Method and device for generating large model training task migration strategy across intelligence computing centers

Abstract

The present disclosure provides a method for generating a large model training task migration strategy across intelligent computing centers, which comprises the following steps: firstly, extracting large model training tasks that can be migrated across intelligent computing centers; then, combining the state of intelligent computing centers and elastic optical networks to calculate the comprehensive migration cost of each task migrating to different nodes; according to a preset condition, matching a target node to generate an initial candidate set; then, screening tasks, calculating migration gains and sorting to form a task group; after executing migration and occupying resources, re-evaluating and iterating the task group, repeatedly scheduling, and finally outputting a task migration strategy across intelligent computing centers. The technical solution provided by the present disclosure can ensure the quality of service of training tasks, reduce the task completion time and non-renewable energy consumption, and improve the system resource utilization efficiency.

Description

Technical Field

[0001] This disclosure relates to the field of large model training task deployment technology, and in particular to a method and apparatus for generating a large model training task migration strategy across intelligent computing centers. Background Technology

[0002] This section is intended to provide background or context for the embodiments of this disclosure as set forth in the claims. The description herein is not intended to be a prior art simply because it is included in this section.

[0003] With the rapid development of large language models and generative artificial intelligence applications, the demand for computing infrastructure for large model training tasks continues to increase. Geographically distributed intelligent computing centers make it possible to schedule large model training tasks across regions. Dynamically migrating long-cycle training tasks from the current intelligent computing center to a center with stronger computing power, shorter queuing times, or more abundant renewable energy could potentially shorten task completion time and reduce the consumption of non-renewable energy.

[0004] When migrating long-cycle, large-scale model training tasks, establishing a migration optical path in a flexible optical network must satisfy routing and resource allocation constraints. If the target intelligent computing center is selected only from the computing side or the energy side, migration decisions that seem advantageous in terms of computing or energy may not be feasible on the network side.

[0005] Existing green scheduling methods rarely make dynamic migration decisions for long-term, large-scale model training tasks in operation, and fail to jointly consider heterogeneous computing power, GPU resource status, task queue status, renewable energy supply, non-renewable energy consumption, and routing and resource constraints in elastic optical networks within the same framework. Summary of the Invention

[0006] In view of this, the purpose of this disclosure is to propose a technical solution for a strategy to migrate large model training tasks across intelligent computing centers, which at least to some extent solves one of the technical problems in related technologies.

[0007] To achieve the above objectives, the first aspect of this exemplary embodiment provides a method for generating a cross-computing center migration strategy for large model training tasks, comprising: obtaining multiple transferable large model training tasks from the multiple large model training tasks based on basic information of the multiple large model training tasks; calculating a comprehensive migration cost for each of the multiple transferable large model training tasks to the multiple available intelligent computing centers based on state information of the multiple available intelligent computing centers and elastic optical networks; and matching a corresponding target intelligent computing center from the multiple available intelligent computing centers for each of the multiple transferable large model training tasks based on preset standards and the comprehensive migration cost. The system obtains an initial set of migration candidates; based on the initial set of migration candidates, it selects transferable large model training tasks that meet preset rules, calculates and sorts them, and constructs the current migration task group; in response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, it re-evaluates the remaining transferable large model training tasks in the initial set of migration candidates to obtain an updated current migration task group; it iteratively executes the steps of migrating the current migration task group and re-evaluating the remaining tasks until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling cycle, thus obtaining the final executable migration strategy for realizing cross-computing center migration.

[0008] Based on the same inventive concept, a second aspect of the exemplary embodiments of this disclosure provides a cross-computing center migration apparatus for large model training tasks, including a training task screening module configured to obtain multiple transferable large model training tasks from the multiple large model training tasks based on basic information of the multiple large model training tasks; a computing center screening module configured to calculate a comprehensive migration cost to the multiple available computing centers for each of the multiple transferable large model training tasks based on the state information of multiple available computing centers and elastic optical networks; and an initial migration candidate set generation module configured to match each of the multiple transferable large model training tasks from the multiple available computing centers based on preset criteria and the comprehensive migration cost. The corresponding target intelligent computing center obtains an initial migration candidate set; the migration task group construction module is configured to select transferable large model training tasks that meet preset rules based on the initial migration candidate set, calculate migration gains and sort them to construct the current migration task group; and the migration strategy construction module is configured to, in response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, re-evaluate the remaining transferable large model training tasks in the initial migration candidate set to obtain an updated current migration task group; and iteratively execute the migration of the current migration task group and the re-evaluation of the remaining tasks until there are no remaining candidate tasks that meet the migration gains and resource constraints in the current scheduling period, to obtain the final executable migration strategy for realizing cross-intelligent computing center migration.

[0009] Based on the same inventive concept, a third aspect of the exemplary embodiments of this disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in the first aspect.

[0010] Based on the same inventive concept, a fourth aspect of the exemplary embodiments of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method as described in the first aspect.

[0011] Based on the same inventive concept, a fifth aspect of the exemplary embodiments of this disclosure provides a computer program product including computer program instructions that, when run on a computer, cause the computer to perform the method as described in the first aspect.

[0012] As can be seen from the above, the cross-computing center migration technology solution for large model training provided in this disclosure can jointly consider the computing power, GPU resources, renewable energy, and elastic optical network spectrum resource status of the computing center. It determines the migration priority of candidate training tasks based on comprehensive migration costs and gains, and generates feasible migration decisions by combining candidate path search and spectrum feasibility checks. This can reduce task completion time and non-renewable energy consumption while ensuring the quality of training tasks, thereby improving system resource utilization efficiency. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the cross-computing center migration strategy method for generating large model training tasks according to an embodiment of this disclosure.

[0015] Figure 2 This is a schematic diagram of a possible optical network topology according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram of the process for calculating the comprehensive portability cost according to an embodiment of this disclosure; Figure 4 This is a schematic diagram of candidate paths in an embodiment of this disclosure; Figure 5 This is a schematic diagram illustrating various possible implementation paths for embodiments of this disclosure; Figure 6This is a schematic diagram illustrating one possible implementation path of an embodiment of this disclosure; Figure 7 This is a schematic diagram of the apparatus for migrating the large model training task across intelligent computing centers according to an embodiment of the present disclosure. Figure 8 This is a schematic diagram of the electronic device for a cross-computing center migration strategy for generating large model training tasks, as described in an embodiment of this disclosure. Detailed Implementation

[0016] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0017] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this application's technical solution, based on the prompt message.

[0018] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0019] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0020] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0021] To make the objectives, technical solutions, and advantages of this disclosure clearer, the principles and spirit of this disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0022] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0023] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar words used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly. The article "a" or "an" preceding an element does not exclude the existence of multiple such elements.

