JP2022539955A - Task scheduling method and apparatus - Google Patents

Abstract

Embodiments of the present disclosure provide a task scheduling method and apparatus. The method includes obtaining task information, the task information including at least one of a task type and resource requirement information of the task, and determining target nodes for each of a plurality of subtasks included in the task based on the task information. assign. [Selection drawing] Fig. 2

Description

The present disclosure relates to the technical field of distributed systems, and more particularly to task scheduling in distributed systems.

<Cross reference to related applications>
This application is a Chinese patent application with application number 202010165543.4 and entitled "Task Scheduling Method and Apparatus" filed on March 11, 2020 and filed on March 11, 2020 , No. 202010165763.7 and entitled "Task Scheduling Method and Apparatus", claims priority from a Chinese patent application, which is incorporated herein by reference in its entirety.

Today, more and more tasks are starting to utilize distributed systems for resource allocation and computation. The content of multiple subtasks may be the same or different, and the resources required by multiple subtasks may be the same or different. Task scheduling in conventional distributed systems generally allocates resource nodes on a per-subtask basis, resulting in low cluster resource utilization and low task execution efficiency.

The present disclosure provides a task scheduling method.

In a first aspect of an embodiment of the present disclosure, obtaining task information, wherein the task information includes at least one of a task type of the task and resource requirement information of the task; assigning respective target nodes to a plurality of sub-tasks included in said task based on .

In some embodiments, the resource requirement information of the task is resource requirement information of a plurality of subtasks included in the task, and the method comprises determining that currently available resources of a distributed system are resources of a plurality of subtasks of the task. The step of allocating each of the plurality of subtasks included in the task with a corresponding target node if the total resource corresponding to the request information is satisfied.

In some embodiments, the method further comprises determining target nodes for a plurality of subtasks of the task based on resource requirement information of the plurality of subtasks, wherein currently available resources of the distributed system are: assigning a respective target node to a plurality of subtasks included in the task when meeting total resources corresponding to resource requirement information of the plurality of subtasks of the task, corresponding to each subtask of the plurality of subtasks. assigning a respective target node to each of a plurality of subtasks included in the task if the target node is successfully determined;

In some embodiments, based on the current resource status information of the distributed system and the resource request information of the plurality of subtasks, target nodes corresponding to the plurality of subtasks are sequentially determined according to a specific order.

Optionally in one example, the resource requirement information of the plurality of sub-tasks comprises resource requirement information of each sub-task of the plurality of sub-tasks, wherein the resource requirement information comprises a requested resource type and quantity of each type of resource, or It also contains other information.

Optionally in one example, the current resource state information of the distributed system includes current state information of a plurality of nodes of the distributed system, wherein the current state information includes whether resources are available, the type and quantity of resources available, It may be used to indicate at least one of load status, topological connectivity information, or may contain other information as well.

In some embodiments, each time a target node to assign to a subtask is determined, the current resource state information of the distributed system is updated, and the target node corresponding to the next subtask is determined based on the updated current resource state information. decide.

In some embodiments, if the target node corresponding to at least one subtask of the plurality of subtasks is not successfully determined, currently available resources of the distributed system are used for each subtask of the plurality of subtasks. determines that it does not meet the resource requirements of

In some embodiments, the order of sequentially determining target nodes of the plurality of subtasks is derived from at least one of priority of subtasks and dependencies between subtasks.

In some embodiments, determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks comprises current resource state information of the distributed system. determining a preselected node set for each subtask of the plurality of subtasks based on the resource requirement information of the plurality of subtasks; and a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks. and selecting.

However, different subtasks may have the same or different preselected node-sets. In some examples, multiple subtasks have the same preselected node-set.

In some embodiments, determining a preselected node set for each subtask of the plurality of subtasks based on current resource state information of the distributed system and resource demand information of the plurality of subtasks comprises: determining a preselected node-set for each subtask based on the current resource state information of each subtask and the resource requirement information of each subtask.

The determination of preselected node-sets for multiple sub-tasks may be independent and independent of each other, eg, executed in parallel or in any order. For example, the preselection node-sets of multiple subtasks are determined based on the same distributed system's current resource state information, ie, the distributed system's current resource state information is not updated during the preselection process. For example, the determination of each subtask's preselected node set is only related to its own resource requirement information and is independent of the resource requirement information of other subtasks.

In some embodiments, determining a pre-selected node-set for each sub-task of the plurality of sub-tasks includes sequentially determining a pre-selected node-set for each sub-task of the plurality of sub-tasks according to a particular order.

In one example, preselected nodes of a preselected node-set of a first subtask of the plurality of subtasks are selected from a preselected nodeset of a second subtask of the plurality of subtasks, wherein the order of the second subtasks is the first subtask. Positioned before subtasks.

In another example, determining a preselected node set for each subtask of the plurality of subtasks based on resource requirement information of the plurality of subtasks.

In some embodiments, selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks comprises selecting a target node for each subtask from a preselected node for each subtask of the plurality of subtasks. sequentially according to a particular order.

Optionally in one example, preferentially selecting a target node of a second subtask of said plurality of subtasks as a target node of a first subtask of said plurality of subtasks, wherein said second subtask is ordered before said first subtask. Located in

In some embodiments, selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks includes selecting a target node for each subtask based on current resource state information for the distributed system. determining a target node of the first subtask from a set of preselected nodes of one subtask; and updating current resource state information of the distributed system based on the resource request information of the first subtask to update the distributed system. determining a target node of the second subtask from a preselected node set of a second subtask of the plurality of subtasks based on the obtained current resource state information.

Optionally in one example, determining a target node of the first subtask from the preselected node-set of the first subtask based on current state information of each preselected node of the preselected nodeset of the first subtask; Update the current state information of the target node of the first subtask according to the resource requirement information of the subtask.

In some embodiments, selecting a target node for each subtask from a preselected node-set for each subtask of the plurality of subtasks includes including in each subtask's preselected node-set based on the task type of the task. and selecting a target node for each subtask from each subtask's preselected node-set based on the scores of the preselected nodes included in each subtask's preselected node-set. including.

Optionally, in one example, determining a scoring scheme for each preselected node of the preselected node-set based on the task type of the task.

In some embodiments, assigning respective target nodes to a plurality of subtasks included in the task based on the task information includes determining a task type of the task; and assigning respective target nodes to a plurality of subtasks included in the task based on.

In some embodiments, the task type of the task is computationally intensive or communication intensive.

In some embodiments, assigning respective target nodes to a plurality of subtasks of said task comprises dividing a plurality of subtasks of said task into at least one group, each group comprising said including at least one sub-task of a plurality of sub-tasks; and assigning each group of said at least one group a respective corresponding target node, wherein each sub-task of the same group is assigned to the same target node. .

In some embodiments, dividing the plurality of subtasks of the task into at least one group comprises dividing the plurality of subtasks of the task into at least one group based on a task type of the task. .

In some embodiments, dividing the plurality of subtasks of the task into at least one group is based on resource requirement information of the plurality of subtasks of the task when the task type of the task is communication intensive. determining a total node quantity required by the plurality of subtasks, and dividing the plurality of subtasks of the task into at least one group based on the total node quantity.

In some embodiments, dividing the plurality of subtasks of the task into at least one group includes dividing each subtask of the plurality of subtasks into one group if the task type of the task is computationally intensive. including the step of

In some embodiments, assigning respective target nodes to each of said at least one group comprises determining respective pre-selected node sets for each of said at least one group; sequentially selecting a target node of each group from the set of preselected nodes of each group of .

In some embodiments, assigning respective target nodes to a plurality of subtasks included in the task based on a task type of the task includes: determining a target node for a subtask; and assigning target nodes to a plurality of subtasks included in the task upon successfully determining a corresponding target node for each of the plurality of subtasks of the task. .

