KR20220002547A - Task Scheduling Method and Apparatus - Google Patents

Abstract

An embodiment of the present invention provides a task scheduling method and apparatus, comprising: acquiring task information including at least one of a task type and resource demand information of the task; A target node may be assigned to each of a plurality of sub-tasks included in the task based on the information on the task.

Description

Task Scheduling Method and Apparatus

[Cross-reference to related applications]

The present invention is a Chinese patent application with the application number 202010165543.4 filed with the Chinese Patent Office on March 11, 2020 and the invention is titled "Task Scheduling Method and Apparatus", and an application filed with the Chinese Patent Office on March 11, 2020 Priority is claimed on the Chinese patent application numbered 202010165763.7 and entitled "Task Scheduling Method and Apparatus", the entire contents of which are incorporated herein by way of reference.

FIELD OF THE INVENTION The present invention relates to the field of distributed systems, and more particularly to task scheduling in distributed systems.

Currently, more and more tasks use distributed systems to perform resource allocation calculations, where many tasks often include multiple subtasks, and the content and required resources of each subtask may be completely the same or different. The task scheduling method in the traditional distributed system generally allocates resource nodes in units of sub-tasks, so both resource utilization and task execution efficiency for the cluster are relatively low.

The present invention provides a task scheduling method.

According to a first aspect of an embodiment of the present invention, there is provided a task scheduling method, the method comprising: obtaining information on a task including at least one of a task type and resource demand information of the task; and allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task.

In some embodiments, the resource demand information of the task is resource demand information of a plurality of sub-tasks included in the task; The method includes: allocating a target node corresponding to each of the plurality of subtasks of the task when the currently available resources in the distributed system satisfy total resources corresponding to resource demand information of the plurality of subtasks of the task; include more

In some embodiments, the method further comprises: determining a target node for a plurality of sub-tasks of the task based on resource demand information of the plurality of sub-tasks; When the currently available resources in the distributed system satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task, allocating a target node corresponding to each of the plurality of subtasks of the task includes: and allocating a target node corresponding to each of a plurality of subtasks of the task when a corresponding target node is successfully determined for each subtask of the subtasks.

In some embodiments, the target node corresponding to the plurality of sub-tasks is sequentially determined according to a specific order based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks.

In an optional example, the resource demand information of the plurality of subtasks includes resource demand information of each subtask among the plurality of subtasks, wherein the resource demand information includes a required resource type and a quantity of each type of resource or may further include other information.

In an optional example, the current resource status information of the distributed system may include current status information of a plurality of nodes in the distributed system, wherein the current status information includes availability of resources, types and quantities of available resources, and loads It indicates at least one of context and topological connection information, or further includes other information.

In some embodiments, after determining the target node for one subtask each time, 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. .

In some embodiments, when a target node for at least one subtask among the plurality of subtasks is not successfully determined, the currently available resources of the distributed system are the resources required by each subtask among the plurality of subtasks. decides that it is not satisfied.

In some embodiments, the sequential determination order of the target nodes of the plurality of sub-tasks is determined based on at least one of a priority of the sub-tasks and a dependency relationship between the sub-tasks.

In some embodiments, the determining of a target node for the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks comprises: the current resource state information of the distributed system and determining a preselection node set of each subtask among the plurality of subtasks based on resource demand information of the plurality of subtasks; and selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks.

Here, the preselection node sets of different subtasks may be the same or different. In some examples, a plurality of sub-tasks may have the same set of pre-selection nodes.

In some embodiments, the determining of the pre-selection node set of each sub-task among the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks comprises: and determining a pre-selection node set of each sub-task based on current resource state information and resource demand information of each sub-task.

Determination of the pre-selection node set of the plurality of sub-tasks is independent and may not have a mutual dependency, for example, may be executed in parallel or may be executed in an arbitrary precedence order. For example, the preselection node set of the plurality of subtasks is determined based on the current resource state information of the same distributed system, that is, the current resource state information of the distributed system may not be updated during the preselection process. For example, the determination of the pre-selection node set of each sub-task is only related to its own resource demand information and is independent of the resource demand information of other sub-tasks.

In some embodiments, the determining of the preselection node set of each subtask among the plurality of subtasks comprises: determining the preselection node set of each subtask of the plurality of subtasks sequentially according to a specific order; include

In one example, a preselection node of a set of preselection nodes of a first subtask of the plurality of subtasks is selected from a set of preselection nodes of a second subtask of the plurality of subtasks, wherein the second subtask The order of is placed before the first sub-task.

In another example, a preselection node set of each subtask among the plurality of subtasks is determined based on resource demand information of the plurality of subtasks.

In some embodiments, the step of selecting the target node of each sub-task from the set of pre-selection nodes of each sub-task among the plurality of sub-tasks includes the dictionary of each sub-task among the plurality of sub-tasks in order according to a specific order. A target node of each subtask is selected from the selection node.

In an optional example, first, a target node of a second subtask among the plurality of subtasks is selected to be a target node of a first subtask among the plurality of subtasks, wherein the order of the second subtask is the second subtask 1 It is located in front of the sub task.

In some embodiments, the selecting of the target node of each sub-task from the pre-selection node set of each sub-task among the plurality of sub-tasks comprises: selecting one of the plurality of sub-tasks based on current resource state information of the distributed system. determining a target node of the first sub-task from a set of pre-selected nodes of the first sub-task; Update the current resource state information of the distributed system according to the resource demand information of the first sub-task, and according to the updated current resource state information of the distributed system, a dictionary of a second sub-task of the plurality of sub-tasks and determining a target node of the second sub-task from the selected node set.

In an optional example, determining a target node of the first subtask in the preselection node set of the first subtask based on current state information of each preselection node in the preselection node set of the first subtask, and The current state information of the target node of the first sub-task is updated based on the required resource information of the first sub-task.

In some embodiments, the step of selecting the target node of each subtask from the preselection node set of each subtask among the plurality of subtasks includes adding the preselection node set of each subtask according to the task type of the task. determining a score of included pre-selection nodes; and selecting a target node of each subtask from the preselection node set of each subtask based on the scores of the preselection nodes included in the preselection node set of each subtask.

In an optional example, a scoring strategy of each pre-selection node in the set of pre-selection nodes is determined based on the task type of the task.

In some embodiments, allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task may include: determining a task type of the task; and allocating a target node corresponding to each of a plurality of sub-tasks of the task based on the task type of the task.

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

In some embodiments, allocating a target node corresponding to each of a plurality of subtasks of the task comprises dividing the plurality of subtasks of the task into at least one group - for each group at least one of the plurality of subtasks Contains one sub-task -; and allocating a corresponding target node to each group in the at least one group, wherein the same target node is allocated to each subtask in the same group.

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, the dividing the plurality of subtasks of the task into at least one group includes: when the task type of the task is communication-intensive, based on resource demand information of the plurality of subtasks of the task determining a total node quantity required for the subtask, and dividing the plurality of subtasks of the task into at least one group based on the total node quantity.