[0024] The principles and spirit of this disclosure will be explained in detail below with reference to several representative embodiments. Invention Overview As described in the background section, the migration of long-cycle, large-model training tasks across intelligent computing centers is not a simple task rescheduling. Such migrations typically require a checkpoint mechanism: training is paused and state data is saved, then the state data is transmitted to the target intelligent computing center via the network, and finally training resumes at the target center. Due to the large amount of data being migrated, cross-computing center migration requires sufficient transmission capacity from the underlying network. Elastic optical networks (OOCs) offer advantages such as high bandwidth, low latency, and flexible spectrum allocation, making them suitable for handling cross-computing center migration traffic for large-scale training tasks. However, establishing migration optical paths within an OOC must satisfy routing and resource allocation constraints. If the target intelligent computing center is selected solely from the computing or energy side without checking the paths and spectrum resources within the OOC, migration decisions that appear computationally or energy-efficient may be impractical on the network side.

[0026] In related technologies, existing scheduling methods mostly focus on the placement of ordinary cloud tasks, batch processing tasks, or static tasks, with less emphasis on dynamic migration decisions for long-term, large-scale model training tasks. While some cross-computing center migration methods consider network transmission limitations, they fail to jointly consider heterogeneous computing power, GPU resource status, task queue status, renewable energy supply, non-renewable energy consumption, and routing and resource constraints in elastic optical networks within the same framework. Therefore, there is an urgent need for a cross-computing center migration method capable of jointly coordinating computing resources, energy resources, and network spectrum resources for long-term, large-scale model training tasks.

[0027] To address the aforementioned issues, this disclosure provides a technical solution for generating a cross-computing center migration strategy for large-scale model training tasks. Specifically, it includes: obtaining multiple transferable large-scale model training tasks based on their basic information; calculating the comprehensive migration cost to multiple available intelligent computing centers for each of the multiple transferable large-scale model training tasks based on the state information of multiple available intelligent computing centers and elastic optical networks; and matching a corresponding target intelligent computing center from the multiple available intelligent computing centers for each of the multiple transferable large-scale model training tasks based on preset standards and the comprehensive migration cost. The process involves obtaining an initial migration candidate set; selecting transferable large-scale model training tasks that meet preset rules based on the initial migration candidate set, calculating and sorting the migration gains, and constructing a current migration task group; responding to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, re-evaluating the remaining transferable large-scale model training tasks in the initial migration candidate set to obtain an updated current migration task group; iteratively executing the migration of the current migration task group and the re-evaluation of remaining tasks until no remaining candidate tasks satisfying migration benefits and resource constraints exist within the current scheduling cycle, resulting in a final executable migration strategy for achieving cross-computing center migration. This technical solution can jointly consider the computing power, GPU resources, renewable energy, and elastic optical network spectrum resource status of the computing center, determine the migration priority of candidate training tasks based on comprehensive migration costs and migration gains, and generate feasible migration decisions by combining candidate path search and spectrum feasibility checks. It can reduce task completion time and non-renewable energy consumption while ensuring the service quality of training tasks, thereby improving system resource utilization efficiency.

[0028] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0029] Exemplary methods The method provided in this disclosure is applicable to training task carrying systems composed of multiple geographically distributed intelligent computing centers and elastic optical networks. The elastic optical network is a core optical network or backbone optical network connecting multiple geographically distributed intelligent computing centers. This invention is based on the characteristic that long-cycle, large-model training tasks are affected by heterogeneous computing power, GPU resource status, task queue status, time-varying renewable energy supply, and elastic optical network spectrum resources during operation. The core concept is that, when making training task migration decisions, firstly, the task completion time and non-renewable energy consumption after migrating the task to different candidate intelligent computing centers are calculated, and a comprehensive migration cost is constructed; then, a pre-selected target intelligent computing center is selected based on the comprehensive migration cost, and the task migration gain is calculated; finally, candidate tasks are sorted according to the migration gain, and a migration decision that satisfies the resource constraints of the intelligent computing center and the network spectrum constraints is generated by combining K candidate path searches and spectrum resource feasibility checks.

[0030] In some embodiments, specifically refer to Figure 1 This disclosure includes a method for generating a strategy for transferring large model training tasks across computing centers, comprising: Step 101: Based on the basic information of multiple large model training tasks, obtain multiple transferable large model training tasks from among the multiple large model training tasks. Specifically, to achieve step 101, first obtain the state information of the geographic distributed intelligent computing center and the elastic optical network, as well as the basic information of the large model training tasks.

[0031] One specific method for obtaining the state information of geographically distributed intelligent computing centers and elastic optical networks can be: through undirected graphs. Let V represent the physical topology of the elastic optical network, where V represents the set of network nodes. Represents a set of fiber optic links. (Used) This represents a set of intelligent computing centers deployed on network nodes, and satisfies... Each intelligent computing center node has computing resources, data transmission and reception capabilities, and an energy supply system; ordinary optical network nodes are only responsible for data transmission.

[0032] For any optical link Its spectrum resources are divided and managed using frequency slots as the basic unit. Assume each optical link is divided into... Each frequency slot, then the link The set of frequency slots on the frequency band is represented as:

[0033] In the link Above, the first The usage status of each frequency slot is determined by... Indicates. When When, it indicates a link The first Each frequency gap at time It is in a state of being occupied; when When this occurs, it indicates that the frequency slot is in an idle state. Simultaneously, the data for each intelligent computing center is obtained. The computational and energy states, including the single-GPU training rate factor. GPU block capacity limit GPU power requirements Current GPU resource usage status, task queue status, and earliest time when a new task can be accepted. Renewable energy supply power And the total power of the current intelligent computing center.

[0034] in, This is used to represent the differences in single-GPU training capabilities among different intelligent computing centers due to differences in GPU models, interconnect architectures, system software stacks, and operating environments.

[0035] A specific method for obtaining basic information about large model training tasks can be: representing the set of active large model training tasks in the system as... For any training task Represent it as

[0036] in, This indicates the moment when the large model training task first enters the system; Indicates the current scheduling time Assign task The remaining computational workload; This represents the total amount of data that needs to be transferred at the time of task migration. Indicates the current computing center where the task is located; This indicates the deadline for the training task, meaning the training time cannot exceed this deadline at the latest. The latency sensitivity factor represents the task. This indicates the number of GPUs required to execute the task.

[0037] specifically refer to Figure 2 , Figure 2 A possible optical network topology is shown, where blue circle nodes represent intelligent computing center nodes, which, in addition to data transmission capabilities, also possess computing resources and energy supply capabilities, while single circle nodes represent ordinary optical network nodes, responsible only for data transmission. To further illustrate the implementation method of this invention, this invention will be based on... Figure 2 The topology is shown in more specific embodiments.