In some embodiments, the method performs delayed assignment for all of a plurality of subtasks of the task if the target node corresponding to at least one subtask of the subtask is not successfully determined. Further including steps.

In some embodiments, the method includes: if the currently available resources of the distributed system do not meet the total resources corresponding to the resource requirement information of the subtasks of the task; It further includes the step of performing delayed allocation for all.

In some embodiments, the method further comprises synchronously scheduling a plurality of subtasks of the task after assigning respective target nodes to the plurality of subtasks included in the task.

In some embodiments, the method is applied to a task scheduling system.

In a second aspect of an embodiment of the present disclosure, an obtaining module for obtaining task information, wherein the task information includes at least one of a task type of the task and resource requirement information of the task; , a first assignment module for assigning respective target nodes to a plurality of subtasks included in the task based on the task information.

In a third aspect of the embodiments of the present disclosure, there is provided a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method of any embodiment.

A fourth aspect of embodiments of the present disclosure includes a memory, a processor, and a computer program stored in said memory and operable on said processor, wherein the method of any embodiment when said processor executes said program. A computing device is provided for implementing

In an embodiment of the present disclosure, task information is obtained, the task information includes at least one of a task type of the task and resource requirement information of the task, and the resource requirement information of the task is a plurality of tasks included in the task. subtask resource request information. Allocation of target nodes based on task type or task resource requirement information. For example, if the currently available resources of the distributed system satisfy the total resources corresponding to the resource requirement information of the subtasks of the task, the multiple subtasks included in the task are assigned respective target nodes. In this way, it is possible to grasp the resource request status of the entire task, implement resource allocation with the granularity of the entire task, and improve the resource utilization rate of the cluster.

It should be noted that the foregoing general description and the following detailed description are for purposes of illustration and interpretation and are not limitations on the present disclosure.

The drawings shown here are incorporated as part of the present specification, and the drawings show embodiments consistent with the present disclosure, and together with the present specification serve to explain the technical solutions of the present disclosure.
1 is a schematic diagram of a distributed system according to an embodiment of the present disclosure; FIG. 4 is a flowchart of a task scheduling method according to an embodiment of the present disclosure; FIG. 4 is a schematic diagram of resource changes in the process of node simulation allocation according to an embodiment of the present disclosure; FIG. 4 is a flow chart of a process for simulation assignment of nodes in accordance with an embodiment of the present disclosure; FIG. 1 is a schematic diagram of a conventional task scheduling process; FIG. FIG. 4 is a schematic diagram of the process of task scheduling according to an embodiment of the present disclosure; FIG. 4 is a schematic diagram of the scheduling logic of an embodiment of the present disclosure; 1 is a structural schematic diagram of a task scheduling device according to an embodiment of the present disclosure; FIG. 1 is a structural schematic diagram of a computing device according to an embodiment of the present disclosure; FIG.

Exemplary embodiments are described in detail below, examples of which are illustrated in the drawings. Where drawings are involved in the following description, the same numbers in different drawings represent the same or similar elements, unless otherwise stated. The embodiments described in the illustrative examples below do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of this disclosure as recited in the claims.

The terminology used in this disclosure is for the purpose of describing particular embodiments and is not intended to be limiting of the disclosure. As used in this disclosure and claims, the singular forms "a", "said" and "the" include the plural unless the context clearly dictates otherwise. As used herein, the term "and/or" includes any and all possible combinations of one or more of the associated items. In addition, the term "at least one" as used herein represents any one of a plurality of types or any combination of at least two of a plurality of types.

It should be noted that although the terms first, second, third, etc. are used to describe various elements in this disclosure, such elements are not limited to these terms. The terms are only used to distinguish elements of the same type from each other. For example, a first element could be termed a second element without departing from the scope of the present disclosure. Similarly, the second element may be referred to as the first element. Depending on the context, the term "if" as used herein can be interpreted as "when" or "if" or "if determined".

Currently, more and more tasks are processed using distributed systems, and such tasks are called distributed tasks. Such distributed tasks require high performance processing resources such as Graphics Processing Units (GPUs), Digital Signal Processors (DSPs), Field Programmable Gate Arrays (FPGAs) among others. It includes demanding tasks, for example deep learning tasks. FIG. 1 illustrates a distributed system of some embodiments, including one or more clusters, each cluster including one or more servers, each server may be considered a node. (Indicated by each black dot in the figure). Each node includes at least one of resources such as a central processing unit (CPU), GPU, memory, disk, network port, and the like. After subtasks are assigned to nodes, the subtasks can be executed by resources on the node, and each node can execute one or more subtasks. The subtasks performed by each node in the same cluster may be the same or different. In a distributed system, both the tasks themselves and the resource requirements of the tasks are completely different from conventional tasks. You need to increase task execution performance and cluster resource utilization.

When allocating resources to tasks in a distributed system, the allocation method affects the efficiency of task execution. For example, in some cases frequent communication is required between multiple subtasks of a task, so it is preferable to assign the multiple subtasks of the task to nodes with low communication costs. In some cases, the process of executing multiple subtasks of a task involves a large amount of computation, so it is preferable to assign the multiple subtasks of the task to nodes with more computing resources.

Furthermore, most of the tasks in conventional distributed systems have the following characteristics. A task generally includes multiple subtasks, each of which may or may not have exactly the same content and required resources, and may require frequent communication between the subtasks. Also, in some cases, multiple subtasks of a task need to be scheduled and executed together, otherwise the task cannot be executed successfully. In some cases, the execution order of each subtask has some influence on the execution status of the entire task. For example, a slave subtask cannot run until the master subtask to which it belongs has finished running.

Currently, the following schedulers (eg, kube-batch) perform batch scheduling for tasks in distributed systems in the following manner. The scheduler attempts to schedule each subtask of a task one by one in order, and after it has completed attempting to schedule each subtask of that task, determines whether the task as a whole meets the user's expected execution state. do. This scheduling method is called gang scheduling.

However, in current implementations of gang scheduling, subtasks are practically used as the granularity for scheduling. might not be the best choice, resulting in poor task execution efficiency and cluster resource utilization. Therefore, scheduling failures often occur when cluster resources are insufficient or task types are complex.

Based on this, embodiments of the present disclosure provide a task scheduling method. As shown in FIG. 2, the method may include steps 201 and 202. FIG.

In step 201, task information is obtained, the task information including at least one of a task type of the task and resource requirement information of the task.

In step 202, a plurality of subtasks included in the task are assigned respective target nodes based on the task information.

The method of the embodiments of the present disclosure may be performed by any electronic device such as a terminal device or a cloud server. In some embodiments, the method is performed by a processor or scheduler. Optionally, the processor or scheduler may run on a cloud platform such as Kubernetes, and specifically set in a container orchestration engine, although embodiments of the present disclosure are not so limited. At step 201, information for one or more tasks is obtained.

The plurality of subtasks included in the task may be all subtasks of the task, or some subtasks of the task, such as subtasks with high resource demand, subtasks with high priority, or mutual dependencies. , subtasks with similar resource types or quantities, and so on. However, subtasks having a mutually dependent relationship refer to a plurality of subtasks having a mutually dependent relationship or a master-slave relationship. For example, subtask A cannot be executed until execution of subtask B is completed. Another example refers to multiple subtasks having the same type, or having at least one specific type, and the like. Said tasks may be neural network training tasks, inference tasks, or other types of deep learning tasks. Optionally, in some embodiments, each subtask of the same task is executed by the same or different nodes of the same cluster of the distributed system. When a distributed system includes multiple clusters, each subtask of the same task is executed by nodes of the same or different clusters. Optionally, a task is assigned to a cluster and then target nodes for subtasks of the task are determined from the cluster. In this way, assigning multiple subtasks of a task to the same cluster for execution helps reduce communication overhead and improve task execution efficiency.