In some embodiments, the dividing the plurality of subtasks into at least one group includes: when the task type of the task is a computationally intensive type, grouping each subtask among the plurality of subtasks into one group include

In some embodiments, allocating a corresponding target node for each group in the at least one group comprises determining a set of preselected nodes for each group in the at least one group, and in the order the at least one group and selecting a target node of the group from a set of pre-selection nodes of each group in the group.

In some embodiments, allocating a target node corresponding to each of the plurality of sub-tasks of the task based on the task type of the task comprises: a target for the plurality of sub-tasks of the task based on the task type of the task. determining a node; and allocating target nodes for the plurality of subtasks included in the task when all corresponding target nodes for the plurality of subtasks of the task are successfully determined.

In some embodiments, the method includes executing delay allocation for all of a plurality of subtasks of the task when a corresponding target node for at least one of the subtasks is not successfully determined. include more

In some embodiments, the method includes: when the currently available resources in the distributed system do not satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task, all of the plurality of subtasks included in the task The method further includes performing lazy allocation.

In some embodiments, the method further comprises allocating a target node corresponding to each of a plurality of subtasks of the task, and then executing synchronous scheduling on the plurality of subtasks of the task.

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

According to a second aspect of an embodiment of the present invention, there is provided an apparatus for scheduling a task, the apparatus comprising: an acquiring module for acquiring information of a task including at least one of a task type and resource demand information of the task; and a first assignment module for allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task.

According to a third aspect of the embodiment of the present invention, there is provided a computer-readable storage medium in which a computer program is stored. When the program is executed by a processor, the method according to any one embodiment is implemented.

According to a fourth aspect of the embodiment of the present invention, there is provided a computer device including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor performs an operation in any one embodiment when the program is executed. implement the following method.

In an embodiment of the present invention, information of a task including at least one of a task type and resource request information of the task is obtained, wherein the resource request information of the task is resource demand information of a plurality of sub-tasks included in the task . Allocation of the target node may be performed based on the task type or resource demand information of the task. For example, when the currently available resources in the distributed system satisfy the total resources corresponding to resource demand information of a plurality of subtasks of the task, a target node corresponding to each of the plurality of subtasks of the task is allocated. In this way, the resource demand situation of the whole task can be obtained, and the resource allocation is executed at the granularity of the whole task, and the resource utilization rate of the cluster is improved.

It should be understood that the above general description and detailed description in the back are merely illustrative and interpretative and not intended to limit the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The drawings herein are attached to and form a part of the present specification, illustrate embodiments consistent with the present invention, and together with the specification interpret the technical solutions of the present invention.
1 is a schematic diagram of a distributed system according to an embodiment of the present invention.
2 is a flowchart of a task scheduling method according to an embodiment of the present invention.
3A is a schematic diagram of resource change during a node mock allocation process according to an embodiment of the present invention.
3B is a flowchart of a simulated node allocation process according to an embodiment of the present invention.
4 is a schematic diagram of a traditional task scheduling process.
5 is a schematic diagram of a task scheduling process according to an embodiment of the present invention.
6 is a schematic diagram of scheduling logic according to an embodiment of the present invention.
7 is a structural schematic diagram of a task scheduling apparatus according to an embodiment of the present invention.
8 is a structural schematic diagram of a computer device of an apparatus according to an embodiment of the present invention.

Exemplary embodiments are described in detail herein and are exemplarily shown in the drawings. When the following description relates to drawings, like numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The embodiments described in the illustrative examples below are not representative of all embodiments consistent with the present invention. To the contrary, these are merely examples of apparatus and methods consistent with some aspects of the invention detailed in the appended claims.

The terminology used in the present invention is only for describing specific embodiments and not for limiting the present invention. It is said that odd-numbered forms of “one,” “the,” and “the same” used in the present invention and the appended claims also include plural forms unless otherwise clearly indicated in the preceding sentence. It should be understood that the term “and/or” includes any or all possible combinations of one or more associated enumerated items.

Although the present invention may use terms such as "first", "second", "third" and the like to describe various pieces of information, it should be understood that such information should not be limited to these terms. These terms are merely intended to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may be referred to as second information, and similarly, the second information may be referred to as first information. This is determined by the language environment, for example, the word "if" used here is "... It can be construed as “when” or “when…” or “in response to a decision.”

Currently, more and more tasks are starting to be processed using distributed systems, and these tasks are called distributed tasks. These distributed tasks include an acute demand for high-performance processing resources such as Graphics Processing Units (GPUs), Digital Signal Processors (DSPs), Field Programmable Gate Arrays (FPGAs), etc. A bar, for example, is a deep learning task. As shown in FIG. 1 , this is a distributed system in some embodiments, wherein the distributed system includes one or a plurality of clusters, each cluster includes one or a plurality of servers, and each server includes one node (in the drawing). each node), and each node contains at least one of resources such as a CPU (Central Processing Unit), GPU, memory, disk, and network port. After allocating a subtask to a node, the subtask can be executed using the node's resources, and each node can execute one or more subtasks. The subtasks executed by each node in the same cluster may be the same or different. In a distributed system, both the task itself and the demand for the resource of the task are completely different from the traditional task. Therefore, the task scheduling method in the distributed system is adjusted accordingly to improve the suitability of the use of cluster resources in the distributed system, and furthermore, the task It improves execution performance and resource utilization of the cluster.

When allocating resources for tasks in a distributed system, different allocation methods affect the execution efficiency of tasks. For example, in some cases, since a plurality of subtasks in one task need to perform frequent communication with each other, the plurality of subtasks in the task are suitable to be assigned to a node having a low communication cost; In some other cases, since a plurality of subtasks in one task generate a large amount of computation during execution, the plurality of subtasks in the task are suitable to be allocated to a node having many computational resources.

And, most of the tasks in the current distributed system have the following characteristics: One task often includes a plurality of sub-tasks, and the content and required resources of each sub-task may be the same or different, and between sub-tasks may need to perform frequent communication. Also, in some cases, a plurality of sub-tasks in one task must be scheduled and executed together, otherwise the task is not normally executed; In some other cases, the execution order of each subtask has a certain effect on the execution status of the entire task. For example, a slave subtask can be executed only after the master subtask to which it depends is completed. have.

Currently, some schedulers (eg, kube-batch) batch schedule tasks in a distributed system according to the following method: The scheduler attempts to schedule one for each sub-task of one task in order, and each After completing the scheduling attempt for the sub-task, it is determined whether the entire task meets the user-expected executable state. This scheduling strategy is called gang scheduling.

However, the current implementation of gang scheduling essentially schedules subtasks at a granular level, and if the scheduler does not obtain the resource demand of the entire task, the node allocated for the subtask may not be the best choice. Therefore, both task execution efficiency and resource utilization for the cluster may be relatively low. Therefore, when cluster resources are tight or task types are complex, a large number of scheduling failures occur.

Based on this, an embodiment of the present invention provides a task scheduling method. As shown in FIG. 2 , the method may include steps 201 and 202 .

Step 201: Obtain task information including at least one of a task type and resource demand information of the task.