[0038] Specifically, a system model can be established first. As described in the above embodiments, the elastic optical network is represented as... The set of network nodes A, D, and F are intelligent computing center nodes, i.e., the set of intelligent computing centers. B, C, and E are ordinary optical network nodes, only responsible for data transmission. The link set is as follows: All spectrum resources on each fiber optic link are divided into 8 frequency slots, numbered from slot 0 to slot 7. The link spectrum occupancy status is monitored through... The occupancy values ​​are indicated by 0, where 0 indicates the slot is idle and 1 indicates the slot is occupied. The link lengths and current spectrum occupancy status are shown in Table 1. Initially, links AF, FE, and ED on path AFED all have common, consecutive slots 1-3; links AB, BC, and CD on path ABCD all have common, consecutive slots 4-6. Therefore, this network state can be used to illustrate the process of multiple training tasks competing for limited spectrum resources after ranking. The system uses Binary Phase Shift Keying (BPSK) as the base modulation format, and the transmission rate of a single slot is... Modulation factor The current scheduling time is .

[0039]

[0040] Table 1 Link Length and Current Slot Occupancy Status In the embodiments described in this disclosure, GPU resources, training rates, task queues, and renewable energy supply information of each intelligent computing center are periodically collected. The specific status of each intelligent computing center at the current scheduling moment is shown in Table 2. Here, intelligent computing center A is the intelligent computing center where the task is currently located. Indicates that the task is included. The total power of the current training task, including the intelligent computing center D and intelligent computing center F, are candidate target intelligent computing centers. As shown in Table 2, intelligent computing center D needs to wait for... The first computing center can accept new tasks, but its single-GPU training speed is high and the supply of renewable energy is relatively abundant; the second computing center does not require queuing, but its training speed is low; the third computing center is the current location of the task, but the supply of renewable energy is limited.

[0041]

[0042] Table 2 Status Information of Intelligent Computing Center In an exemplary implementation of this embodiment, it is assumed that there are three large model training tasks currently running in the current intelligent computing center A, denoted as j1, j2, and j3. Their basic information is consistent with the task models described in the invention, specifically as shown in Table 3. Assume that all three training tasks satisfy the maximum number of transfers constraint and the minimum transfer interval constraint; therefore, all three can enter the candidate target intelligent computing center evaluation process. The system sets the maximum allowed transfer time for each transfer training task. This is used to calculate the minimum number of consecutive frequency slots required for migration.

[0043]

[0044] Table 3 Information on the three training tasks to be evaluated in Intelligent Computing Center A The above embodiments are used to obtain the information required to implement step 101. However, those skilled in the art should understand that the above embodiments are merely exemplary and not limiting. The training tasks and data information required to implement the technical solutions in this disclosure, as well as the specific methods for obtaining the information, are not limited to those described in the above embodiments.

[0045] Specifically, the basic information of the aforementioned large model training tasks includes the transfer constraints of the multiple large model training tasks. Step 101 may include: Choose one large model training task from multiple large model training tasks.

[0046] Obtain the maximum number of transfers and minimum transfer interval constraints for the selected large model training task.

[0047] In response to the maximum number of transfers and minimum transfer interval constraints of the selected large model training task satisfying the preset conditions, the selected large model training task is added to the set for multiple transferable large model training tasks.

[0048] And it iterates through multiple large model training tasks to obtain multiple transferable large model training tasks.

[0049] Because this invention employs a periodic migration mechanism, the method disclosed herein performs migration evaluation with a fixed scheduling period. To avoid excessive data transmission, state updates, and resource reconfiguration overhead caused by frequent migrations of training tasks, it is necessary to first check whether the task meets migration constraints. These migration constraints include a maximum migration count constraint and a minimum migration interval constraint. Let... Indicates the maximum number of migrations allowed for the task. Indicates the time of the task Whether migration occurs or not, the number of migrations for the task during the entire operation should not exceed [number missing]. At the same time, in length of Within the specified time window, a task may undergo migration at most once. If a task does not meet the above migration constraints, it will not be included in the migration candidate evaluation in the current scheduling cycle; if it meets the above migration constraints, the candidate target intelligent computing center evaluation will continue.

[0050] Step 102: Based on the state information of multiple available intelligent computing centers and elastic optical networks, calculate the comprehensive migration cost to multiple available intelligent computing centers for each task in multiple transferable large model training tasks.

[0051] specifically refer to Figure 3 Based on the state information of multiple available intelligent computing centers and elastic optical networks, the comprehensive migration cost for each task in multiple transferable large model training tasks to multiple available intelligent computing centers is calculated, including: Step 301: Select a transferable large model training task from multiple transferable large model training tasks.

[0052] Step 302: Calculate the task completion time and non-renewable energy consumption for migrating the selected transferable large model training task to all available intelligent computing centers. Specifically, this step may include the following: Calculate the number of consecutive bandwidths and network transmission time required for the transferable training task. For training tasks that satisfy the transfer constraints... Traversing candidate intelligent computing centers and The system sets a maximum allowed transfer time for status data for each migration task. This is used to determine the minimum number of consecutive frequency slots required for the transfer. It depends on the amount of transfer data required for the training task. Modulation factor corresponding to modulation format, single-slot reference transmission rate and migration transmission time constraints Calculate the number of consecutive frequency slots required for migration. .in, Indicates the task Migrating data within the maximum allowed transfer time The minimum number of consecutive frequency slots required to complete transmission within a given timeframe. In determining... Afterwards, when the task originates from the current intelligent computing center Migration to candidate intelligent computing centers At that time, its state data Transmission via a flexible optical network, actual network transmission time It can be calculated using formula (1): (1) Based on Figure 2In the illustrated embodiment, this disclosure further elaborates on the implementation method of this step. Specifically, based on the above, the amount of data transmitted for training task j1 is... for The number of consecutive frequency slots required for training task j1 is Therefore, when training task j1 is transferred in the elastic optical network, at least 3 consecutive frequency slots need to be allocated.

[0053] Calculate the queuing time for transferable training tasks after migrating to all available intelligent computing centers. When the training task data arrives at a candidate intelligent computing center, if the GPU resources of the target intelligent computing center have not yet been released, the task needs to enter the queue and wait. The candidate intelligent computing center at time... The earliest time when a new task can be accepted is recorded as The task will then be migrated to the candidate intelligent computing center. After the queue waiting time The calculation can be performed according to formula (2). Indicates the current scheduling time Next Candidate Intelligent Computing Center The earliest moment when one can accept new tasks.