The task information may be obtained based on user input or settings, or may be obtained based on task analysis or the like. The task information may include the task type and/or resource requirement information of the task. Specifically, the resource request information of the task is resource request information of a plurality of subtasks included in the task. In some embodiments, the task information includes priority information of the task, priority information of multiple subtasks included in the task, dependency information between subtasks, user-provided information, in addition to task type and resource requirement information. information, history information, calculation amount information, communication amount information, and the like.

The resource request information of the subtask may include a resource type and/or resource quantity required for execution of the subtask, and the resource type is a resource such as CPU, GPU, DSP, FPGA, memory, disk, network port, etc. including but not limited to at least one of

The priority information of the task includes, but is not limited to, the waiting time before the task is scheduled and/or the total resources required by the plurality of subtasks of the task. Latency to schedule refers to the period from the time a task is received by the scheduler to the time it is scheduled, and the task with the longer latency to be scheduled is more demanding of total resources. By setting a high priority with the task, the task is prevented from waiting for scheduling for a long time.

The priority information of the plurality of subtasks included in the task is used to determine the priority of the plurality of subtasks included in the task. In some embodiments, scheduling of sub-tasks can be performed based at least on the priority information of each sub-task in one scheduling cycle or resource allocation process.

The dependency information between the subtasks is used to determine a dependency relationship or a master-slave relationship between a plurality of subtasks included in the task. depends on subtask A, or subtask A is a master task and subtask B is a slave task of subtask A.

The user-provided information is used to determine computational and communication costs of subtasks of the task.

The history information may include past scheduling information and/or past execution status. A task type of the task in past scheduling can be determined based on the historical information.

In addition to the above information, according to the actual application scenario, the task information may contain other information, which will not be repeated here.

In some embodiments, assigning 202 respective target nodes to a plurality of subtasks included in the task based on the task information includes the step 202 of assigning respective target nodes to a plurality of subtasks included in the task, wherein the currently available resources of the distributed system are limited to the plurality of subtasks of the task. Allocating a respective target node to each of a plurality of subtasks included in the task if the total resource corresponding to the resource requirement information of the subtask is satisfied. The total resource corresponding to the resource requirement information of the multiple subtasks of the task may be the total resource required of the multiple subtasks of the task, optionally if the multiple subtasks are all subtasks of the task. and the total resource is the total resource required by the task. optionally, the currently available resources of the distributed system are sufficient to support execution of the plurality of sub-tasks if the currently available resources of the distributed system satisfy the plurality of sub-tasks or the total resource requirements of the tasks; indicates that Optionally, if the currently available resources of the distributed system do not meet the total resource requirements of the multiple subtasks or tasks, the currently available resources of the distributed system may at best perform some of the tasks of the multiple subtasks. Not supported, indicating that execution of all subtasks cannot be supported.

In an embodiment of the present disclosure, multiple subtasks included in a task may be assigned corresponding target nodes only if the currently available resources of the distributed system satisfy the total resource requirements of the multiple subtasks of the task. assign. As a result, a task is used as a unit of resource allocation, and a plurality of subtasks included in the task are collectively allocated to nodes. Therefore, if the currently available resources of the distributed system are insufficient, only some of the subtasks of a task cannot successfully allocate resources to some of the subtasks, so that only some of the subtasks of that task can use the entire resources in the distributed system. It helps avoid hogging situations and makes rational use of resources.

In some embodiments, target nodes are assigned to the multiple subtasks based on the current resource state information of the distributed system and the resource requirement information of the multiple subtasks. Each subtask may be assigned a target node that satisfies its resource requirements, and the target nodes for different subtasks may be the same or different.

In some embodiments, determining a target node for the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks, corresponding to each subtask of the plurality of subtasks. If the target node is successfully determined, assign a corresponding target node to each of the plurality of subtasks included in the task. The above process of determining target nodes for the plurality of subtasks is a process of simulation allocation, which is a virtual resource allocation process rather than an actual execution resource allocation process. The simulation allocation process currently computes and determines whether to allocate multiple subtasks to nodes of a distributed system. By performing a simulated allocation of nodes for subtasks of the task, it is possible to determine whether the distributed system currently meets the resource requirements of the subtasks.

the current resource status information of the distributed system is used to display the currently available resources of the distributed system and may include at least one of the quantity, type and distribution of the currently available resources of the distributed system; Or, it may include the current state of multiple nodes of the distributed system, such as availability, whether there are resources available, the type and quantity of resources available, and the like. The resource requirement information of the plurality of subtasks may include resource requirement information of each subtask of the plurality of subtasks, for example, at least one of a requested resource type and a resource quantity of a resource type, for example, a request GPUs, DSPs, CPUs, etc., and embodiments of the present disclosure are not limited thereto.

In some embodiments, determining whether there is a target node in the distributed system to meet the requirements of the subtasks based on the resource request information of the multiple subtasks and the current resource state information of the distributed system; or A target node for each subtask of the plurality of subtasks is sequentially determined according to a specific order. For example, when each subtask of a plurality of subtasks finds a target node in the distributed system that satisfies its requirements, it determines that the currently available resources of the distributed system satisfy the total resource requirements of the plurality of subtasks. As an example, if at least one subtask of the plurality of subtasks cannot find a target node satisfying its requirements in the distributed system, it determines that the currently available resources of the distributed system do not meet the total resource requirements of the plurality of subtasks.

Unlike the conventional node determination of a single subtask, in the embodiments of the present disclosure, the target nodes of multiple subtasks are sequentially determined according to a specific order, and the determination of the target node of the next subtask is based on the previous subtask. May be affected by the determined target node. In some embodiments, after the target node corresponding to one subtask is determined each time, update the current resource state information of the distributed system, for example, update the current state information of the determined target node, and then , avoiding the problem of resource contention by determining the target node corresponding to the next subtask based on the updated current resource state information.

It should be noted that updating the current resource state information here is used in the above simulation allocation or target node determination process, and is not an actual update but a virtual update, and a subtask node simulation allocation or It may be a virtual change situation of the currently available resources of the distributed system after performing the determination of the target node once. Since the actual assignment of subtasks has not yet taken place, the currently available resources of the distributed system have not actually changed. For example, after the target node of one subtask is determined, the target node is virtually assigned to the subtask, and some resources corresponding to the resources required by the subtask on the target node are virtually allocated to the subtask. You can think of it as binding. As such, the virtually bound resource cannot be assigned to another subtask.

The above specific order may be a preset order. Alternatively, in some examples, the particular order is determined based on the priorities of the plurality of subtasks, e.g., determining target nodes corresponding to higher priority subtasks, then lower priority subtasks. Determine the target node corresponding to the subtask. The priority may be determined based on one or more factors such as the type of subtask, the dependencies between the subtasks and other subtasks, e.g. For subtasks, the priority of a master subtask is higher than the priority of a slave subtask, although embodiments of the present disclosure are not so limited. Or, in some other examples, the particular order is determined based on a dependency or master-slave relationship of the subtasks. Since a slave subtask cannot run without a master subtask, preferentially assigning the corresponding target node to the master subtask helps improve the execution efficiency of the overall task. Alternatively, the particular order may be determined based on other factors, although embodiments of the present disclosure are not limited in this respect.

In some embodiments, if the distributed system has a target node that satisfies the resource requirements of each subtask of said plurality of subtasks, the currently available resources of the distributed system contain resource request information of the plurality of subtasks of said task. otherwise, the currently available resources of the distributed system do not meet the total resources corresponding to the resource requirement information of the plurality of subtasks of the task.

In the example shown in FIGS. 3A and 3B, assuming that the task includes subtask 1, subtask 2 and subtask 3 in order of priority, first, the currently available resource information of the distributed system and the resource request information of subtask 1 are based on, perform a simulation of allocating nodes to subtask 1, obtain the target node of subtask 1, then update the current resource state information of the distributed system, and then update the current available resource information of the distributed system and Based on the resource request information of subtask 2, perform a simulation of assigning nodes to subtask 2, obtain the target node of subtask 2, then update the current resource state information of the distributed system, and further update the distributed system again. and the resource requirement information of subtask 3, a simulation of allocating nodes to subtask 3 is performed to obtain the target nodes of subtask 3.