Step 202: Allocate a target node for each of the plurality of sub-tasks of the task according to the information of the task.

The method according to an embodiment of the present invention may be executed by any electronic device such as a terminal device or a cloud server. In some embodiments, the method is executed by a processor or a scheduler, wherein, optionally, the processor Alternatively, the scheduler may be executed in a cloud platform such as Kubernetes, and specifically, may be installed in a container arranging engine, but the embodiment of the present invention is not limited thereto. In step 201, information on one or a plurality of tasks may be acquired.

The plurality of sub-tasks included in the task may be all sub-tasks of the task, and some sub-tasks of the tasks, for example, a sub-task with a large resource demand, a sub-task with a high priority, or a sub-task having a dependency relationship with each other , a subtask with a relatively close resource demand type or quantity, or the like. Here, the sub-tasks having a mutual dependency relationship refer to a plurality of sub-tasks having a mutual dependency relationship or a master-slave relationship. For example, the execution of the sub task A can be started only after the execution of the sub task B is completed, and For example, it may be a plurality of sub-tasks having the same type or at least one specific type. The task may be a training task of a neural network, an estimation task, or other types of deep learning tasks. In some alternative embodiments, each sub-task of the same task may be executed by the same or different nodes in the same cluster in the distributed system. When a distributed system includes a plurality of clusters, each sub-task of the same task can be executed by nodes in the same or different clusters. Optionally, a task may be assigned to a cluster first, and then target nodes used for a plurality of subtasks of the task may be determined in the cluster again. In this way, by allocating a plurality of sub-tasks of the task to the same cluster, it is advantageous to reduce communication overhead and improve the execution efficiency of the task.

The information of the task may be obtained based on a user input or setting, or obtained based on analysis of the task. The information on the task may include task type and/or resource demand information of the task. Here, the resource demand information of the task is specifically, resource demand information of a plurality of sub-tasks included in the task. In some embodiments, the information of the task in addition to the task type and resource demand information includes priority information of the task, priority information of a plurality of sub-tasks included in the task, dependency information between sub-tasks, and information provided by a user. , may further include one or more of past information, calculation amount information, communication amount information, and the like.

Here, the resource demand information of the sub-task may include a resource type and/or resource quantity required to execute the sub-task, and the resource type is CPU, GPU, DSP, FPGA, memory, disk, and network port. It may include at least one of resources, such as, but is not limited thereto.

The priority information of the task may include, but is not limited to, a length of time waiting for the task to be scheduled and/or a total resource required for the plurality of sub-tasks among the tasks. The length of time waiting to be scheduled indicates the length of time from the time when the scheduler receives a task to the time the task is scheduled. Priority is given to tasks with relatively long waiting to be scheduled and tasks with relatively large required total resources. By setting the priority high, it is possible to prevent a task from being placed in a state waiting for scheduling for a long time.

Priority information of a plurality of sub-tasks included in the task is used to determine priorities of a plurality of sub-tasks included in the task. In some embodiments, in one scheduling cycle or resource allocation process, scheduling of sub-tasks may be executed based on at least priority information of each sub-task.

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

The information provided by the user is used to determine the amount of computation and communication of a plurality of sub-tasks of the task.

The past information may include past scheduling information and/or past execution status. A task type of the task may be determined during a past scheduling process based on the past information.

In addition to the above information, the information of the task may further include other information according to an actual application scenario, which is not further described herein.

In some embodiments, in step 202, allocating a target node to each of the plurality of subtasks included in the task based on the information of the task includes: currently available resources in the distributed system of the plurality of subtasks of the task. and allocating a target node corresponding to each of a plurality of sub-tasks of the task when the total resource corresponding to the resource demand information is satisfied. The total resource corresponding to the resource demand information of the plurality of subtasks of the task may be a total resource required for the plurality of subtasks of the task. Optionally, when the plurality of subtasks are all subtasks of the task, the The total resource may be a total resource required for the task. Optionally, it indicates that currently available resources in the distributed system can support execution of the plurality of subtasks if the currently available resources of the distributed system satisfy the plurality of subtasks or the total resources required for the tasks. Optionally, if the currently available resources of the distributed system do not satisfy the plurality of subtasks or the total resources required for the tasks, the currently available resources of the distributed system support the execution of only some tasks among the plurality of subtasks and delay the execution of all subtasks. Indicates that support is not available.

In an embodiment of the present invention, a target node corresponding to each of the plurality of subtasks of the task may be allocated only when the currently available resources in the distributed system satisfy the total resources required for the plurality of subtasks of the task. In this way, by dividing a node together for a plurality of subtasks included in a task as a unit of resource allocation, when the currently available resources in the distributed system are insufficient, some subtasks of the tasks cannot allocate resources successfully. It is advantageous to improve the rational use of resources by preventing a case where only some subtasks in a task occupy all the resources in the distributed system.

In some embodiments, a target node may be allocated to a plurality of sub-tasks based on current resource state information of the distributed system and resource demand information of the plurality of sub-tasks. Here, target nodes capable of satisfying respective resource demands may be allocated to each sub-task, and target nodes of different sub-tasks may be the same or different.

In some embodiments, a target node may be determined for the plurality of sub-tasks based on current resource state information of the distributed system and resource demand information of the plurality of sub-tasks, and each sub-task among the plurality of sub-tasks may be selected. When a target node corresponding to the target node is successfully determined, a target node corresponding to each of the plurality of sub-tasks of the task may be allocated. The process of determining a target node for the plurality of sub-tasks is a mock allocation process. The mock allocation is a virtual resource allocation process, not an actual resource allocation process. It is used to determine whether a node in the current distributed system can be allocated. By executing the mock assignment of nodes for a plurality of subtasks of the task, it can be determined whether the distributed system currently satisfies the resources required for the plurality of subtasks.

Here, the current resource status information of the distributed system is for displaying currently available resources in the distributed system, and includes at least one of the quantity, type, and distribution of the currently available resources of the distributed system, or the current status of a plurality of nodes in the distributed system. The status may include, for example, availability or existence of available resources, the type and quantity of available resources, and the like. The resource demand information of the plurality of subtasks includes resource demand information of each subtask among the plurality of subtasks, for example, at least one of a required resource type and a resource quantity of any one resource type, for example, a required GPU, DSP, One or a plurality of CPUs may be included, but embodiments of the present invention are not limited thereto.

In some embodiments, based on the resource demand information of the plurality of sub-tasks and the current resource state information of the distributed system, it is determined whether there is a target node satisfying the demand of the plurality of sub-tasks in the distributed system, or according to a specific order The target node of each subtask among the plurality of subtasks can be determined sequentially. For example, if each sub-task among a plurality of sub-tasks finds a target node that satisfies each demand in the distributed system, it can be determined that the currently available resources of the distributed system satisfy the total resources required for the plurality of sub-tasks. do. Also, for example, if at least one sub-task among the plurality of sub-tasks does not find a target node that satisfies its demand in the distributed system, it means that the currently available resources of the distributed system do not satisfy the total resources required for the plurality of sub-tasks. can decide

Unlike the traditional determination of the node of one sub-task, in the embodiment of the present invention, the target nodes of the plurality of sub-tasks are sequentially determined according to a specific order, and the determination of the target node of the subsequent sub-task is determined by the previous sub-task. It can be influenced by the target node determined for In some embodiments, after determining a corresponding target node for one subtask each time, the current resource status information of the distributed system, for example, the determined current status information of the target node is updated, and then the updated current resource Resource conflict is prevented by determining a target node corresponding to the next sub-task based on the state information.

