belongs to the field of divisible computing resource management technology, and in particular relates to a method for consistent allocation of multi-replica computing power slices based on the inheritance of shared resource identifiers. Background Technology

[0002] Currently, to improve the utilization of GPU cloud resources, a single physical device is often divided into multiple independent computing capacity slices for multiple training replicas to share. However, in existing technologies, the allocation of computing capacity slices in multi-replica training scenarios lacks a unified standard, and the coordination among various functional units (such as slice allocation, replica management, and fault handling) is insufficient. Furthermore, when a replica fails and needs to be rebuilt, it is often necessary to fully reallocate the computing capacity slices of all replicas, which can easily disrupt the original shared resource relationships, leading to training performance fluctuations and increased memory fragmentation. At the same time, operations such as resource reclamation and capacity compensation lack a unified execution standard, resulting in poor coordination between the operations of various functional units. In addition, existing solutions cannot effectively inherit the original shared resource identifiers and slice allocation relationships when replicas fail and need to be rebuilt, which can easily lead to resource waste and allocation chaos, making it difficult to meet the requirements of multi-replica training tasks for consistency, stability, and efficiency in computing capacity allocation.

[0003] There is an urgent need for improvement, so we propose a consistent allocation method for multi-replica computing power slices based on the inheritance of shared resource identifiers. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a method for consistent allocation of multi-replica computing power slices based on the inheritance of shared resource identifiers.

[0005] In view of this, the present invention provides a method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance, comprising the following steps: S1. When allocating computing power slices for the first time for a multi-replica training task, the shared identifier management generates a shared resource identifier for each replica participating in resource sharing and establishes the association between the replica and the shared resource identifier. S2. The running replica records the parent replica identifier, inheritance generation, and consumed capacity of the computing power slice for each replica. S3. When any replica fails and needs to be rebuilt, the failure status of the replica is identified by the fault detection in the reconstruction and compensation process and marked as a replica to be rebuilt. The shared resource identifier corresponding to the parent replica of the failed replica is retrieved. S4. Determine whether the original shared computing power slice corresponding to the faulty copy is available by inheritance judgment in reconstruction and compensation. If it is available, the new copy after reconstruction directly inherits the original shared resource identifier. S5. If the original shared computing power slice is unavailable, the local rebinding in the reconstruction and compensation process will only perform the local rebinding operation on the faulty replica subset, and the inheritance generation of the faulty replica will be updated synchronously. S6. Based on the consumed capacity receipts of each replica, selectively perform partial recovery of computing power slices, capacity compensation, or resource rebalancing operations by status and recovery. S7. Monitor the length of the replica inheritance chain in real time. When the length of the inheritance chain exceeds the preset threshold, trigger the resource consolidation operation of the computing power slice of the entire group of replicas.

[0006] Furthermore, in step S1, the slice allocator divides a single physical device into multiple independent and allocable computing capacity slices, which are stored in the computing capacity slice pool. At the same time, the replica group generation unit splits and generates multiple training replicas according to the training task resource specifications, the number of parallel replicas, and the minimum video memory requirements. Then, the shared identifier management configures a unified shared resource identifier for replicas that have the same resource sharing topology.

[0007] Furthermore, in step S2, when recording the information corresponding to each replica, the running replica synchronously records the physical device number, slice number, slice capacity, replica activity level and real-time running status of each replica, and constructs a resource association management table, storing all associated data in the resource association management table, and the resource association management table is stored in the running replica.

[0008] Furthermore, in step S3, after identifying the failure status of the replica through fault detection of reconstruction and compensation and marking it as a replica to be rebuilt, the association data of the parent replica identifier and the shared resource identifier of the replica to be rebuilt, which is pre-stored in the running replica, is retrieved.

[0009] Furthermore, the specific conditions for determining the availability of the original shared computing power slice in step S4 include: the original shared slice is not occupied by other services, the physical device to which the slice belongs is in good health, and the remaining capacity of the physical device and the network topology location both meet the requirements for running the training task. The determination operation is performed by the inheritance determination of reconstruction and compensation.

[0010] Furthermore, the local rebinding operation in step S5 specifically includes: firstly searching for continuous available capacity in the adjacent slice area of ​​the original physical device; if the original physical device does not have matching continuous capacity, then expanding the search range to other physical device slice areas in the topology proximity to perform capacity matching; the local rebinding operation is performed by the reconstruction and compensation local rebinding.