If the simulation allocation of nodes to subtask 1, subtask 2 and subtask 3 are all successful, the currently available resources of the distributed system satisfy the demands of each subtask of the plurality of subtasks, i.e. the demands of the plurality of subtasks. determines that it satisfies the total resources to be used. If the simulated assignment of a node to subtask 1, subtask 2 or subtask 3 fails, e.g., the distributed system cannot find a target node that satisfies its required resources, then the currently available resources of the distributed system may fall short of the plurality of nodes. Determines that the subtask's total resource requirements cannot be met.

a target node corresponding to each subtask determined in the simulation allocation when the simulation allocation determines that the currently available resources of the distributed system meet the resource requirements of each subtask of the plurality of subtasks; It may be assigned to each of the subtasks.

In some embodiments, the target node for each subtask may be determined based on the current state information of each node of the distributed system, e.g., sorting or scoring each node based on the current state information of each node. The order or score nodes are assigned sequentially to the subtasks according to a specific order.

In some embodiments, when determining target nodes for multiple subtasks of the task, the target nodes corresponding to each of the multiple subtasks are determined through two processes of preselection and optimal selection.

In the pre-selection process, for each sub-task of the plurality of sub-tasks, based on the resource requirement information of the plurality of sub-tasks, each pre-selection node is preliminarily selected from the distributed system to satisfy its requirement, or Eliminate nodes that do not meet the requirements. For example, determining a preselected node set for each subtask of the plurality of subtasks based on the current state information of the plurality of nodes of the distributed system and the resource request information of the plurality of subtasks; contains at least one preselected node. The preselected node-sets for different subtasks may contain the same nodes or different nodes.

The optimal selection process selects a target node for each subtask from each subtask's preselected node set. For example, an optimal node from the subtask's preselected node set is selected as the target node for the subtask. Optionally, in some examples, based on the current state information of each preselected node in the subtask's preselected node-set, e.g., current load situation, type of available resources, quantity of available resources, utilization A target node is determined based on at least one of the distribution of possible resources and topological connectivity between nodes and topological connectivity with other nodes. The types of available resources include, but are not limited to, at least one of communication ports, disk storage space, memory, CPU, GPU, and the like. The quantity of available resources may be the quantity of CPUs, remaining disk capacity, remaining memory capacity, and the like. The distribution of available resources is also called resource fragmentation situation, that is, the distribution position of currently available resources, for example, how many discs the remaining capacity of a disc is distributed over. . The topology structure of each node of the distributed system is, for example, bus topology, star topology, ring topology, tree topology, and the like.

In some embodiments, the preselection process determines a set of preselected nodes for each of the subtasks based on current resource state information of the distributed system and resource requirement information of each subtask, i.e., a plurality of the tasks. do not interfere with each other, and the selection of the preselection node set of each subtask does not depend on the resource requirement information of other subtasks, but is only related to the resource requirement information of that subtask. Accordingly, determination of preselected node sets for multiple subtasks may be performed in parallel or in any order. In this way, preselected nodes can be quickly determined to improve the overall efficiency of simulation allocation.

For example, assuming a task includes subtask 1, subtask 2 and subtask 3 in order of priority, the preselection process may perform the following operations in parallel. determining a preselected node set for subtask 1 based on the resource request information of subtask 1 and the current resource state information of the distributed system; Determining a preselected node-set and determining a preselected node-set for subtask 3 based on the resource requirement information of subtask 3 and the current resource state information of the distributed system. For example, the preselected node-set corresponding to subtask 1 includes {node 1, node 2, node 3}, and the preselected node-set corresponding to subtask 2 includes {node 2, node 5, node 6, node 7}. , and the preselected node-set corresponding to subtask 3 includes {node 6, node 7}.

In some embodiments, the preselection process sequentially selects preselected node-sets of multiple subtasks according to a particular order. In an optional example, determine the preselected node-set for the next subtask based on the preselected node-set for the previous subtask. For example, the next subtask's preselected nodes are selected from the previous subtask's preselected node set. The number of the previous subtasks may be one or more, for example, the previous subtasks are all subtasks before the next subtask, and the previous subtask and the next subtask are: They may be adjacent or separated by at least one subtask, although embodiments of the present disclosure are not so limited.

In some embodiments, resource requirement information of a plurality of subtasks are collectively considered to determine a preselected node-set for a plurality of subtasks, for example, selecting a common preselected node-set for the plurality of subtasks. , the preselected nodes included in the common preselected node set satisfy the resource requirements of at least two subtasks of the plurality of subtasks.

In some embodiments, the optimal selection process determines target nodes for multiple subtasks in parallel or combines with preselected node sets for multiple subtasks to jointly determine target nodes for multiple subtasks. . For example, the target nodes of at least some of the subtasks are determined by intersecting preselected node sets of at least some of the subtasks, but embodiments of the present disclosure are not limited thereto.

In some embodiments, target nodes corresponding to the plurality of subtasks are sequentially determined from the preselected node sets of the plurality of subtasks according to a particular order, for example, based on priorities of the plurality of subtasks. In one example option, each time a target node corresponding to a subtask is determined, the current state information of that target node is updated. For example, the current state information of the target node is updated based on the resource requirement information corresponding to the subtask. In this way, in the process of determining the target node of the succeeding subtask, the part corresponding to the resource required by the subtask on the target node is not available, and the smooth execution of the multiple subtasks is guaranteed. Optionally, in another example, the target node for the next subtask is selected based on the target node selected by the previous subtask. For example, unless the next subtask's preselected node-set does not contain the target node or other significant elements of that target node satisfy the next subtask's requirements, the next subtask will be is preferentially selected as its target node, although embodiments of the present disclosure are not so limited.

Said pre-selection and optimal selection process may be a process of simulated allocation when resources are allocated to multiple sub-tasks synchronously. Said pre-selection and optimal selection process may be the actual node allocation process when allocating resources to multiple sub-tasks asynchronously. Similarly, the updating of current resource state information involved in the above process is the actual updating process performed due to changes in the currently available resources of the distributed system after the execution of node assignment to subtasks has been completed once. may

Embodiments of the present disclosure may select a target node from a preselected node set according to a predetermined scheme. For example, the target node is selected based on the current resource state information of the distributed system or the current state information of each preselected node and the resource requirement information of the plurality of subtasks. In some embodiments, the score of the preselected nodes in each subtask's preselected node-set is determined, e.g., the current state information of each preselected node in the subtask's preselected node-set and the resource requirement information of the subtask. and determining the score of each preselected node based on and, based on the score of the preselected nodes in each subtask's preselected node set, extracting a target node for each subtask from the preselected node set for each subtask. select. For example, the preselected node with the highest score from at least one preselected node included in the preselected node set is selected as the target node. Alternatively, select target nodes based on both score and other factors.

Continuing with the previous example, taking the example of sequentially determining the target nodes of multiple subtasks based on the task priority, first determine the score of each preselected node in the preselected node set of subtask 1, and the score of node 1 is 80, the score of node 2 is 90, and the score of node 3 is 70, determine node 2 as the target node for subtask 1 and update the current state information of node 2. Then determine the score for each preselected node in the preselected node set for subtask 2, where the score for node 2 is determined based on the updated current state information for node 2. Assuming that node 2 has a score of 60, node 5 has a score of 80, node 6 has a score of 75, and node 7 has a score of 70, determine node 5 as the target node for subtask 2, and determine node 5's current state Update information. Finally, determine the score of each preselected node in the preselected node set for subtask 3, and assuming node 6 has a score of 70 and node 7 has a score of 60, determine node 6 as the target node for subtask 3. do.