It should be explained that the updating of the current resource state information mentioned here is used in the mock assignment or target node determination process, and is a hypothetical update that is not true. After execution or determination of the target node, it indicates the virtual change status of the currently available resources in the distributed system. Since the actual allocation of the sub-tasks is not executed, the currently available resources in the distributed system do not actually change. For example, after determining the target node of one subtask, the target node is virtually assigned to the subtask, and some resources corresponding to the resources required by the subtask in the target node are virtually allocated to the subtask. As it can be regarded as binding, the resource virtually bound in this way cannot be re-allocated to other subtasks.

The specific order may indicate a preset order. Alternatively, in some examples, the specific order may be determined according to the priorities of the plurality of sub-tasks. For example, first a corresponding target node is determined for a sub-task having a high priority, and then the priority is changed again. A corresponding target node is determined for a lower subtask, wherein the priority may be determined by one or more factors such as a subtask type, a dependency relationship between a subtask and other subtasks, and the like, for example, In two sub-tasks in which a dependency relationship or a master-slave relationship exists, the priority of the master sub-task is higher than that of the slave sub-task, but the embodiment of the present invention is not limited thereto. Alternatively, in some other examples, the specific order may be determined according to a dependency relationship or a master-slave relationship of the sub-tasks. Since the slave sub-task cannot start execution when the master sub-task is lacking, a target node corresponding to the master sub-task is first allocated to improve the execution efficiency of the entire task. Alternatively, the specific order may be determined by other factors, and embodiments of the present invention are not limited thereto.

In some embodiments, when a target node of a resource required for each subtask among the plurality of subtasks exists in the distributed system, the currently available resource in the distributed system is a total resource corresponding to the resource demand information of the plurality of subtasks in the task determines that it satisfies; Otherwise, it is determined that the currently available resources in the distributed system do not satisfy the total resources corresponding to the resource demand information of a plurality of subtasks of the task.

As in the example shown in FIGS. 3A and 3B , if it is assumed that one task includes sub-task 1, sub-task 2, and sub-task 3 in the order of priority from high to low, first, the currently available resources of the distributed system obtain the target node of the sub-task 1 by allocating nodes for the sub-task 1 according to the information and the resource demand information of the sub-task 1, and then update the current resource status information in the distributed system; After being updated again, the target node of sub-task 2 is obtained by mock-allocating nodes for sub-task 2 based on the currently available resource information of the distributed system and the resource demand information of sub-task 2, and then the current resource status information in the distributed system is updated. and; Then, the target node of the sub-task 3 is obtained by mock-allocating the node for the sub-task 3 based on the updated information on the currently available resources of the distributed system and the resource demand information of the sub-task 3.

If all nodes for sub-task 1, sub-task 2 and sub-task 3 are all successfully mock-allocated, the currently available resources of the distributed system are required for the demand of each sub-task among the plurality of sub-tasks, that is, required for the plurality of sub-tasks. It is determined that the total resource is satisfied. If the mock allocation of nodes for sub-task 1, sub-task 2, and sub-task 3 fails, for example, when a target node that satisfies the required resource cannot be found in the distributed system, the current state of the distributed system It is determined that the available resources do not satisfy the total resources required for the plurality of sub-tasks.

When it is determined based on the mock assignment that the currently available resource of the distributed system satisfies the resource required for each subtask among the plurality of subtasks, a target node corresponding to each subtask determined in the mock assignment is assigned to each of the subtasks. It can be assigned to a task.

In some embodiments, a target node may be determined for each sub-task based on current status information of each node in the distributed system, for example, arranging or scoring each node based on the current status information of each node Also, according to a specific order, it is possible to allocate a node with a high arrangement rank or a high score for a plurality of sub-tasks in order.

In some embodiments, when determining a target node for a plurality of subtasks of the task, a target node corresponding to each of the plurality of subtasks may be determined through two processes of preselection and optimization.

Here, in the pre-selection process, based on the resource demand information of the plurality of sub-tasks, for each sub-task among the plurality of sub-tasks, at least one pre-selection node that satisfies its demand in each distributed system is initially selected, or Remove nodes that cannot satisfy their demand in a distributed system. For example, a preselection node set of each subtask among the plurality of subtasks is determined based on current state information of a plurality of nodes in the distributed system and resource demand information of the plurality of subtasks, wherein each subtask The set of preselection nodes of the task includes at least one preselection node. The same or different nodes may be included in the set of preselection nodes of different subtasks.

In the optimization process, a target node of each subtask is selected from a set of preselected nodes of each subtask. For example, one most suitable node from the preselection node set of the subtask may be selected as the target node of the subtask. In some optional examples, the current state information of each preselection node included in the preselection node set of the subtask, for example, the current load situation, the available resource type, the available resource quantity, the available resource distribution, and the topology among the nodes. The target node may be determined based on at least one of a connection relationship and a topological connection relationship with other nodes. The available resource type includes, but is not limited to, at least one of a communication port, a disk storage space, a memory, a CPU, a GPU, and the like. The available resource quantity may be a CPU quantity, a remaining disk capacity, a remaining memory capacity, and the like. The available resource distribution is also referred to as a resource fragmentation situation, that is, the current distribution position of available resources, for example, how many disks have the remaining disk capacity distributed. The topological structure of each node in the distributed system may be, for example, a bus-type topology, a star-type topology, a ring-type topology, and a tree-type topology.

In some embodiments, during the preselection process, the preselection node set of each subtask may be determined based on the current resource state information of the distributed system and the resource demand information of each subtask, that is, a plurality of The pre-selection process of each sub-task in the sub-task does not interfere with each other, and the selection of the pre-selection node set of each sub-task does not depend on the resource demand information of other sub-tasks, but only relates to the resource demand information of the sub-task. Correspondingly, the determination of the set of preselection nodes of the plurality of subtasks may also be executed in parallel or in any precedence order. In this way, the decision of the preselection node can be executed quickly, improving the overall efficiency of the mock assignment.

For example, assuming that a task includes sub-task 1, sub-task 2, and sub-task 3 in the order of priority from high to low, the following operations can be performed in parallel during the pre-selection process: Based on the resource demand information and the current resource status information of the distributed system, the preselection node set of the subtask 1 is determined, and the preselection of the subtask 2 is based on the resource demand information of the subtask 2 and the current resource status information of the distributed system The node set is determined, and the preselection node set of the subtask 3 is determined based on the resource demand information of the subtask 3 and the current resource status information of the distributed system, and the {node1 , node2, node3} are included, the preselection node set corresponding to subtask 2 includes {node2, node5, node6, node7}, and the preselection node set corresponding to subtask3 is Assume that {node6 and node7} gets included.