[0054] (2) Calculate the execution time of the transferable training task after migrating it to all available intelligent computing centers. This is assuming the training task is in a candidate intelligent computing center. After acquiring sufficient GPU resources, the training phase resumes. Considering the differences in single-GPU training capabilities among different computing centers, the task is performed at candidate computing centers. Execution time It can be calculated according to formula (3).

[0055] (3) Calculate the task completion time after the transferable training task is migrated to all available intelligent computing centers. Remaining completion time after migration to candidate intelligent computing center It consists of three parts: network transmission time, queuing time, and execution time after resuming training. It can be calculated according to formula (4). Specifically, it starts from the moment the task initially enters the system. Start calculating, task At the current scheduling time Migration to candidate intelligent computing centers The task completion time after that The calculation can be performed according to formula (5). Simultaneously, the migration must be completed before the deadline. .

[0056] (4) (5) Based on Figure 2 In the illustrated embodiment, this disclosure further describes a specific implementation method for this step. Specifically, if training task j1 is migrated from intelligent computing center A to intelligent computing center D, the network transmission time is calculated according to formula (1). Because the earliest time that the intelligent computing center D can accept new tasks. Therefore, according to formula (2), training task j1 still needs to wait after arriving at the intelligent computing center D. According to formula (3), the execution time of training task j1 after resuming training in the intelligent computing center D is: According to formula (4), we can obtain According to formula (5), the total completion time of the task can be obtained. Therefore, the intelligent computing center D satisfies the deadline constraint of training task j1.

[0057] based on Figure 2 The illustrated embodiment also includes calculating the number of consecutive frequency slots required for training task j2 and the task completion time after migration to the intelligent computing center D. The amount of data transmitted for training task j2 is... for The number of consecutive frequency slots required for training task j2 is .

[0058] If training task j2 is moved from intelligent computing center A to intelligent computing center D, then the network transmission time is calculated according to formula (1) as follows: According to formula (2), the waiting time for training task j2 after arriving at the intelligent computing center D is: According to formula (3), the execution time of training task j2 after training resumes at the intelligent computing center D is: According to formulas (4) and (5), we can obtain ,at the same time Therefore, the intelligent computing center D satisfies the deadline constraint of training task j2.

[0059] Calculate the non-renewable energy consumption during the execution phase of the transferable training task after it has been migrated to all available intelligent computing centers. For any candidate intelligent computing center, Indicates at time At the Intelligent Computing Center The set of training tasks running in the process. Indicates intelligent computing center Total power demand Calculate according to formula (6), Indicates task The training power requirements. In cases where multiple training tasks share renewable energy supply, renewable energy power is allocated to tasks according to their power requirements as a proportion of the total power requirements of the training tasks after migration. Allocated to tasks renewable energy power Calculate according to formula (7). Task In candidate intelligent computing centers Non-renewable energy consumption during the implementation phase Calculate according to formula (8).

[0060] (6) (7) (8) And calculate the non-renewable energy consumption during the migration phase of the transferable training task. Migration across computing centers itself generates network transmission energy consumption. Let... The average network migration power during the migration phase is represented by the task. Migration to candidate intelligent computing centers Non-renewable energy consumption during the migration phase Calculated according to formula (9). Total non-renewable energy consumption after the task is migrated to the candidate intelligent computing center. Calculate according to formula (10).

[0061] (9) (10) Step 303: Calculate the comprehensive migration cost of migrating the selected transferable large-scale model training task to all available intelligent computing centers, based on task completion time and non-renewable energy consumption. The comprehensive migration cost is constructed based on task completion time and non-renewable energy consumption. Calculated according to formula (11), As a weighting factor, and ; and These represent the normalized baseline values ​​for task completion time and non-renewable energy consumption, respectively.

[0062] (11) And in step 304, traverse multiple transferable large model training tasks to obtain the comprehensive migration cost of migrating multiple transferable large model training tasks to multiple available intelligent computing centers.

[0063] Based on Figure 2 In the illustrated embodiment, this disclosure further elaborates on the implementation method of the above steps. The training power requirement for training task j1 is... For candidate target intelligent computing center D, This indicates the power of the training task before the migration of training task j1. Therefore, the total power of the training task after migrating training task j1 to the intelligent computing center D is: In the intelligent computing center D, the renewable energy supply power... Based on the proportion of task power demand to the total power of the training task after migration, according to formula (7), the renewable energy power allocated to training task j1 is: Therefore, the power required by non-renewable energy sources for training task j1 during its execution phase at the intelligent computing center D is [value missing]. According to formula (8), the non-renewable energy consumption during the execution phase is... Let the average network migration power during the migration phase be... According to formula (9), the non-renewable energy consumption during the migration phase is... According to formula (10), the total non-renewable energy consumption after training task j1 is migrated to the intelligent computing center D is Let the weighting factor be... Normalized baseline value of task completion time Normalized baseline value of non-renewable energy consumption According to formula (11), the comprehensive migration cost after training task j1 is transferred to the intelligent computing center D is: .

[0064] based on Figure 2 The illustrated embodiment further includes determining the pre-selected target computing center for training task j1 and the migration gain by calculating the combined costs of keeping training task j1 executed at computing center A and migrating it to computing center F, respectively. If training task j1 is kept executed at computing center A, then , , If training task j1 is migrated to the intelligent computing center F, then , , .because The minimum value is selected, therefore, the intelligent computing center D is chosen as the pre-selected target intelligent computing center for training task j1, i.e. According to formula (12), the transfer gain of training task j1 is: ,because Furthermore, since the pre-selected target intelligent computing center D is different from the current intelligent computing center A, training task j1 is added to the transfer candidate set. .

[0065] based on Figure 2 The illustrated embodiment also includes calculating the non-renewable energy consumption, overall migration cost, and migration gain of training task j2. The training power requirement of training task j2 is... According to formula (6), the total power of the training task j2 after being migrated to the intelligent computing center D is: According to formula (7), the renewable energy power allocated to training task j2 is Therefore, the power required by non-renewable energy sources for training task j2 during its execution phase at the intelligent computing center D is [amount missing]. According to formulas (8) to (10), we can obtain , According to formula (11), the comprehensive migration cost after training task j2 is transferred to the intelligent computing center D is: Using the same method, it can be seen that training task j2 is maintained while being executed at the intelligent computing center A. , , When training task j2 is migrated to the intelligent computing center F, , , .because The minimum target computing center for training task j2 is F. According to formula (12) in the invention, its transfer gain is... ,because Furthermore, since the pre-selected target intelligent computing center F is different from the current intelligent computing center A, training task j2 is added to the transfer candidate set. .