In some embodiments, the task type of the task is determined based on the task information obtained in step 201, and target nodes are assigned to multiple subtasks based on the task type of the task. The task type of the task can be set according to the actual situation. For deep learning tasks such as training deep learning models such as neural networks, inference, etc., the tasks are optionally divided into two types: communication-intensive tasks and computation-intensive tasks. A communication-intensive task is a task in which communication between subtasks is frequent in the process of task processing, and a computation-intensive task is a task with a large amount of calculation in the process of task processing.

Specifically, the task type of a task can be determined based on the overall status of multiple subtasks of the task. Optionally, in one example, determining a task type for a task based on computational and communication complexity of multiple subtasks of the task. For example, if the complexity is greater than a preset complexity threshold, the task is considered computationally intensive. As another example, the task is considered a communication-intensive task if the traffic is greater than a preset traffic threshold. As another example, a task may be determined as a communication-intensive task if both computational complexity and communication complexity exceed corresponding preset thresholds. Optionally, in another example, determine the task type based on past scheduling experience. For a given task, if the scheduler determined the task as a computation-intensive task in past scheduling, the task is determined as a computation-intensive task, and the scheduler determined the task as a communication-intensive task in the past scheduling. Once determined as a task, the task is determined as a communication-intensive task. optionally determining a task type of the task based on the circumstances of at least one past execution of the task, e.g., during the most recent one or more past executions, to be used in the computation of the task; determining the task as a computationally intensive task when there is a large amount of time or computing resources to perform the task; If there are many resources, the task is determined as a communication-intensive task. Alternatively, another method may be used to determine the task type of the task. Alternatively, the task type of the task is determined based on user-provided information, such as user-provided task type information, or user-provided task calculation amount and/or communication amount information. Embodiments of the present disclosure do not specifically limit the implementation of task type determination.

Embodiments of the present disclosure may efficiently execute the overall task by determining target nodes for multiple subtasks based on the task type. For example, for a communication-intensive task, multiple subtasks of the task can be assigned to as few target nodes as possible to reduce communication overhead. As another example, for a computationally intensive task, optionally the resources on the target node only need to meet the demands of multiple subtasks of said task, or preferentially selecting nodes with superior computational performance, or or to optimize the communication overhead on the premise that the computing resources meet the user's demand as much as possible.

Optionally, in some examples, the above pre-selection and/or optimal selection process is performed based on the task type of the task.

In some embodiments, the above pre-selection process is performed based on the task type of the task, e.g., for communication intensive tasks, the type of resources available and/or the quantity of resources available is low. Eliminate nodes that clearly do not satisfy the requirements of many subtasks of multiple subtasks, such as Nodes with many resource types and/or many in quantity, or nodes whose resource types match the resources required by all of the subtasks, are preferentially selected as preselected nodes. Optionally, in some instances, instead of determining separate preselected node-sets for each subtask of a communication-intensive task, one preselected node-set may be determined for some or all subtasks of multiple subtasks. Good, but embodiments of the present disclosure are not so limited.

In some other embodiments, the above optimal selection process is performed based on task type. For example, the task type determines the target node selection method. For example, in the process of determining the target node of one subtask, preferentially select the target node determined by the previous subtask. For example, if a task includes subtask 1 and subtask 2, and the target nodes of subtask 1 include node 1 and node 2, subtask 2 may optimally select node 1 and node 2 as target nodes. Optionally, in some examples, based on the current state information of each node in the subtask's preselected node-set and the task type of the task, calculate the score of each preselected node in the subtask's preselected node-set. decide. In some other embodiments, other methods may be used to determine the score of each preselected node included in the preselected node set of the subtask, and the description is omitted here for redundancy.

For example, in the case of a communication-intensive task, it is desirable to reduce the number of target nodes as much as possible to reduce the communication cost between subtasks. Nodes that do have a higher score than nodes with distributed resources. For example, if the available disk capacity of preselected node 1 of a subtask is distributed over two disks and the available disk capacity of preselected node 2 is distributed over one disk, then preselected node 2 of that subtask is preselected Higher score than node 1. In another example, for computationally intensive tasks, nodes with high computational performance generally score higher than those with low computational performance. For example, if the processor of preselection node 1 of the subtask contains one processor core and the processor of preselection node 2 of the subtask contains two processor cores, preselection node 2 of the subtask will score more than preselection node 1. is high.

In the above process, an example is given in which subtasks are subject to resource allocation. In some embodiments, a plurality of sub-tasks of said task may be divided into at least one group, wherein each group includes at least one sub-task of said plurality of sub-tasks; Each group is assigned a corresponding target node, provided that each subtask in the same group is assigned to the same target node.

The number of groups may be the same as the number of nodes required by the task, i.e., when assigning multiple subtasks of a task to N target nodes, divide the multiple subtasks of the task into N groups. The number of subtasks in each group can be the same or different. For example, assign multiple subtasks at random, divide subtasks with similar priorities into one group, divide subtasks with or without dependencies into one group, or subtasks with large differences in required resource quantities. are divided into one group, or subtasks that require the same resource type are divided into one group, and the embodiments of the present disclosure do not specifically limit the method of group division. Determining target nodes on a group-by-group basis in this manner can improve the efficiency of the resource allocation process without significantly impacting the efficiency of task execution.

In some embodiments, subtasks of the task are divided into at least one group based on a task type of the task. For example, in the case of a communication-intensive task, subtasks of the task can be assigned to a minimum number of target nodes to reduce communication costs between subtasks. For example, by dividing all subtasks of a plurality of subtasks into the same group, all subtasks are executed on the same target node. Optionally, in some examples, determining a total node quantity required of a plurality of subtasks of the task based on resource requirement information of a plurality of subtasks of the task; subtasks into at least one group. As another example, in the case of a computation-intensive task, each subtask of a plurality of subtasks of the task is grouped into a group, that is, the target node is determined in units of subtasks.

For example, assuming that one task contains four subtasks, each subtask requires one GPU, and the quantity of available GPUs in each node of the distributed system is two, then the four subtasks of said task are: Requires at least two nodes to run. Therefore, if the task is a communication-intensive task, the four subtasks may be evenly divided into two groups, but the grouping may be done randomly, or the resource requirement information of the subtasks, etc. can be done on the basis of If the task is a computationally intensive task, the four subtasks may be divided into four groups, each subtask being one group.

In the case of division into groups, optionally, the determination of the target nodes of the subtasks above also applies to the determination of the target nodes of the groups, which will be explained in detail below.

For example, in some embodiments, if each group of at least one group of the task has a corresponding target node, the currently available resources of the distributed system demand multiple subtasks of the task. Decide to meet total resources. In some other embodiments, if there are no target nodes corresponding to at least one group of said task, the currently available resources of the distributed system do not meet the total resource requirements of the plurality of subtasks of said task. and decide.

As another example, based on current resource status information of the distributed system and resource request information of each group of the at least one group, sequentially determine a target node for each group of the at least one group according to a specific order. . The resource requirement information of the group may be determined based on the resource requirement information of at least one subtask included in the group, for example, the resources required by the group are of the same type required by all subtasks included in the group. sum of resources. Optionally, said particular order is determined based on at least one of priorities and dependencies of said plurality of groups, or other factors. The priority of the group may be determined based on the priorities of the two or more subtasks included in the group, for example determining the priority of the group as the highest priority of the two or more subtasks, or such as determining the priority as the average of the priorities of two or more subtasks. Dependencies between groups may be determined based on dependencies between subtasks included in different groups, e.g., suppose one subtask in group 1 depends on one subtask in group 2, then group 1 is group 2, etc., although embodiments of the present disclosure are not so limited.

In some embodiments, determining a preselected node set for each group based on current resource status information of the distributed system and resource request information for each group, and determining the preselected node set for each group from the preselected node set for the group. Target nodes are sequentially determined according to a predetermined order.