In some embodiments, during the preselection process, a preselection node set of a plurality of subtasks may be sequentially selected according to a specific order. In an optional example, the pre-selection node set of the subsequent sub-task is determined based on the pre-selection node set of the previous sub-task. For example, the pre-selection node of the next sub-task is selected from the set of pre-selection nodes of the previous sub-task, where the quantity of the preceding sub-task may be one or two or more, for example, The preceding subtask may be all subtasks before the following subtask, and the preceding subtask may be adjacent to the next subtask or may be spaced apart by at least one subtask. Examples are not limited thereto.

In some embodiments, the preselection node set of the plurality of subtasks may be determined by comprehensively considering resource demand information of the plurality of subtasks, for example, a common preselection node set for the plurality of subtasks However, the preselection node included in the common preselection node set may satisfy the resource demand of at least two subtasks among the plurality of subtasks.

In some embodiments, during the optimization process, target nodes of a plurality of sub-tasks may be determined in parallel or a target node of a plurality of sub-tasks may be comprehensively determined in conjunction with a set of candidate nodes of the plurality of sub-tasks. For example, the intersection of candidate node sets of at least some subtasks among a plurality of subtasks is obtained to determine a target node of the at least some subtasks, but the embodiment of the present invention is not limited thereto.

In some embodiments, a target node corresponding to the plurality of subtasks may be sequentially determined from a set of preselected nodes of the plurality of subtasks according to a specific order, for example, based on the priorities of the plurality of subtasks . Here, in an optional example, whenever a target node corresponding to each sub-task is determined, the current state information of the target node is updated. For example, the current state information of the target node is updated based on the resource demand information corresponding to the subtask. In this way, in the process of subsequently determining the target node of the subtask, the portion of the target node corresponding to the resource required by the subtask becomes unavailable, so that it is possible to ensure that a plurality of subtasks are executed smoothly. have. In another optional example, the target node of the subsequent subtask is selected based on the target node selected by the previous subtask. For example, unless the target node is not included in the set of candidate nodes of the subsequent subtask or other important factors of the target node do not satisfy the demand of the subsequent subtask, the subsequent subtask takes precedence The target node of the sub task may be selected as its own target node, and the embodiment of the present invention is not limited thereto.

When resource allocation is performed synchronously for a plurality of sub-tasks, the pre-selection and optimization process may be a simulated allocation process. When resource allocation is asynchronously executed for a plurality of subtasks, the preselection and optimization process may be an actual node allocation process. Similarly, the updating of the current resource state information mentioned in the above process may be an actual update process executed by a change in currently available resources in the distributed system after a node assignment of one sub task is completed.

In an embodiment of the present invention, a target node may be selected from a set of pre-selection nodes based on a certain strategy. For example, the target node may be selected based on current resource status information of the distributed system or current status information of each preselection node and resource demand information of a plurality of subtasks. In some embodiments, it is possible to determine the score of the preselection nodes included in the preselection node set of each subtask, for example, current state information of each preselection node in the preselection node set of the subtask and the subtask The score of each preselection node is determined based on the resource demand information of the task, and each subtask is determined from the preselection node set of each subtask based on the score of the preselection node included in the preselection node set of each subtask. You can select the target node of the task. For example, a preselection node having the highest score among at least one preselection node included in the preselection node set may be selected as the target node. Alternatively, the selection of the target node may be jointly executed based on the score and other factors.

Continuing with the previous example, if target nodes of a plurality of sub-tasks are sequentially determined based on task priorities, the score of each pre-selection node in the set of pre-selection nodes of sub-task 1 can be determined first, If it is assumed that the scores of node 1, node 2, and node 3 are 80, 90, and 70, respectively, node 2 is determined as the target node of sub task 1 and the current state information of node 2 is updated. Thereafter, the score of each pre-selection node in the set of pre-selection nodes of sub-task 2 is determined, wherein the score of node 2 is determined based on the updated current state information of node 2, and if node 2 and node 5 , assuming that the scores of node 6 and node 7 are 60, 80, 75, and 70, respectively, node 5 is determined as the target node of sub task 2 and the current state information of node 5 is updated. Finally, the score of each pre-selection node in the set of pre-selection nodes of sub-task 3 is determined. If it is assumed that the scores of nodes 6 and 7 are 70 and 60, respectively, node 6 is the target node of sub-task 3 to be decided by

In some embodiments, the task type of the task may be determined based on the information of the task obtained in step 201, and target nodes may be allocated for a plurality of sub-tasks based on the task type of the task. Here, the task type of the task may be divided according to the actual situation. In deep learning tasks such as training and reasoning of deep learning models such as neural networks, tasks can be selectively divided into communication-intensive tasks and computation-intensive tasks. Here, the communication-intensive task refers to a task in which communication between sub-tasks is relatively frequent during the task processing process, and the calculation-intensive task refers to a task with a relatively large amount of computation during the task processing process.

Specifically, the task type of the task may be determined based on the overall status of the plurality of sub-tasks of the task. As an optional example, the task type of the task may be determined based on the amount of computation and the amount of communication of a plurality of sub-tasks of the task. For example, when the amount of calculation is greater than a preset calculation amount threshold, the task is recognized as a calculation-intensive task; Also, for example, when the communication amount is greater than a preset communication amount threshold, the task may be recognized as a communication-intensive task. Also, for example, if both the amount of calculation and the amount of communication exceed a corresponding preset threshold, it may be determined that the task is a communication-intensive task. As another optional example, the task type may be determined based on past scheduling experience. For a specific task, if the scheduler determines the task as a computationally intensive task during the past scheduling process, determines the task as a computationally intensive task; If the scheduler determines that the task is a communication-intensive task during the past scheduling process, it determines the task as a communication-intensive task. Optionally, the task type of the task may be determined based on information on at least one past execution of the task, for example, if, during the recent one or several past executions, the calculation of the task is determining the task as a computationally intensive task when the time required or computational resources are large; Also, for example, if the communication of the task occupies a large amount of time or communication resources during one recent or several past execution processes, the task may be determined as a communication-intensive task. Alternatively, the task type of the task may be determined based on other methods. Alternatively, the task type of the task may be determined based on information provided by the user, for example, task type information provided by the user or information on the amount of computation and/or communication amount of the task provided by the user. Embodiments of the present invention are not limited to the specific implementation for determining the task type.

In an embodiment of the present invention, the target node of the plurality of sub-tasks may be determined based on the task type so that the entire task is efficiently executed. For example, in a communication-intensive task, communication overhead can be reduced by allocating a plurality of sub-tasks in the task to target nodes with a small possible number; Also, for example, in a computationally intensive task, the resource of the target node only needs to satisfy the demand of a plurality of subtasks of the task, or a node with good computational performance is preferentially selected, or a fragment of the node Priority use of fragmented resources, or optimization of communication overhead under the premise that available computational resources meet user demand, and the like.