[0066] Step 103: Based on preset standards and comprehensive transfer costs, match the corresponding target intelligent computing center from multiple available intelligent computing centers for each of the multiple transferable large model training tasks to obtain an initial transfer candidate set.

[0067] In one specific embodiment, step 103 may include the following steps: Based on preset standards and the comprehensive transfer cost of all available intelligent computing centers, a target intelligent computing center is matched from the candidate intelligent computing centers for each transferable large model training task. Specifically, for the training task... After traversing all candidate intelligent computing centers, the intelligent computing center with the lowest overall migration cost is selected as the pre-selected target intelligent computing center. .

[0068] Based on all available intelligent computing centers and their respective matched target intelligent computing centers, a migration gain is calculated. A migration candidate set is then constructed based on the migration gain. The overall cost of keeping the task running in the current intelligent computing center is calculated. and migrate it to the pre-selected target intelligent computing center. The overall migration cost afterward Compare the task's migration gain. Calculate according to formula (12): (12) in, This represents the latency sensitivity factor for the task, used to adjust the impact of transfer learning gains on training stability. For tasks with low latency sensitivity, Taking a larger value reduces the transfer gain score, avoiding overly frequent or aggressive transfers; for tasks with high latency sensitivity, Choose a smaller value to make it easier to schedule and migrate for energy savings or speedup benefits. and Then the task will be added to the set of all migration candidate tasks in the current network. Otherwise, the task will remain executed at the original intelligent computing center during the current scheduling cycle.

[0069] Candidate tasks are sorted according to transfer gain to obtain a candidate task sequence. After calculating the transfer gain for all active training tasks, the candidates are sorted according to transfer gain. From largest to smallest, the migration candidate set The training tasks with higher potential benefits are prioritized to receive resources from the target intelligent computing center and network spectrum.

[0070] based on Figure 2 The illustrated embodiment also includes calculating the transfer evaluation result of training task j3. The calculation process for training task j3 is similar to that for training tasks j1 and j2, and will not be elaborated here. According to formulas (1) to (12), the number of consecutive frequency slots required for training task j3 is... The evaluation results for execution at computing center A, migration to computing center D, and migration to computing center F are shown in Table 4. Migration to computing center F has the lowest overall cost, therefore, F is the pre-selected target computing center for training task j3. Due to the latency sensitivity factor of training task j3... Therefore, according to formula (12), its migration gain is ,because Furthermore, since the pre-selected target intelligent computing center F is different from the current intelligent computing center A, training task j3 is also added to the transfer candidate set. .

[0071]

[0072] Table 4. Transfer evaluation results for training task j3 Candidate paths are generated based on the candidate task sequence. The migration candidate set is then processed sequentially according to the sorting results. The training task in the [system / process]. For any candidate transfer task, [it is performed] within its current intelligent computing center. With the pre-selected target intelligent computing center Between, adopt The shortest path algorithm generates at most 10 candidate paths.

[0073] And based on the candidate paths, an initial migration candidate set is generated. In this step, the candidate path set is first... The paths in the spectrum are checked for spectrum resource feasibility in sequence. For a given candidate path... If there exists a continuous frequency gap, for example, the first... The frequency gap to the first One frequency gap, and path All links on If all conditions are met, then the routing and resource allocation scheme is satisfied. In this step, the spectrum resource feasibility check includes three types of constraints: the first type is the spectrum consistency constraint, that is, the same migration request occupies the same numbered range of frequency slots on all links of the end-to-end path; the second type is the spectrum continuity constraint, that is, the same migration request occupies consecutive frequency slots on each link; and the third type is the spectrum non-overlap constraint, that is, different migration requests sharing the same link cannot occupy the same frequency slot. In one implementation, the First-Fit algorithm can be used for spectrum search, that is, searching for available consecutive frequency slots that satisfy the above constraints sequentially from low-numbered frequency slots to high-numbered frequency slots.

[0074] (13) Then, during the generation of the initial migration candidate set, if a path and slot combination that satisfies the constraints is found on a certain candidate path, the task is... Target Intelligent Computing Center Candidate paths The allocated continuous frequency band range constitutes the migration decision and is added to the initial migration candidate set. After processing the sorted migration candidate set Then, the initial migration candidate set is set. Construction complete. Collection Each element in the dataset includes at least the training task to be transferred, the source computing center, the target computing center, the transfer path, the allocated frequency slot, the start time of the transfer, and the resource status update information after the transfer is completed.

[0075] Based on Figure 2 In the illustrated embodiment, this disclosure further elaborates on the implementation method of this step. The candidate transfer evaluation results for the three training tasks, after the above calculations, are summarized in Table 5. Since the transfer gain of all three training tasks is greater than 0, and their pre-selected target computing centers are not equal to their current computing center A, all three are added to the transfer candidate set. Then, the migration candidate set is sorted from largest to smallest according to migration gain, resulting in... In scenarios involving concurrent migration of multiple training tasks, the available spectrum resources in the network are limited. Without prioritizing migration gains, low-gain training tasks may occupy consecutive frequency slots on critical links, thus blocking the migration of high-gain training tasks. This invention calculates the combined migration cost and migration gain first, and then ranks candidate training tasks according to their migration gains, ensuring that more worthwhile training tasks receive priority access to network spectrum resources.

[0076]

[0077] Table 5. Transfer cost and transfer gain for the three training tasks Step 104: Based on the initial transfer candidate set, select transferable large model training tasks that meet preset rules, calculate and sort the transfer gains, and construct the current transfer task group. Specifically, to avoid changes in the resource status of the intelligent computing center and network resources after the transfer of previous tasks affecting the decision-making of subsequent tasks, the initial transfer candidate set is... A grouped iterative execution mechanism is adopted.

[0078] In step 105, in response to the current migration task group completing the generation of executable migration decisions and occupying the corresponding resources, the remaining transferable large model training tasks in the initial migration candidate set are re-evaluated to obtain the updated current migration task group; the re-evaluation steps of the current migration task group and the remaining tasks are executed cyclically until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling period, and the final executable migration strategy for realizing cross-computing center migration is obtained.