In an embodiment of the present disclosure, first determine the target node for the entire task, for example, perform the above preselection and optimal selection process, finally obtain the information of the target node corresponding to the entire task, and then Tasks are divided into subtask levels and bound, that is, target nodes are bound on a subtask basis.

In some embodiments, the target node corresponding to at least one subtask of the plurality of subtasks, or when the currently available resources of the distributed system do not meet the total resource requirements of the plurality of subtasks of the task. If the determination is unsuccessful, a delayed assignment is made to all of the subtasks included in the task. Said delayed allocation may be to allocate a plurality of subtasks included in said task in the next allocation cycle. In the next allocation cycle, subtasks of said task are assigned respective target nodes only if the currently available resources of the distributed system satisfy the total resource requirements of the subtasks of said task. . If the currently available resources of the distributed system still do not meet the total resource requirements of the multiple subtasks of the task in the next allocation period, then the delayed allocation continues.

In some embodiments, each time a target node corresponding to a subtask or group is determined, a virtual assignment to the target node corresponding to the current subtask or group, or a virtual assignment on the target node corresponding to the current subtask or group Make virtual bindings to resources. The above virtual allocation or A virtual binding is considered invalid, so that the corresponding target node and its resources can be assigned to other tasks.

In some embodiments, after assigning respective target nodes to the subtasks of the task, the subtasks of the task are synchronously scheduled.

It should be noted that, in the embodiments of the present disclosure, allocation means handing out the subtasks to the corresponding target nodes, and scheduling means that the corresponding target nodes execute the subtasks after handing the subtasks to the corresponding target nodes. is.

As shown in FIG. 4, in the conventional scheduling method, to realize gang scheduling, the scheduler processes multiple subtasks of a task in order, and one subtask is scheduled (in the embodiment of the present disclosure, As soon as possible (rather than simulated assignment as described), the subtask is actually assigned to the corresponding target node and the resources on the target node are used to schedule the subtask. When assigning each subtask of a task to a task, the currently available resources dynamically change and gradually decrease. Such a scheduling method is called a task-by-task scheduling method. call. Moreover, when scheduling subtasks, the scheduling situations and resource requirements of multiple subtasks of the same task are not comprehensively considered, so each assignment is an independent behavior. However, in practice, there is often some relationship between multiple subtasks within the same task. If it is assigned to multiple nodes with large values, it will greatly affect the execution efficiency of the entire task.

In the embodiments of the present disclosure, subtasks are assigned and scheduled using tasks as granularity. As shown in FIG. 5, we first get all the tasks still by priority, and for each task we consider the entirety of the task holistically, and determine if the currently available resources of the distributed system are of the multiple subtasks of the task. Allocating each of the sub-tasks in the task to their corresponding target nodes only if the total required resources are met, and then synchronously scheduling the sub-tasks, i.e. resource allocation and task-by-task unit. Do scheduling.

Unlike previous task-by-task scheduling methods based on subtask requirements, embodiments of the present disclosure consider the entire task holistically when selecting nodes for a task. By using the minimum unit of scheduling, the interrelationships and requirements of each subtask within a task can be better utilized, and task execution efficiency can be improved by selecting a more suitable scheduling node for the entire task from the cluster. And can improve the resource utilization efficiency of the cluster.

FIG. 6 shows a schematic diagram of the scheduling logic of some embodiments of the present disclosure. To reduce the scheduling complexity, we make the following assumptions for each task to be scheduled.
(1) Each task includes multiple subtasks.
(2) Each subtask is homogeneous in resource requirements, i.e., has the same resource requirements.
(3) When each user submits a task, the requested resource amount of each subtask is specified at the time of submission, and the number of subtasks and the requested resource amount of each subtask do not change during the entire process of task processing.

First, a preselection process is performed to eliminate nodes that do not fulfill the user's task. Specifically, it includes the following steps.
(1) For each subtask of a task, based on the current resource state information of the distributed system and the resource requirement information of each subtask, preselected node sets for the subtask may be respectively determined. Nodes outside the pre-selected node-set of the sub-task are of course unable to support the execution of the sub-task. Here, a task may be called a job, and a subtask may be called a task. The preselected node-sets for different subtasks may be the same or different.

Optionally, in some examples, determine the preselected nodes for the next subtask from the set of preselected nodes for the previous subtask, where optionally the order of the subtasks is determined based on the priority of the subtasks. For example, assuming a distributed system includes nodes 1 through 5, and a task includes subtask 1, subtask 2, and subtask 3, determine the preselected nodes for subtask 1, e.g., node 1, node 2, and subtask 3. It is node 3. Then, determine the preselected nodes for subtask 2 from node 1, node 2, and node 3, eg, node 2 and node 3. Next, determine the preselected node for subtask 3 from nodes 2 and 3, for example node 3. To improve the efficiency of determining a preselected node set and the efficiency of task assignment in the case of a task requiring a certain amount of communication, especially a communication-intensive task, by such a method of determining a preselected node set. can be done.

(2) verifying each preselected node of the preselected node-set to determine if the total resources on that preselected node satisfy the total resource requirements of all subtasks assigned to that preselected node; If yes, perform the optimal selection process, otherwise perform the pre-selection process again.

This step is optional and verification of preselected nodes can improve the overall reliability of task assignment. Before performing the pre-selection process, the node set may be split, i.e. each node may be split into groups, and the basis for group splitting may be the cluster to which the node belongs, i.e. nodes on different clusters. into different groups. With node-set partitioning, different tasks are assigned to specific groups of nodes.

A best fit selection process is then performed to select the best target node for the user's task. Specifically, the following steps may be included.
(1) from the information provided by the user and the past experience of the scheduler in the user's task, the computational complexity and communication complexity of the task training process are obtained, and the user's task is computationally intensive based on the computational complexity and communication complexity; or communication-intensive type.

Optionally, in some examples, task type determination may be performed during or prior to the preselection process described above, determining a preselected node-set for subtasks based on task type.

Optionally, in some examples, the optimal selection process determines target nodes for multiple subtasks based on task type. If it is a communication-intensive task, place (assign) the task on the minimum number of physical nodes (i.e. target nodes), specifically, the quantity of physical nodes assigned depends on the available resources of the nodes and the task. If the above placement requirements are not met, all of the multiple subtasks of the task are subject to delayed allocation. If it is a computationally intensive task, relax the above placement constraint to allocate if a node with sufficient resources to satisfy the task request can be selected without meeting the fewest physical nodes, otherwise lazy allocation.

(2) Scoring and sorting the cluster nodes from large to small according to the surplus resources to obtain the score of the mapping from the set of preselected nodes to the task, and the score of the preselected nodes is the node of the cluster of the distributed system. It may be determined based on load and node resource fragmentation conditions. The node load is the total amount of available resources and the amount of resource usage at each node of the cluster at present. A resource fragmentation situation is the distribution of resources on a node.

(3) For all the nodes selected in the above step, sequentially place the tasks of the job in descending order of score (from high score to low score) until they cannot be placed or all tasks in the job are placed. do.

(4) Check the state of the task. If a job has extra tasks that are not placed, it indicates insufficient cluster resources and failed allocation. Release the resource occupied by the task and perform delayed allocation. If there are no surplus tasks in the job (job) and the task type is communication-intensive, it is determined whether the minimum number of nodes is satisfied. If satisfied, the allocation succeeds; if not satisfied, the allocation fails, frees the resource and performs a delayed allocation.

The methods of embodiments of the present disclosure are applied to task scheduling systems or cloud platforms, and the task scheduling systems may be implemented based on platforms such as Kubernetes.

As can be appreciated by those skilled in the art, in specific embodiments of the method, although the steps are described in a particular order, execution in that order is not essential and is not limiting, and each step is determined by its function and internal logic.