In some optional examples, the preselection and/or optimization process may be executed based on the task type of the task.

In some embodiments, the pre-selection process is executed based on the task type of the task. For example, in a communication-intensive task, a large number of sub-tasks in a plurality of sub-tasks, such as a small amount of available resource types and/or available resources Eliminate nodes that cannot clearly satisfy the task demand, and control, for example, the number of nodes included in the preselection node set within a certain range, wherein nodes having a large number of resource types and/or a large number are prioritized Alternatively, a node whose resource type relatively matches the total demand resources of the plurality of sub-tasks is used as the pre-selection node. In some optional examples, instead of determining a single set of preselection nodes for each subtask, one set of preselection nodes may be determined for some or all subtasks among a plurality of subtasks of the communication-intensive task. The embodiment is not limited thereto.

In some other embodiments, the optimization process is executed based on the task type. For example, determine a strategy for selecting a target node based on the task type. For example, in the process of determining the target node of one subtask, the target node determined for the previous subtask is first selected. For example, if sub-task 1 and sub-task 2 are included in a task, and node 1 and node 2 are included in the target node of sub-task 1, sub-task 2 can select node 1 and node 2 as their target nodes. . In some optional embodiments, the score of each pre-selection node included in the pre-selection node set of the sub-task is determined based on the task type and the current state information of each node included in the pre-selection node set of the sub-task. can In some other embodiments, the score of each preselection node included in the preselection node set of the subtask may be determined according to other methods, which will not be described further herein.

In one example, for a communication-intensive task, the number of target nodes is reduced as much as possible to reduce the communication cost between the sub-tasks, so in the sub-task of the communication-intensive task, the score of the node where the resource distribution is concentrated is generally The distribution is higher than the distributed nodes. For example, since the usable disk capacity of the preselection node 1 of one subtask is distributed over two disks, and the usable disk capacity of the preselection node 2 is distributed on one disk, the preselection of the subtask is Node 2 has a higher score than pre-selection Node 1. In another example, for a computationally intensive task, the score of a node with strong computational power is generally higher than that of a node with weak computational capability. For example, if the processor of the preselection node 1 of one subtask includes one processor core, and the processor of the preselection node2 of the subtask includes two processor cores, the preselection node2 of the subtask includes two processor cores. has a higher score than preselection node 1.

The above process describes a sub-task as a resource allocation object. In some embodiments, it is also possible to divide a plurality of subtasks of the task into at least one group, wherein each group includes at least one subtask of the plurality of subtasks; Allocate a respective corresponding target node for each group in at least one group of the task, wherein the same target node is allocated to each subtask in the same group.

The number of groups is the same as the number of nodes required by the task. That is, when a plurality of subtasks of a task are allocated to N target nodes, the plurality of subtasks of the task can be divided into N groups. Here, the number of sub-tasks in each group may be the same or different. For example, a plurality of subtasks are randomly assigned, subtasks with similar priorities are divided into one group, or subtasks with or without dependency are divided into one group, or resources A subtask having a large difference in demand quantity is divided into one group, or a subtask having the same requested resource type is divided into one group, etc. The embodiment of the present invention does not limit the specific grouping method. In this way, it is possible to improve the efficiency of the resource allocation process under the premise that the target node is not significantly affected on the execution efficiency of the task by determining the target node for each group.

In some embodiments, a plurality of sub-tasks of the task may be divided into at least one group based on the task type of the task. For example, in a communication-intensive task, the communication cost between the sub-tasks can be reduced by allocating a plurality of sub-tasks of the task to a target node having a small number, for example, all sub-tasks among the plurality of sub-tasks. You can divide tasks into the same group so that all subtasks run on the same target note. In some optional examples, a total number of nodes required for a plurality of sub-tasks of the task may be determined according to resource demand information of the plurality of sub-tasks of the task; A plurality of sub-tasks of the task may be divided into at least one group based on the total number of nodes. Further, for example, in a computation intensive task, each subtask among a plurality of subtasks of the task can be grouped into one group. In other words, the target node can be determined for each subtask as a unit.

For example, if one task includes 4 subtasks, each subtask requires one GPU, and it is assumed that the number of GPUs available for each node in the distributed system is 2, then at least 2 nodes 4 subtasks of the above tasks should be executed using . Therefore, if the task is a communication-intensive task, the four sub-tasks are divided into two groups on average, where the group may be randomly executed or executed based on resource demand information of the sub-task, etc. . If the task is a computationally intensive task, these four subtasks can be divided into four groups, and each subtask can be made into one group.

When a group is divided, optionally, the determination of the target node of the sub-task can also be applied to the determination of the target node of the group, which will be introduced in detail in Xiamen.

For example, in some embodiments, if there are target nodes all corresponding to each group in at least one group of the task, the currently available resources in the distributed system satisfy the total resources required for a plurality of subtasks of the task. decide In some other embodiments, when a target node corresponding to at least one group of the task does not exist, it is determined that currently available resources in the distributed system do not satisfy the total resources required for a plurality of subtasks of the task.

Also, for example, based on the current resource status information of the distributed system and the resource demand information of each group in the at least one group, a target node may be determined for each group of the at least one group in order according to a specific order. have. The resource demand information of the group may be determined based on the resource demand information of at least one subtask included in the group. For example, the resource demanded by the group is the same type of resource demanded by all the subtasks included in the group. It can be the sum of resources. Here, the specific order may be selectively determined according to at least one or other factors among priorities and dependencies of the plurality of groups. The priority of the group may be determined based on the priorities of two or more subtasks included in the group, for example, the priority of the group is determined as the highest priority of the two or more subtasks, or Determining the priority as the average priority of two or more subtasks, and so on. The dependency relationship between groups is determined based on the dependency relationship between sub-tasks included in different groups. For example, if it is assumed that one sub-task in group 1 depends on one sub-task in group 2, Determining that group 1 depends on group 2, etc., the embodiment of the present invention is not limited thereto.

In some embodiments, the preselection node set of each group is also determined based on the current resource state information of the distributed system and the resource demand information of each group, and in a predetermined order from the preselection node set of each group in order A target node of the group may be determined.

In an embodiment of the present invention, a target node is first determined for all tasks. For example, the preselection and optimization operations are executed to finally obtain information on the target node corresponding to the entire task, and then perform the task again. is divided into sub-task level and bound once, that is, the target node is bound by sub-task unit.

In some embodiments, when the currently available resources in the distributed system do not satisfy the total resources required for a plurality of subtasks of the task, or determining a corresponding target node for at least one subtask of the plurality of subtasks If this is not successful, the delay assignment may be executed for all sub-tasks included in the task. The delayed assignment may be to perform assignment for a plurality of sub-tasks included in the task in a next assignment cycle. In the next allocation period, it is still possible to allocate the target node corresponding to each of the plurality of subtasks of the task only when the currently available resources in the distributed system satisfy the total resources required for the plurality of subtasks of the task. If, in the next allocation period, the currently available resources in the distributed system still do not satisfy the total resources required for a plurality of subtasks of the task, the delayed allocation is continued.