[0011] Furthermore, during the execution of the partial rebinding operation, for replicas with high capacity usage, low operational activity, and weak short-term computing needs, a partial recovery of computing power slices or a reduction of resource quotas is performed. This operation frees up continuous slice capacity for the reconstruction of faulty replicas. The partial recovery of computing power slices or reduction of resource quotas is synchronously fed back to the computing power slice pool of the functional module group to update the resource status of the computing power slice pool.

[0012] Furthermore, in step S6, the capacity compensation refers to the capacity compensation module of the status and reclamation, which dynamically replenishes the resource quota based on the actual consumed capacity of the replica. The resource rebalancing refers to the reclamation strategy module of the status and reclamation, which performs operations that take into account both the fairness of multi-replica computing power resource occupation and the locality of resource deployment. The capacity compensation and resource rebalancing operations are jointly executed by the capacity compensation module and the reclamation strategy module in the status and reclamation.

[0013] Furthermore, the above-mentioned multi-replica computing power slice consistency allocation method based on shared resource identifier inheritance is adapted to the four-replica training task scenario, specifically by splitting the training task into four parallel training replicas.

[0014] Furthermore, the computing power slices are stored in a computing power slice pool, specifically including slice A, slice B, slice C, and slice D.

[0015] The beneficial effects of this invention are: 1. This invention prioritizes the inheritance of the original shared resource identifier during replica failure reconstruction, without altering the computing power slice mapping relationship of normal replicas. This effectively avoids replica reconstruction disrupting the resource sharing topology among multiple replicas, significantly reducing fluctuations in sharing relationships and communication disturbances and performance jitters during the training process. 2. This invention abandons the traditional mode of unified reallocation of the entire replica group, and only performs local rebinding and resource adjustment for the faulty replica subset. It only triggers the whole group reorganization after the inheritance chain exceeds the limit, which significantly reduces the number of times the computing power slice is relocated and reduces the system overhead of cluster resource scheduling. 3. In the process of resource recycling and scheduling, this invention prioritizes the principle of physical device topology locality, and ensures the fairness of multi-replica resource occupation through capacity compensation and resource rebalancing mechanisms. At the same time, it reasonably recycles and relocates low-activity and high-occupancy resources, effectively revitalizes idle computing power, and comprehensively improves the overall utilization rate of computing power resources. 4. This invention avoids the rapid accumulation of memory fragmentation caused by frequent resource shuffling through a combination mechanism of identifier inheritance, local rebinding, and controlled grouping, effectively slowing down the fragmentation growth rate and significantly improving the long-term stability and reliability of the divisible computing power resource pool. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating a method for consistent allocation of multi-replica computing power slices based on the inheritance of shared resource identifiers proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0018] In the description of this application, it should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. For ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0019] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and are not limited in number; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0020] It should be noted that in the description of this application, the directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this application. The directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0021] It should be noted that, in this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0022] Reference Figure 1 A method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance includes the following steps: S1. When allocating computing power slices for the first time for a multi-replica training task, the shared identifier management generates a shared resource identifier for each replica participating in resource sharing and establishes the association between the replica and the shared resource identifier. S2. The running replica records the parent replica identifier, inheritance generation, and consumed capacity of the computing power slice for each replica. S3. When any replica fails and needs to be rebuilt, the failure status of the replica is identified by the fault detection in the reconstruction and compensation process and marked as a replica to be rebuilt. The shared resource identifier corresponding to the parent replica of the failed replica is retrieved. S4. Determine whether the original shared computing power slice corresponding to the faulty copy is available by inheritance judgment in reconstruction and compensation. If it is available, the new copy after reconstruction directly inherits the original shared resource identifier. S5. If the original shared computing power slice is unavailable, the local rebinding in the reconstruction and compensation process will only perform the local rebinding operation on the faulty replica subset, and the inheritance generation of the faulty replica will be updated synchronously. S6. Based on the consumed capacity receipts of each replica, the status and recycling selectively perform partial recycling of computing power slices, capacity compensation, or resource rebalancing operations. During the resource recycling process, the locality of resource deployment, the fairness of multi-replica allocation, and the overall computing power resource utilization rate are taken into account. S7. Real-time monitoring of replica inheritance chain length. When the inheritance chain length exceeds the preset threshold, trigger the resource reorganization operation of the entire group of multi-replica computing power slices. This can control the scheduling frequency, reduce the redistribution of the entire group, effectively slow down the growth of memory fragmentation, and improve the overall operational stability of the divisible computing power resource pool.