[0079] In one specific embodiment, step 105 may include the following steps: In response to the current migration task group completing the generation of executable migration decisions and occupying the corresponding resources, update the resource status, task queue status, earliest time when new tasks can be accepted, energy status, and elastic optical network spectrum occupancy status of the source intelligent computing center and the target intelligent computing center to obtain the updated target intelligent computing center status. For large model training tasks that were not executed in the initial set of transferable candidates, the comprehensive transfer cost and transfer gain are recalculated based on the updated state of the target intelligent computing center. Based on preset update rules, update the transferable large model training tasks in the initial transfer candidate set; and Based on the updated migration strategy group and the corresponding comprehensive migration cost and migration gain, the current migration task group and the remaining task re-evaluation steps are executed cyclically until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling cycle, and the final executable migration strategy for realizing cross-computing center migration is obtained. The preset update rules include: if a transferable large model training task does not have transfer benefits or does not meet the current resource constraints, then it will not be transferred in the current transfer loop; if a transferable large model training task has transfer benefits and meets the resource constraints, then it will be retained in the transfer tasks to be executed and will enter the next set of transfers.

[0080] Specifically, for each group in the initial migration candidate set, based on the migration gain ranking, priority is given to executing migration tasks that are distinct from the pre-selected target intelligent computing centers, and the resource status, task queue status, earliest available time for accepting new tasks, and network spectrum occupancy status of the intelligent computing centers are updated. Subsequently, for For migration tasks that have not yet been executed, the overall migration cost and migration gain are recalculated based on the updated system state. If a task no longer has migration benefits or does not meet the current resource constraints, it will not be migrated in the current scheduling cycle; if it still has migration benefits and meets the resource constraints, it will be retained in the pending migration tasks and enter the next set of migration judgments.

[0081] After completing the initial migration candidate set After grouped iterative execution and dynamic reassessment, the system outputs the final set of migration tasks actually executed and the set of tasks whose migration is postponed within the current scheduling period. For tasks whose migration is postponed within the current scheduling period, a new migration assessment will be conducted after the next scheduling period arrives, taking into account the updated status of the intelligent computing center, the renewable energy supply status, and the elastic optical network spectrum resource status.

[0082] Based on Figure 2 In the illustrated embodiment, this disclosure further elaborates on the implementation method of this step. Let K=3. Based on the sorting results obtained in the above steps, since the pre-selected target computing centers are different for different training tasks, it is necessary to generate candidate paths separately for each training task's source computing center and pre-selected target computing center, rather than using the same set of candidate paths for all training tasks. For training task j1, its source computing center is A, and its pre-selected target computing center is D. At most 3 candidate paths are generated, denoted as... .like Figure 4 As shown, path p1 is from node A to node F to node E to node D, with a path length of 500 + 600 + 500 = 1600 km; path p2 is from node A to node B to node C to node D, with a path length of 500 + 600 + 500 = 1600 km; path p3 is from node A to node B to node E to node D, with a path length of 500 + 700 + 500 = 1700 km. For training tasks j2 and j3, the source computing center is A, and the pre-selected target computing center is F. Candidate paths from A to F are generated respectively, denoted as... Among them, path q1 is from node A to node F, with a path length of 500km; path q2 is from node A to node B to node F, with a path length of 500+400=900km; path q3 is from node A to node B to node E to node F, with a path length of 500+700+600=1800km. Subsequently, the feasibility of paths and spectrum resources is checked sequentially according to the sorting result [j1,j2,j3].

[0083] Then, based on the sorting results obtained from the above steps... The pre-selected target intelligent computing centers for training tasks j1, j2, and j3 are D, F, and F, respectively. Candidate tasks are traversed from largest to smallest based on migration gain, with higher-priority tasks having different pre-selected target intelligent computing centers being prioritized for inclusion in the current batch migration decision set. Therefore, training tasks j1 and j2 are prioritized for routing and spectrum resource allocation decisions in the current batch; training tasks j3 and j2 both have pre-selected target intelligent computing centers of F, so they are not included in the current batch migration decision and will be reserved for re-evaluation based on the updated system status.

[0084] Based on the above results, Figure 2 In the illustrated embodiment, training task j1 is processed first, and the network spectrum state is updated. The pre-selected target computing center for training task j1 is D, and it requires 3 consecutive frequency slots. First, the candidate path p1 is checked as node A to node F to node E to node D. According to the initial spectrum state in Table 1, links AF, FE, and ED all have the first to third consecutive free frequency slots. According to formula (13), the links on path p1 need to satisfy , , , e∈{AF,FE,ED}. Therefore, path p1 satisfies the routing and resource allocation scheme of training task j1. For example... Figure 5 As shown, at this point, the migration decision generated for training task j1 is: source computing center A, target computing center D, migration path from node A to node F to node E to node D, and allocation of frequency slots 1-3. Subsequently, the states of frequency slots 1-3 on links AF, FE, and ED are updated from 0 to 1, indicating that these frequency slots have been occupied by training task j1. After the update, the spectrum states of the relevant links are... , , After the update, some of the public contiguous frequency slot resources on path AFED were consumed, and subsequent training tasks need to continue to perform feasibility checks based on the updated network topology.

[0085] Then, training task j2 is processed and the network spectrum state is updated. The pre-selected target computing center for training task j2 is F, and it requires two consecutive frequency slots. Therefore, the candidate path from A to D is no longer checked; instead, the candidate path from A to F corresponding to training task j2 is checked. First, candidate path q1 is checked as node A to node F. Since training task j1 has already occupied the first to third frequency slots on link AF, the updated spectrum state of link AF is... Therefore, it can be seen that the 6th and 7th frequency slots of link AF are still idle, and the 6th and 7th frequency slots form a continuous frequency slot block of length 2, which satisfies the 2 continuous frequency slots required for training task j2. According to formula (13), for link AF on path q1, it satisfies , , Therefore, as Figure 6 As shown, path q1=AF satisfies the routing and resource allocation scheme for training task j2. The migration decision generated for training task j2 is: source computing center A, target computing center F, migration path from node A to node F, and allocation of frequency slots 6-7. Subsequently, the state of frequency slots 6-7 on link AF is updated from 0 to 1, indicating that these frequency slots have been occupied by training task j2. After the update, the spectrum state of link AF is... .

[0086] Subsequently, the remaining training task j3 is dynamically re-evaluated. After training tasks j1 and j2 complete their migration and update their system status, training task j3, which has not yet entered the current batch migration decision, is re-evaluated. If training task j3 still has migration benefits in the updated system status, its candidate path and spectrum resource feasibility are further checked. The pre-selected target computing center for training task j3 is F, and it requires 3 consecutive frequency slots. First, path q1=AF is checked. Since all frequency slots on link AF are occupied, path q1 cannot meet its migration requirements. Path q2=ABF is checked next. There are no three consecutive free frequency slots in common between links AB and BF, so q2 is also not feasible. Path q3=ABEF is checked next. Similarly, there are no consecutive free frequency slots in common between links AB, BE, and EF. Therefore, training task j3 cannot obtain a routing and resource allocation scheme that meets the constraints in the current scheduling cycle, and migration is not performed temporarily. It continues to run in computing center A and waits for re-evaluation in the next scheduling cycle.