In some embodiments, after each determination of the target node corresponding to the current sub-task or group, the virtual allocation of the target node corresponding to the current sub-task or group is executed, or the target node corresponding to the current sub-task or group is executed. Virtual binding of resources can be implemented. When the currently available resources in the distributed system do not satisfy the total resources required for a plurality of subtasks of the task, for example, when a corresponding target node cannot be found for a subtask, the virtual allocation or virtual binding is performed. It can be regarded as a failure and the corresponding target node and its resources can be allocated to other tasks.

In some embodiments, after allocating a target node corresponding to each of the plurality of subtasks of the task, synchronous scheduling may be performed on the plurality of subtasks of the task.

It should be explained that, in the embodiment of the present invention, assignment refers to allocating a subtask to a corresponding target node, and scheduling refers to allocating a subtask to a corresponding target node, and then the corresponding target node executes the subtask say to do

As shown in Fig. 4, in the traditional scheduling method, in order to implement gang scheduling, the scheduler processes a plurality of sub-tasks among tasks in order, and when one sub-task is scheduled, a target corresponding to the sub-task immediately It is actually allocated to a node (not a simulated allocation according to the embodiment of the present invention), and the resource of the target node is occupied and scheduled for the subtask. In one task, the currently available resource when allocated by each subtask in the task is dynamically changed and gradually decreases. This scheduling method is referred to as a task by task scheduling method. In addition, when scheduling of sub-tasks is performed, the scheduling situation and resource demand of a plurality of sub-tasks in the same task are not fully considered, resulting in that all allocations are independent actions each time. However, in a practical case, there is often some association between a plurality of subtasks in the same task. For example, some tasks waiting to be scheduled often involve a large amount of communication demand. If some subtasks of one task If the communication overhead is allocated to a plurality of nodes, it will have a distinct effect on the execution efficiency of the entire task.

In the embodiment of the present invention, sub-task allocation and scheduling are performed at task granularity. As shown in Fig. 5, first, all tasks are still acquired according to the priority, and for each task, the whole task is considered comprehensively, and the currently available resources in the distributed system are the total resources required for a plurality of sub-tasks of the task. A target node corresponding to each of the plurality of sub-tasks of the task is allocated only when . .

Unlike executing the task-by-task scheduling method according to the demand of the sub-task, the embodiment of the present invention considers the entire task as a whole when selecting a node for the task, and considers the entire task as the minimum unit of scheduling. It can better utilize the interrelationship and demand situation of each subtask, and select a more suitable scheduling node in the cluster for the whole task, improving the execution efficiency of the task and the resource utilization efficiency of the cluster.

6 , this is a schematic diagram of scheduling logic according to some embodiments of the present invention. In order to reduce scheduling complexity, the following assumptions may be made for each scheduled waiting task.

(1) Each task has a plurality of sub-tasks;

(2) each sub-task is homogeneous in resource demand, that is, the demand for resources is the same;

(3) The task submitted by each user specifies the resource demand for each sub-task when submitted, and the quantity of the sub-task and the resource demand for each sub-task remain unchanged during the entire task processing process.

First, a preselection process is executed to remove nodes that do not satisfy the user's task. Specifically, it includes the following steps.

(1) For each subtask of the task, based on the current resource state information of the distributed system and the resource demand information of each subtask, a set of preselection nodes of the subtask is determined, respectively, and the preselection node of the subtask A node outside the set cannot necessarily support the execution of the subtask on that node. Here, a task may be referred to as a job and a sub-task may be referred to as a task. The set of preselection nodes of different subtasks may be the same or different.

In some optional examples, a preselection node of a subsequent subtask may be determined from a set of preselection nodes of a previous subtask, wherein the order of the subtasks may optionally be determined based on the priority of the subtasks . For example, if it is assumed that nodes 1 to 5 are included in the distributed system, and one task includes sub-task 1, sub-task 2, and sub-task 3, first, a preselection node for sub-task 1 is determined and It can be assumed that node1, node2 and node3 are; Next, the pre-selection nodes of the sub-task 2 are determined in the nodes 1, 2 and 3, assuming that they are the nodes 2 and 3; Again, it can be assumed that it is node 3 by determining the pre-selection node of sub task 3 in node 2 and node 3 again. Through such a method of determining the pre-selection node set, it is possible to improve the efficiency of determining the pre-selection node set for a task with a certain amount of communication demand, particularly for a communication-intensive task, and further improve the task allocation efficiency.

(2) it is possible to verify each preselection node in the preselection node set to determine whether the total resource on the preselection node satisfies the total resource demands of all subtasks assigned to the preselection node. If it is satisfied, the optimization process is run; if not, the preselection process is run again.

Here, this step is optional, and verification of the pre-selection node can improve the reliability of the overall task assignment. Before executing the pre-selection process, first, node set partitioning can be performed, that is, when each node is divided into groups, the criterion for dividing the group into groups can be the cluster to which the node belongs, that is, nodes on different clusters are divided into different groups. share By performing node set partitioning, different tasks can be assigned only to nodes of a specified group.

Next, the optimization process is executed, and the target node most suitable for the user's task is selected. Specifically, it may include the following steps.

(1) Acquire the amount of computation and communication during the training process of the task based on the information provided by the user in the user task and the past experience of the scheduler, and determine whether the user's task is of the computation-intensive type or the communication-intensive type based on the amount of computation and communication.

In some optional examples, the determination of the task type may be performed during or before the preselection process, and the preselection node set of the subtask may be determined based on the task type.

In some optional examples, target nodes of the plurality of sub-tasks are determined during the optimization process based on the task type. If it is a communication-intensive task, the task is placed (i.e., allocated) on the physical node with the smallest quantity (i.e., the target node), and the number of specifically allocated physical nodes is jointly determined by the node's available resources and the resources required for the task. , if this arrangement constraint cannot be satisfied, delay assignment is performed for all subtasks in the task. If it is a computationally intensive task, the allocation is executed if it can satisfy the task demand by selecting a node with sufficient resources even if it does not satisfy the fewest physical nodes by relaxing these placement constraints; otherwise, the allocation is postponed.

(2) For the cluster nodes, scoring and arranging them in order from largest to smallest according to surplus resources to obtain a mapping score from the set of pre-selection nodes to tasks, where the score of the pre-selection node is equal to the node load in the cluster of the distributed system It can be decided according to the node resource fragmentation situation. The node load refers to the total amount of resources currently available and resource usage of each node in the cluster at the current time. The resource fragmentation situation refers to resource distribution among nodes.

(3) For all the nodes selected in the above step, the tasks of the tasks are arranged in order from the highest score to the lowest score until they cannot be settled or all tasks of the task are arranged.

(4) The task state is detected. If there are still no surplus tasks assigned to the task, it explains that the allocation failed due to insufficient cluster resources, and the allocation is delayed by releasing the resources occupied by the task. If there are no redundant tasks in the job and the task type is communication-intensive, it is determined whether the smallest node demand is satisfied. If it is satisfied, the allocation was successful. If not, the allocation fails and the allocation is delayed by releasing the resource.