[0023] This application standardizes the entire process of shared resource identifier generation, replica information recording, fault detection and handling, and resource recovery and compensation, avoiding resource shuffling and training performance jitter during replica reconstruction. It also establishes an inheritance chain monitoring mechanism, reducing memory fragmentation and resource waste, ensuring the stability and consistency of multi-replica training tasks, and guaranteeing the feasibility and reproducibility of the technical solution. This solves the problems of chaotic steps and lack of unified standards in existing technologies. Furthermore, the application allows for custom configuration of the inheritance chain preset threshold based on the size of the computing power resource pool and the number of replica tasks, adapting to different computing power cluster deployment scenarios. After replica reconstruction inherits the original sharing relationship, there is no need to adjust the slice mapping relationship of other normally operating replicas, significantly reducing the sharing relationship jitter caused by replica reconstruction and effectively reducing communication disturbances and resource fluctuations during the training process.

[0024] In the example of this application, in step S1, the slice allocator divides a single physical device into multiple independent and allocable computing capacity slices, which are stored in the computing capacity slice pool. At the same time, the replica group generation unit splits and generates multiple training replicas according to the training task resource specifications, the number of parallel replicas, and the minimum video memory requirements. Then, the shared identifier management configures a unified shared resource identifier for replicas with the same resource sharing topology. The underlying physical graphics processor is used as the computing power carrying hardware, providing hardware foundation support for computing power slice division and multi-replica training task operation.

[0025] As a preferred example of the present invention, by clarifying the role of the slice allocator and the replica generation logic, the source and storage location of computing capacity slices are clearly defined. At the same time, the process of replica generation and shared resource identifier configuration is standardized, avoiding the problem of slice allocation and replica usage being disconnected, improving the rationality of resource allocation, providing a foundation for subsequent replica operation and fault handling, and reducing resource idleness and allocation chaos.

[0026] In the example of this application, in step S2, when the running replica records the information corresponding to each replica, it synchronously records the physical device number, slice number, slice capacity, replica activity level and real-time running status of each replica, and constructs a resource association management table. All related data are centrally stored in the resource association management table, and the resource association management table is stored in the running replica. The resource association management table fully collects all related fields such as replica identifier, parent replica identifier, shared resource identifier, and inheritance generation, so as to realize unified control and traceability of data in all dimensions.

[0027] As a preferred example of the present invention, a complete resource association management system is constructed by synchronously recording the physical device information, operating status and associated data of the replicas. This facilitates subsequent fault diagnosis, data traceability and replica operation monitoring. At the same time, it standardizes the storage method of the resource association management table, ensuring that all replica-related data is searchable and traceable, thereby improving the maintainability and standardization of the technical solution.

[0028] In the example of this application, in step S3, after identifying the failure status of the replica through the fault detection of reconstruction and compensation and marking it as a replica to be rebuilt, the association data of the parent replica identifier and the shared resource identifier pre-stored in the running replica is retrieved. The replica failure causes cover three typical scenarios: node restart, abnormal container exit, and host hardware failure, which fully cover the replica failure scenarios in the actual production environment.

[0029] As a preferred example of the present invention, by explicitly storing the associated data of the replica to be rebuilt in the running replica, it is ensured that the parent replica identifier and shared resource identifier can be quickly retrieved during fault reconstruction, thereby improving fault handling efficiency, avoiding data search chaos, ensuring the smoothness of the replica reconstruction process, and reducing the time spent on fault handling.

[0030] In the example of this application, the specific conditions for determining the availability of the original shared computing power slice in step S4 include: the original shared slice is not occupied by other services, the physical device to which the slice belongs is in good health, and the remaining capacity and network topology of the physical device meet the requirements for running the training task. The determination operation is performed by the inheritance determination of reconstruction and compensation. The new replica directly uses the original shared resource architecture, which reduces the sharing relationship jitter caused by replica reconstruction from the root. After the partial rebinding is completed, the updated slice number, shared resource identifier mapping relationship, inheritance generation and capacity status of the faulty replica are synchronously written into the resource management table, so as to retain data basis for the consistency allocation logic of subsequent replica reconstruction and reuse.