[0087] After the current batch migration execution and dynamic reassessment of remaining tasks, the migration tasks actually executed in this scheduling cycle are training task j1 and training task j2. Training task j1 migrates from intelligent computing center A to intelligent computing center D via the migration path AFED, allocating slots 1-3; training task j2 migrates from intelligent computing center A to intelligent computing center F via the migration path AF, allocating slots 6-7. Training task j3 cannot obtain a routing and resource allocation scheme that satisfies the constraints under the updated network spectrum status, and will not be migrated in the current scheduling cycle. Upon the arrival of the next scheduling cycle, based on the latest intelligent computing center resource status, renewable energy supply status, and elastic optical network spectrum resource status, a reassessment will be made regarding whether training task j3 meets the migration conditions. If training task j3 meets the migration conditions, it will be migrated; otherwise, it will not be migrated in this cycle and will wait for the next scheduling time to migrate.

[0088] The cross-computing center migration method for large model training provided by this invention can jointly consider the computing power, GPU resources, renewable energy, and spectrum resource status of the computing center. Based on a comprehensive migration cost and migration gain, it determines the migration priority of candidate training tasks and generates feasible migration decisions by combining candidate path search and spectrum feasibility checks. This method can reduce task completion time and non-renewable energy consumption while ensuring the quality of training tasks, thereby improving system resource utilization efficiency.

[0089] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0090] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0091] Exemplary device Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a cross-computing center migration device for large model training tasks.

[0092] refer to Figure 7 The cross-computing center migration device for the large model training task includes: The training task selection module is configured to obtain multiple transferable large model training tasks based on the basic information of multiple large model training tasks. The intelligent computing center selection module is configured to calculate the comprehensive migration cost for each of the multiple transferable large model training tasks to multiple available intelligent computing centers based on the state information of multiple available intelligent computing centers and the elastic optical network. The initial migration candidate set generation module is configured to match the corresponding target intelligent computing center from multiple available intelligent computing centers for each of the multiple transferable large model training tasks based on preset criteria and the comprehensive migration cost to obtain an initial migration candidate set. The task group construction module is configured to select transferable large model training tasks that meet preset rules based on the initial migration candidate set, calculate and sort the transfer gains, and construct the current migration task group; and the migration strategy construction module is configured to, in response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, re-evaluate the remaining transferable large model training tasks in the initial migration candidate set to obtain the updated current migration task group; and cyclically execute the steps of migrating the current migration task group and re-evaluating the remaining tasks until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling period, to obtain the final executable migration strategy for realizing cross-computing center migration.

[0093] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0094] The apparatus described above is used to implement the cross-computing center migration method for the corresponding large model training task in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0095] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the cross-computing center migration method for large model training tasks described in any of the above embodiments.

[0096] Figure 8This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0097] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0098] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0099] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0100] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0101] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0102] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0103] The electronic devices described above are used to implement the cross-computing center migration method for the corresponding large model training task in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0104] The memory 1020 stores machine-readable instructions executable by the processor 1010. When the electronic device is running, the processor 1010 communicates with the memory 1020 via the bus 1030, causing the processor 1010 to execute the following instructions during operation: Based on the basic information of multiple large-scale model training tasks, multiple transferable large-scale model training tasks are obtained. Based on the state information of multiple available intelligent computing centers and elastic optical networks, the comprehensive migration cost for each of the multiple transferable large-scale model training tasks to migrate to multiple available intelligent computing centers is calculated. Based on preset standards and comprehensive migration costs, each of the multiple transferable large-scale model training tasks is matched with a corresponding target intelligent computing center from multiple available intelligent computing centers to obtain an initial migration candidate set. According to the initial migration candidate set, transferable large-scale model training tasks that meet preset rules are selected, migration gains are calculated and sorted to construct the current migration task group. In response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, the remaining transferable large-scale model training tasks in the initial migration candidate set are re-evaluated to obtain an updated current migration task group. The steps of migrating the current migration task group and re-evaluating the remaining tasks are executed iteratively until there are no remaining candidate tasks that meet the migration gains and resource constraints in the current scheduling cycle, resulting in the final executable migration strategy for realizing cross-intelligent computing center migration.

[0105] By employing the cross-computing center migration method for large model training provided in this disclosure, the method can jointly consider the computing power, GPU resources, renewable energy sources, and the spectrum resource status of elastic optical networks within the computing center. Based on a comprehensive migration cost and gain, it determines the migration priority of candidate training tasks and generates feasible migration decisions by combining candidate path search and spectrum feasibility checks. This approach can reduce task completion time and non-renewable energy consumption while ensuring the quality of training tasks, thereby improving system resource utilization efficiency.

[0106] Exemplary program product Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the cross-computing center migration method for large model training tasks as described in any of the above embodiments.

[0107] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0108] The aforementioned non-transitory computer-readable storage media can be any available medium or data storage device that a computer can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0109] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the cross-computing center migration method for large model training tasks as described in any of the embodiments in the exemplary method section above, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0110] Based on the same inventive concept, corresponding to the cross-computing center migration method for large model training tasks described in any of the above embodiments, this disclosure also provides a computer program product, which includes computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer to cause the computer and / or the processors to perform the cross-computing center migration method for large model training tasks. Corresponding to the execution entity for each step in each embodiment of the cross-computing center migration method for large model training tasks, the processor executing the corresponding step can belong to the corresponding execution entity.

[0111] The computer program product of the above embodiments is used to enable the computer and / or the processor to perform the cross-computing center migration method for large model training tasks as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0112] Those skilled in the art will recognize that embodiments of this disclosure can be implemented as a system, method, or computer program product. Therefore, this disclosure can be implemented as entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this disclosure can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0113] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (not exhaustive) of a computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0114] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0115] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0116] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0117] It should be understood that each block of a flowchart and / or block diagram, as well as combinations of blocks in a flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine that, when executed by a computer or other programmable data processing device, creates means for implementing the functions / operations specified in the blocks of the flowchart and / or block diagram.

[0118] These computer program instructions may also be stored in a computer-readable medium that enables a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce a product comprising an instruction apparatus that implements the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0119] Computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, such that the instructions that execute on the computer or other programmable apparatus can provide a process for implementing the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0120] Furthermore, although the operations of the methods of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be executed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0121] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0122] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0123] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0124] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0125] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0126] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

[0127] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. The scope of the appended claims is to be interpreted in the broadest sense, thereby encompassing all such modifications and equivalent structures and functions.