As shown in FIG. 7, in an embodiment of the present disclosure,
an obtaining module 701 for obtaining task information, the task information including at least one of a task type of the task and resource requirement information of the task;
A task scheduling apparatus is further provided comprising a first allocation module 702 for allocating respective target nodes to a plurality of subtasks included in the task based on the task information.

In some embodiments, the resource requirement information of the task is resource requirement information of a plurality of subtasks included in the task, and the first allocation module 702 further determines that the currently available resources of the distributed system are the task is used to allocate corresponding target nodes to a plurality of sub-tasks included in the task, respectively, if the total resource corresponding to the resource requirement information of the plurality of sub-tasks in the task is satisfied.

In some embodiments, the apparatus comprises a first determining module for determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource request information of the plurality of subtasks. Further comprising, the first assignment module for assigning a respective target node to a plurality of subtasks included in the task if successfully determining a target node corresponding to each subtask of the plurality of subtasks. used for

In some embodiments, the apparatus comprises a first determining module for determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource request information of the plurality of subtasks. Further comprising, the first assignment module for assigning a respective target node to a plurality of subtasks included in the task if successfully determining a target node corresponding to each subtask of the plurality of subtasks. used for

In some embodiments, the first determining module sequentially selects target nodes corresponding to a plurality of subtasks according to a specific order based on current resource state information of the distributed system and resource request information of the plurality of subtasks. used to determine

Optionally in one example, the resource requirement information of the plurality of sub-tasks comprises resource requirement information of each sub-task of the plurality of sub-tasks, wherein the resource requirement information comprises a requested resource type and quantity of each type of resource, or It also contains other information.

Optionally in one example, the current resource state information of the distributed system includes current state information of a plurality of nodes of the distributed system, wherein the current state information includes whether resources are available, the type and quantity of resources available, It may be used to indicate at least one of load status, topological connection information, or may contain other information as well.

In some embodiments, the apparatus updates current resource state information of the distributed system each time a target node to assign to a subtask is determined, and responds to the next subtask based on the updated current resource state information. further comprising an update module for determining target nodes to be updated.

In some embodiments, the order of sequentially determining target nodes of the plurality of subtasks is derived from at least one of priority of subtasks and dependencies between subtasks.

In some embodiments, the first determining module is configured to sequentially determine target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource request information of the plurality of subtasks. a determining unit; and an updating unit for updating current resource state information of the distributed system based on the resource requirement information of the one subtask each time the target node of the one subtask is determined.

In some embodiments, if the apparatus fails to successfully determine a target node corresponding to at least one subtask of the plurality of subtasks, the currently available resources of the distributed system are reduced to the plurality of subtasks. Further comprising a second decision module for determining that the resource requirements of each subtask of the subtasks are not met.

In some embodiments, the first determining unit determines a preselected node set for each subtask of the plurality of subtasks based on current resource state information of the distributed system and resource request information of the plurality of subtasks. and a selection subunit for selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks.

The preselected node-sets for different subtasks may be the same or different. In some examples, multiple subtasks have the same preselected node-set.

In some embodiments, the determining subunit is used to determine a preselected node set for each subtask based on current resource status information of the distributed system and resource requirement information for each subtask.

The determination of preselected node-sets for multiple sub-tasks may be independent and independent of each other, eg, executed in parallel or in any order. For example, the preselection node-sets of multiple subtasks are determined based on the same distributed system's current resource state information, ie, the distributed system's current resource state information is not updated during the preselection process. For example, the determination of each subtask's preselected node set is only related to its own resource requirement information and is independent of the resource requirement information of other subtasks.

In some embodiments, the determining subunit is used to sequentially determine a preselected node set for each subtask of the plurality of subtasks according to a specific order, provided that the preselected node set for a first subtask of the plurality of subtasks is determined sequentially. A pre-selected node of a selected node-set is selected from a pre-selected node-set of a second sub-task of said plurality of sub-tasks, wherein said second sub-task precedes said first sub-task in rank.

In some embodiments, sequentially selecting a target node of each of said sub-tasks from a pre-selected node of each sub-task of said plurality of sub-tasks according to a specific order, provided that a target node of a second sub-task of said plurality of sub-tasks is selected from said plurality of sub-tasks. is preferentially selected as the target node of the first subtask of the subtask of , but the rank of said second subtask is positioned before said first subtask.

In some embodiments, the selection subunit determines a target node of the first subtask from a preselected node set of a first subtask of the plurality of subtasks based on current resource state information of the distributed system; updating current resource state information of the distributed system based on the resource request information of the first subtask; and preselecting a second subtask of the plurality of subtasks based on the updated current resource state information of the distributed system. It is used to determine the target node of the second subtask from the node set.

Optionally in one example, determining a target node of the first subtask from the preselected node-set of the first subtask based on current state information of each preselected node of the preselected nodeset of the first subtask; Update the current state information of the target node of the first subtask according to the resource requirement information of the subtask.

In some embodiments, the selection subunit determines a score of preselected nodes included in each subtask's preselected node-set based on the task type of the task; is used to select a target node for each subtask from a set of preselected nodes for each subtask based on the score of the preselected nodes.

Optionally, in one example, determining a scoring scheme for each preselected node of the preselected node-set based on the task type of the task.

In some embodiments, the apparatus further includes a third determining module for determining a task type of the task, wherein target nodes corresponding to the plurality of subtasks are determined based on the task type of the task. be done.

In some embodiments, the first allocation module is a group dividing unit for dividing a plurality of subtasks of the task into at least one group, each group dividing at least one subtask of the plurality of subtasks. and an allocation unit for allocating each group of the at least one group with a respective target node, wherein each subtask of the same group is allocated to the same target node.

In some embodiments, the group dividing unit is used to divide a plurality of subtasks of the task into at least one group based on a task type of the task.

In some embodiments, the task type of the task is computationally intensive or communication intensive.

In some embodiments, when the task type of the task is a communication-intensive type, the grouping unit, based on resource request information of the plurality of subtasks of the task, A quantity is determined and used to divide a plurality of subtasks of the task into at least one group based on the total node quantity.

In some embodiments, each subtask of the plurality of subtasks is a group if the task type of the task is computationally intensive.

In some embodiments, determining a respective preselected node-set for each group of said at least one group, and sequentially selecting target nodes for said group from the preselected node-set for each group of said at least one group.

In some embodiments, the first assignment module determines target nodes for multiple subtasks of the task based on a task type of the task, and assigns corresponding target nodes for each of the multiple subtasks of the task. It is also used to assign target nodes to sub-tasks included in said task if the determination is successful.

In some embodiments, the apparatus performs delayed assignment for all of a plurality of subtasks of the task if it fails to successfully determine a target node corresponding to at least one subtask of the subtask. further includes a delay module for

In some embodiments, the delay module delays the execution of multiple subtasks of the task when the currently available resources of the distributed system do not meet the total resources corresponding to the resource requirement information of the multiple subtasks of the task. Lazy allocation for all.

In some embodiments, the apparatus further includes a scheduling module for synchronously scheduling a plurality of subtasks of the task after allocating respective target nodes corresponding to the plurality of subtasks included in the task. .

In some embodiments, the apparatus is applied in a task scheduling system.

In some embodiments, the functions or modules provided by the apparatus according to the embodiments of the present disclosure can be used to perform the methods described in the above method embodiments, the implementation details of which are described in the above method embodiments. , and the description is omitted here for the sake of simplicity.

The apparatus embodiments described above are exemplary, and the modules described as separate components may or may not be physically separate, and the components shown as modules may be physically separate. modules may or may not be static, i.e. co-located, or distributed over multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the objectives of the technical solution herein. A person skilled in the art can understand and implement this without any novelty.