[0031] As a preferred example of the present invention, by refining the judgment conditions, the available scope of the original shared slice is made clearer, avoiding ambiguity in judgment. At the same time, the execution responsibility of inheritance judgment is clarified, ensuring the uniformity and standardization of judgment operation, providing a clear basis for resource matching in subsequent copy reconstruction, and reducing the ambiguity of the technical solution.

[0032] In the example of this application, the local rebinding operation in step S5 specifically includes: firstly searching for continuous available capacity in the adjacent slice area of ​​the original physical device; if the original physical device does not have matching continuous capacity, then expanding the search range to other physical device slice areas in the topology proximity to perform capacity matching; the local rebinding operation is performed by the local rebinding of reconstruction and compensation.

[0033] As a preferred example of the present invention, by clearly defining that the partial rebinding operation is performed by the partial rebinding of reconstruction and compensation, the operation process is standardized, ensuring that the partial rebinding process is reproducible and traceable, avoiding the problem of inconsistent rebinding operation standards for different copies, and providing an execution basis for resource adjustments in the subsequent partial rebinding process.

[0034] In the example of this application, during the execution of the partial rebinding operation, for replicas with high capacity usage, low operational activity, and weak short-term computing needs, a partial recovery of computing power slices or a reduction of resource quotas is performed. This operation frees up continuous slice capacity for the reconstruction of faulty replicas. The partial recovery of computing power slices or reduction of resource quotas is synchronously fed back to the computing power slice pool of the functional module group to update the resource status of the computing power slice pool, effectively reducing disordered resource arrangement, delaying the growth of memory fragmentation, and improving the long-term stability of the computing power resource pool.

[0035] As a preferred example of the present invention, by adjusting the resources of high-occupancy, low-activity replicas, available computing power capacity is freed up, which not only ensures the reconstruction needs of faulty replicas, but also realizes the rational use of resources. At the same time, the operation is fed back to the computing power slicing pool to ensure that the resource status is updated in real time, reducing memory fragmentation and resource waste, improving the utilization rate of computing power resources, and realizing dynamic optimization of resources.

[0036] In the example of this application, in step S6, the capacity compensation refers to the capacity compensation module of the status and reclamation, which dynamically replenishes the resource quota based on the actual consumed capacity of the replica. The resource rebalancing refers to the reclamation strategy module of the status and reclamation, which performs the operation that takes into account the fairness of the multi-replica computing power resource occupation and the locality of resource deployment. The capacity compensation and resource rebalancing operations are jointly executed by the capacity compensation module and the reclamation strategy module in the status and reclamation.

[0037] As a preferred example of the present invention, by clarifying the executing entities and collaborative relationships of the two operations, operational confusion is avoided, while the execution process is standardized, ensuring that capacity compensation and resource rebalancing operations are feasible and reproducible, taking into account the fairness and rationality of resource allocation, and improving the rigor of the technical solution.

[0038] In the example of this application, the above-mentioned multi-replica computing power slice consistency allocation method based on shared resource identifier inheritance is adapted to the four-replica training task scenario. Specifically, the training task is split into four parallel training replicas, and two physical graphics processors that meet the requirements of video memory specifications are selected as hardware support. Each physical graphics processor is divided into several computing power slices for the four training replicas to share and occupy. The computing power slices are stored in the computing power slice pool, specifically including slice A, slice B, slice C, and slice D.

[0039] As a preferred example of the present invention, by limiting the training scenario to four replicas, the application scenario of the technical solution is made more specific, and at the same time, clear scenario support is provided for the use of subsequent computing power slicing, ensuring that the technical solution is accurately matched with the actual application needs, and improving the practicality and pertinence of the solution.

[0040] In the example of this application, the computing power slices are stored in a computing power slice pool, specifically including slice A, slice B, slice C, and slice D. Two physical graphics processors that meet the video memory specifications are selected as hardware support, and each physical graphics processor is divided into several computing power slices for four training replicas to share and occupy.

[0041] As a preferred example of the present invention, by clearly defining the storage location and usage scenarios of the slices, the role of the computing power slices becomes clearer, ensuring that the slice resources can support the operation of the replicas, further improving the completeness of the technical solution and enhancing the rationality and pertinence of resource utilization.