Claims

1. A method for generating a strategy to transfer training tasks for large models across intelligent computing centers, comprising: Based on the basic information of multiple large model training tasks, multiple transferable large model training tasks are obtained from the multiple large model training tasks. Based on the state information of multiple available intelligent computing centers and elastic optical networks, the comprehensive migration cost to the multiple available intelligent computing centers is calculated for each task in the multiple transferable large model training tasks. Based on the preset criteria and the comprehensive transfer cost, each of the multiple transferable large model training tasks is matched with a corresponding target intelligent computing center from the multiple available intelligent computing centers to obtain an initial set of transfer candidates. Based on the initial transfer candidate set, select the transferable large model training tasks that meet the preset rules, calculate and sort the transfer gains, and construct the current transfer task group; as well as In response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, the remaining transferable large model training tasks in the initial migration candidate set are re-evaluated to obtain an updated current migration task group; the re-evaluation steps of the current migration task group and the remaining tasks are executed cyclically until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling period, and the final executable migration strategy for realizing the cross-computing center migration is obtained.

2. The method according to claim 1, wherein, The basic information of the multiple large model training tasks includes the transfer constraints of the multiple large model training tasks. Based on the basic information of multiple large model training tasks, the following multiple transferable large model training tasks are obtained from the multiple large model training tasks: Choose one large model training task from the multiple large model training tasks; Obtain the maximum number of transfers and minimum transfer interval constraints for the selected large model training task; In response to the maximum number of transfers constraint and the minimum transfer interval constraint of the selected large model training task satisfying a preset condition, the selected large model training task is added to the set of the plurality of transferable large model training tasks; and The multiple large model training tasks are traversed to obtain the multiple transferable large model training tasks.

3. The method according to claim 1, wherein, Based on the state information of multiple available intelligent computing centers and elastic optical networks, the comprehensive migration cost for each task in the multiple transferable large model training tasks to be migrated to the multiple available intelligent computing centers is calculated, including: Select one transferable large model training task from the plurality of transferable large model training tasks; Calculate the task completion time and non-renewable energy consumption for migrating the selected transferable large model training task to all of the multiple available intelligent computing centers; Based on the task completion time and the non-renewable energy consumption, calculate the comprehensive migration cost of migrating the selected transferable large model training task to all of the multiple available intelligent computing centers; and The multiple transferable large model training tasks are traversed to obtain the comprehensive migration cost of migrating the multiple transferable large model training tasks to the multiple available intelligent computing centers.

4. The method according to claim 1, wherein, Based on preset standards and the comprehensive transfer cost, for each of the multiple transferable large model training tasks, a corresponding target intelligent computing center is matched from the multiple available intelligent computing centers to obtain an initial transfer candidate set, including: Based on preset standards and the comprehensive migration cost of all available intelligent computing centers, a target intelligent computing center is matched from the candidate intelligent computing centers for each of the transferable large model training tasks; The migration gain is calculated based on all available intelligent computing centers and their respective matched target intelligent computing centers; Construct a migration candidate set based on the migration gain; The candidate tasks are sorted according to the migration gain to obtain a candidate task sequence; Based on the aforementioned candidate task sequence, candidate paths are generated; and Based on the candidate paths, an initial migration candidate set is generated.

5. The method according to claim 1, wherein, In response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, the remaining transferable large model training tasks in the initial migration candidate set are re-evaluated to obtain an updated current migration task group; the re-evaluation steps of migrating the current migration task group and remaining tasks are executed cyclically until there are no remaining candidate tasks that satisfy the migration benefits and resource constraints within the current scheduling period, resulting in the final executable migration strategy for implementing the cross-computing center migration, including: In response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, the resource status, task queue status, earliest time to accept new tasks, energy status, and elastic optical network spectrum occupancy status of the source intelligent computing center and the target intelligent computing center are updated to obtain the updated target intelligent computing center status. For the transferable large model training tasks that were not executed in the initial set of transfer candidates, the comprehensive transfer cost and transfer gain are recalculated based on the updated target intelligent computing center state. Based on preset update rules, update the transferable large model training tasks in the initial transfer candidate set; and Based on the updated migration strategy group and the corresponding comprehensive migration cost and migration gain, the migration current task group and remaining task re-evaluation steps are executed cyclically until there are no remaining candidate tasks that meet the migration benefits and resource constraints in the current scheduling period, so as to obtain the final executable migration strategy for realizing the cross-computing center migration. The preset update rules include: if the transferable large model training task does not have transfer benefits or does not meet the current resource constraints, then it will not be transferred in the current transfer loop; if the transferable large model training task has transfer benefits and meets the resource constraints, then it will be retained in the transfer tasks to be executed and enter the next set of transfers.

6. The method according to claim 3, wherein, The calculation of the task completion time and non-renewable energy consumption for migrating the transferable large model training task to all available intelligent computing centers includes: Calculate the number of consecutive frequency slots and network transmission time required for the transferable training task; Calculate the queuing time after the transferable training task is migrated to all available intelligent computing centers; Calculate the execution time of the transferable training task after it has been migrated to all available intelligent computing centers; Calculate the task completion time after the transferable training task is migrated to all available intelligent computing centers; Calculate the non-renewable energy consumption during the execution phase of the transferable training task after it has been migrated to all available intelligent computing centers; and Calculate the non-renewable energy consumption during the migration phase of the transferable training task.

7. A cross-computing center migration device for large model training tasks, comprising: The training task selection module is configured to obtain multiple transferable large model training tasks from the multiple large model training tasks based on the basic information of multiple large model training tasks. The intelligent computing center screening module is configured to calculate the comprehensive migration cost to the multiple available intelligent computing centers for each of the multiple transferable large model training tasks based on the state information of multiple available intelligent computing centers and elastic optical networks. The initial migration candidate set generation module is configured to match the corresponding target intelligent computing center from the multiple available intelligent computing centers for each of the multiple transferable large model training tasks in the multiple transferable large model training tasks based on preset criteria and the comprehensive migration cost to obtain the initial migration candidate set. The transfer task group construction module is configured to select the transferable large model training tasks that meet the preset rules according to the initial transfer candidate set, calculate the transfer gain and sort them, and construct the current transfer task group. as well as The migration strategy construction module is configured to, in response to the current migration task group completing the generation of executable migration decisions and occupying corresponding resources, re-evaluate the remaining transferable large model training tasks in the initial migration candidate set to obtain an updated current migration task group; and cyclically execute the migration of the current migration task group and the re-evaluation of the remaining tasks until there are no remaining candidate tasks that satisfy the migration benefits and resource constraints in the current scheduling period, thereby obtaining the final executable migration strategy for realizing the cross-computing center migration.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as claimed in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 6.

10. A computer program product comprising computer program instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images