Apparatus embodiments herein may be used, for example, in computing devices such as servers or terminal equipment. Apparatus embodiments can be implemented in software, and can be implemented in hardware or a combination of software and hardware. Taking a software implementation as an example, a logically valid device reads into memory corresponding computer program instructions in non-volatile memory by the processor processing the file in which it resides, and causes the processor to read them from memory and act on them. It consists of From the hardware level, as shown in FIG. 8, which is a hardware structural diagram of a computer device including the apparatus of the present specification, processor 801, memory 802, network interface 803, and non-volatile memory 804 shown in FIG. In addition, the server or electronic device where the device of the embodiment is located may generally include other hardware, depending on the actual function of the computer device, and will not be described here.

Further, embodiments of the present disclosure further provide a computer storage medium storing a computer program, which implements the method of any embodiment when the program is executed by a processor.

And, in embodiments of the present disclosure, a computer comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, wherein the processor executes the method of any embodiment when the program is executed. A device is further provided.

The present disclosure can be a computer program product embodied in one or more storage media (including but not limited to disk storage devices, CD-ROMs, optical storage devices, etc.) containing program code. Computer-readable storage media includes permanent and temporary, removable and non-removable media that can be used to store information in any manner or technology. Information may be computer readable instructions, data structures, program modules, or other data. Examples of computer storage media are 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 technology, optical disc (CD-ROM), digital versatile disc (DVD) or other optical storage device, magnetic cassette, magnetic Including, but not limited to, tape storage or other magnetic storage device, or any other non-transmission medium, used to store information accessible by a computing device.

Other embodiments of the present disclosure will readily occur to those skilled in the art upon reading and practicing the specification disclosed herein. The present disclosure includes any variations, uses, or adaptive changes thereof, and these variations, uses, or adaptive changes are consistent with the spirit of the present disclosure, and technical fields not disclosed in the present disclosure. well-known common sense or ordinary technical means. It is intended that the specification and examples be considered exemplary, with the scope and spirit of the disclosure being limited by the claims.

It should be noted that this disclosure is not limited to the precise structures described above and shown in the drawings, and various modifications and changes may be made without departing from its scope. It is the claims that define the scope of the disclosure.

The above are merely preferred embodiments of the present disclosure and are not intended to be limiting of the present disclosure. Any amendment, equivalent replacement, improvement, etc. without departing from the spirit of the present disclosure shall fall within the protection scope of the present disclosure.

Claims (23)

obtaining task information, the task information including at least one of a task type of the task and resource requirement information of the task;
A task scheduling method, comprising: allocating respective target nodes corresponding to a plurality of subtasks included in the task based on the task information. the resource request information of the task is resource request information of a plurality of subtasks included in the task;
assigning target nodes corresponding to each of a plurality of subtasks included in the task based on the task information;
assigning respective target nodes to a plurality of subtasks included in the task when the currently available resources of the distributed system satisfy the total resources corresponding to the resource requirement information of the plurality of subtasks of the task. 2. The method of claim 1, wherein: determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks;
assigning respective target nodes to a plurality of subtasks included in the task when the currently available resources of the distributed system satisfy the total resources corresponding to the resource requirement information of the plurality of subtasks of the task; ,
3. The step of allocating respective target nodes to a plurality of subtasks included in the task when determining a target node corresponding to each subtask of the plurality of subtasks is successful. The method described in . determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks;
sequentially determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource request information of the plurality of subtasks;
updating current resource state information of the distributed system based on the resource requirement information of the one subtask each time a target node of a subtask is determined. the method of. If the target node corresponding to at least one subtask of the plurality of subtasks is not successfully determined, the currently available resources of the distributed system do not meet the resource requirements of each subtask of the plurality of subtasks. 5. A method according to claim 3 or 4, characterized in that it determines that determining target nodes for the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks;
determining a preselected node-set for each subtask of the plurality of subtasks based on current resource state information of the distributed system and resource requirement information of the plurality of subtasks;
selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks. determining a preselected node-set for each subtask of the plurality of subtasks,
sequentially determining a preselected node set of each subtask of the plurality of subtasks according to a specific order, wherein the preselected node set of the first subtask of the plurality of subtasks is the preselected node of the second subtask of the plurality of subtasks; 7. The method of claim 6, comprising the step of selecting from a preselected node set of subtasks, wherein the order of the second subtask precedes the first subtask. selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks,
sequentially selecting a target node of each subtask from a preselected node set of each subtask of the plurality of subtasks according to a specific order, wherein a target node of a second subtask of the plurality of subtasks is selected as a target node of a second subtask of the plurality of subtasks; 8. A method according to claim 6 or 7, comprising the step of preferentially selecting as a target node of one subtask, the order of said second subtask being positioned before said first subtask. selecting a target node for each subtask from a preselected node set for each subtask of the plurality of subtasks,
determining scores for preselected nodes in a preselected node set for each subtask based on the task type of the task;
selecting a target node for each sub-task from the pre-selected node set for each sub-task based on scores of the pre-selected nodes included in the pre-selected node set for each sub-task. 9. The method of any one of 6-8. 10. The order of sequentially determining target nodes for the plurality of subtasks is determined based on at least one of priority of the subtasks and dependency between the subtasks. 1. The method according to item 1. assigning target nodes corresponding to each of a plurality of subtasks included in the task based on the task information;
determining a task type for the task;
11. A method according to any one of claims 1 to 10, comprising assigning respective target nodes to a plurality of sub-tasks included in said task based on the task type of said task. . assigning respective target nodes to subtasks of a plurality of said tasks,
dividing a plurality of subtasks of said task into at least one group, each group comprising at least one subtask of said plurality of subtasks;
assigning each group of the at least one group a respective target node, wherein each sub-task of the same group is assigned to the same target node. A method according to any one of paragraphs. dividing a plurality of subtasks of the task into at least one group;
13. The method of claim 12, comprising dividing a plurality of subtasks of the task into at least one group based on a task type of the task. 14. A method as claimed in any one of claims 11 to 13, wherein the task type of the task is computationally intensive or communication intensive. dividing a plurality of subtasks of the task into at least one group;
when the task type of the task is communication-intensive, determining the total number of nodes required by the plurality of subtasks based on the resource request information of the plurality of subtasks of the task; and based on the total number of nodes, dividing a plurality of subtasks of said task into at least one group;
and/or
15. A method according to any one of claims 12 to 14, comprising grouping each subtask of said plurality of subtasks if the task type of said task is computationally intensive. . assigning respective target nodes to a plurality of subtasks included in the task based on a task type of the task,
determining target nodes for a plurality of subtasks of the task based on a task type of the task;
and assigning a target node to a plurality of subtasks included in the task upon successfully determining a corresponding target node for each of the plurality of subtasks of the task. 16. The method of any one of 15. further comprising performing delayed assignment for all of a plurality of subtasks included in the task if a corresponding target node for at least one subtask of the subtask is not successfully determined. 17. The method of claim 16. performing delayed allocation for all of the plurality of subtasks included in the task when the currently available resources of the distributed system do not meet the total resources corresponding to the resource requirement information of the plurality of subtasks of the task. 18. The method of any one of claims 2-17, further comprising: 19. The method of any one of claims 1 to 18, further comprising synchronously scheduling a plurality of subtasks of the task after assigning respective target nodes to the plurality of subtasks included in the task. The method described in section. 20. A method according to any preceding claim, applied to a task scheduling system. an acquisition module for acquiring task information, wherein the task information includes at least one of a task type of the task and resource requirement information of the task;
A task scheduling device, comprising: a first allocation module for allocating respective target nodes to a plurality of subtasks included in the task based on the task information. 21. A computer readable storage medium having stored thereon a computer program, the method according to any one of claims 1 to 20 being implemented when said program is executed by a processor. medium. A computer device comprising a memory, a processor, and a computer program stored in said memory and operable on said processor, the method of any one of claims 1 to 20 when said processor executes said program. A computing device for implementing

Images