[0042] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance, characterized in that, Includes the following steps: S1. When allocating computing power slices for the first time for a multi-replica training task, the shared identifier management generates a shared resource identifier for each replica participating in resource sharing and establishes the association between the replica and the shared resource identifier. S2. The running replica records the parent replica identifier, inheritance generation, and consumed capacity of the computing power slice for each replica. S3. When any replica fails and needs to be rebuilt, the failure status of the replica is identified by the fault detection in the reconstruction and compensation process and marked as a replica to be rebuilt. The shared resource identifier corresponding to the parent replica of the failed replica is retrieved. S4. Determine whether the original shared computing power slice corresponding to the faulty copy is available by inheritance judgment in reconstruction and compensation. If it is available, the new copy after reconstruction directly inherits the original shared resource identifier. S5. If the original shared computing power slice is unavailable, the local rebinding in the reconstruction and compensation process will only perform the local rebinding operation on the faulty replica subset, and the inheritance generation of the faulty replica will be updated synchronously. S6. Based on the consumed capacity receipts of each replica, selectively perform partial recovery of computing power slices, capacity compensation, or resource rebalancing operations by status and recovery. S7. Monitor the length of the replica inheritance chain in real time. When the length of the inheritance chain exceeds the preset threshold, trigger the resource consolidation operation of the computing power slice of the entire group of replicas.

2. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 1, characterized in that, In step S1, the slice allocator divides a single physical device into multiple independent and allocable computing capacity slices. The computing capacity slices are stored in the computing capacity slice pool. At the same time, the replica group generation unit splits and generates multiple training replicas according to the training task resource specifications, the number of parallel replicas, and the minimum video memory requirements. Then, the shared identifier management configures a unified shared resource identifier for replicas with the same resource sharing topology.

3. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 1, characterized in that, In step S2, when recording the information corresponding to each replica, the running replica synchronously records the physical device number, slice number, slice capacity, replica activity level and real-time running status of each replica, and constructs a resource association management table, storing all associated data in the resource association management table, and the resource association management table is stored in the running replica.

4. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 1, characterized in that, In step S3, after identifying the failure status of the replica through fault detection of reconstruction and compensation and marking it as a replica to be rebuilt, the association data of the parent replica identifier and the shared resource identifier of the replica to be rebuilt, which is pre-stored in the running replica, is retrieved.

5. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 1, characterized in that, The specific conditions for determining the availability of the original shared computing power slice in step S4 include: the original shared slice is not occupied by other services, the physical device to which the slice belongs is in good health, and the remaining capacity of the physical device and the network topology location both meet the requirements for running the training task. The determination operation is performed by the inheritance determination of reconstruction and compensation.

6. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 1, characterized in that, The local rebinding operation in step S5 specifically includes: firstly searching for continuous available capacity in the adjacent slice area of ​​the original physical device; if the original physical device does not have matching continuous capacity, then expanding the search range to other physical device slice areas in the topology proximity to perform capacity matching; the local rebinding operation is performed by the local rebinding of reconstruction and compensation.

7. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 6, characterized in that, During the execution of the partial rebinding operation, for replicas with high capacity usage, low operational activity, and weak short-term computing needs, a partial recovery of computing power slices or a reduction of resource quotas is performed. This operation frees up continuous slice capacity for the reconstruction of faulty replicas. The partial recovery of computing power slices or reduction of resource quotas is synchronously fed back to the computing power slice pool of the functional module group to update the resource status of the computing power slice pool.

8. The method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 7, characterized in that, In step S6, the capacity compensation refers to the capacity compensation module of the status and reclamation, which dynamically replenishes the resource quota based on the actual consumed capacity of the replica. The resource rebalancing refers to the reclamation strategy module of the status and reclamation, which performs operations that take into account both the fairness of the computing power resource occupation of multiple replicas and the locality of resource deployment. The capacity compensation and resource rebalancing operations are jointly executed by the capacity compensation module and the reclamation strategy module in the status and reclamation.

9. A method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 8, characterized in that, The above-mentioned multi-replica computing power slice consistency allocation method based on shared resource identifier inheritance is adapted to the four-replica training task scenario, specifically by splitting the training task into four parallel training replicas.

10. A method for consistent allocation of multi-replica computing power slices based on shared resource identifier inheritance according to claim 9, characterized in that, The computing power slices are stored in a computing power slice pool, specifically including slice A, slice B, slice C, and slice D.

Images