The method according to an embodiment of the present invention may be applied to a task arranging system or a cloud platform, and the task arranging system may be implemented based on a platform such as Kubernetes.

A person of ordinary skill in the art, in the described method of a particular implementation, does not imply that the drafting order of each step is strictly executed, the order in which it is executed forms any limitation on the implementation process, and that the It will be appreciated that the particular execution order should be determined by its function and possibly native logic.

7, an embodiment of the present invention further provides a task scheduling apparatus, the apparatus comprising:

an acquiring module 701 for acquiring information on a task including at least one of a task type and resource demand information of the task;

and a first assignment module 702 for allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task.

In some embodiments, the resource demand information of the task is resource demand information of a plurality of sub-tasks included in the task; The first allocation module 702 is further configured to, when the currently available resources in the distributed system satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task, a target corresponding to each of the plurality of subtasks of the task Allocate a node.

In some embodiments, the apparatus further comprises a first determining module, configured to determine a target node for the plurality of sub-tasks based on current resource state information of the distributed system and resource demand information of the plurality of sub-tasks do; The first allocation module allocates a target node corresponding to each of the plurality of subtasks of the task when a corresponding target node is successfully determined for each subtask among the plurality of subtasks.

In some embodiments, the apparatus further comprises a first determining module, configured to determine a target node for the plurality of sub-tasks based on current resource state information of the distributed system and resource demand information of the plurality of sub-tasks do; The first allocation module allocates a target node corresponding to each of the plurality of subtasks of the task when a corresponding target node is successfully determined for each subtask among the plurality of subtasks.

In some embodiments, the first determining module determines a target node corresponding to a plurality of sub-tasks in order according to a specific order based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks do.

In an optional example, the resource demand information of the plurality of subtasks includes resource demand information of each subtask among the plurality of subtasks, wherein the resource demand information includes a required resource type and a quantity of each type of resource or may further include other information.

In an optional example, the current resource status information of the distributed system may include current status information of a plurality of nodes in the distributed system, wherein the current status information includes availability of resources, types and quantities of available resources, and loads It indicates at least one of context and topological connection information, or further includes other information.

In some embodiments, the device is configured to: after determining a target node for one sub-task each time, update the current resource status information of the distributed system, and based on the updated current resource status information, the target corresponding to the next sub-task and an update module for determining the node.

In some embodiments, the sequential determination order of the target nodes of the plurality of sub-tasks is determined based on at least one of a priority of the sub-tasks and a dependency relationship between the sub-tasks.

In some embodiments, the first determining module is configured to sequentially determine a target node for the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks. decision unit; an update unit configured to update the current resource state information of the distributed system according to the resource demand information of the one subtask after each time determining the target node of the one subtask.

In some embodiments, the device is configured to, when the target node for at least one subtask among the plurality of subtasks is not successfully determined, the currently available resources of the distributed system are allocated to each subtask of the plurality of subtasks. and a second determining module for determining that does not satisfy the required resource.

In some embodiments, the first determining unit is configured to: determine a preselection node set of each subtask among the plurality of subtasks based on the current resource state information of the distributed system and the resource demand information of the plurality of subtasks a decision sub-unit for; and a selection sub-unit for selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks.

Here, the preselection node sets of different subtasks may be the same or different. In some examples, a plurality of sub-tasks may have the same set of pre-selection nodes.

In some embodiments, the determining sub-unit determines the pre-selection node set of each sub-task based on the current resource state information of the distributed system and the resource demand information of each sub-task.

Determination of the pre-selection node set of the plurality of sub-tasks is independent and may not have a mutual dependency, for example, may be executed in parallel or may be executed in an arbitrary precedence order. For example, the preselection node set of the plurality of subtasks is determined based on the current resource state information of the same distributed system, that is, the current resource state information of the distributed system may not be updated during the preselection process. For example, the determination of the pre-selection node set of each sub-task is only related to its own resource demand information and is independent of the resource demand information of other sub-tasks.

In some embodiments, the determining subunit determines a set of preselection nodes of each subtask among the plurality of subtasks in an order according to a specific order, wherein the preselection of a first subtask among the plurality of subtasks The preselection node of the node set is selected from a preselection node set of a second subtask among the plurality of subtasks, wherein the order of the second subtask is located before the first subtask.

In some embodiments, a target node of each subtask is selected from a preselection node of each subtask of the plurality of subtasks in order according to a specific order, wherein first a target of a second subtask of the plurality of subtasks A node is selected as a target node of a first sub-task among the plurality of sub-tasks, wherein the order of the second sub-task is located before the first sub-task.

In some embodiments, the selection sub-unit is configured to: determine a target node of the first sub-task from a set of pre-selection nodes of a first sub-task among the plurality of sub-tasks based on the current resource state information of the distributed system; Update the current resource state information of the distributed system according to the resource demand information of the first sub-task, and according to the updated current resource state information of the distributed system, a dictionary of a second sub-task of the plurality of sub-tasks A target node of the second sub-task is determined from the selected node set.

In an optional example, determining a target node of the first subtask in the preselection node set of the first subtask based on current state information of each preselection node in the preselection node set of the first subtask, and The current state information of the target node of the first sub-task is updated based on the required resource information of the first sub-task.

In some embodiments, the selection sub-unit is configured to: determine a score of a pre-selection node included in a set of pre-selection nodes of each sub-task according to the task type of the task; A target node of each subtask is selected from the set of preselection nodes of each subtask based on the scores of the preselection nodes included in the preselection node set of each subtask.

In an optional example, a scoring strategy of each pre-selection node in the set of pre-selection nodes is determined based on the task type of the task.

In some embodiments, the apparatus further comprises: a third determining module for determining a task type of the task; Here, the target node corresponding to the plurality of sub-tasks is determined based on the task type of the task.

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

In some embodiments, the group unit divides the plurality of sub-tasks of the task into at least one group according to the 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 communication-intensive, the group unit determines a total number of nodes required for the plurality of sub-tasks according to resource demand information of the plurality of sub-tasks of the task, and A plurality of sub-tasks of the task are divided into at least one group according to the total number of nodes.

In some embodiments, when the task type of the task is a computation intensive type, each subtask among the plurality of subtasks is set as a group.

In some embodiments, a set of preselected nodes is determined for each group of the at least one group, and in order, a target node of the group is selected from the set of preselection nodes of each group of the at least one group.

In some embodiments, the first allocation module is further configured to: determine a target node for a plurality of sub-tasks of the task based on a task type of the task; When all corresponding target nodes for the plurality of subtasks of the task are successfully determined, target nodes are allocated for the plurality of subtasks included in the task.

In some embodiments, the apparatus is a delay module configured to, when not successfully determining a corresponding target node for at least one of the sub-tasks, execute delay assignment for all of the plurality of sub-tasks of the task. further includes

In some embodiments, when the currently available resources in the distributed system do not satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task, the delay module is configured to: Execute lazy assignment.

In some embodiments, the apparatus further comprises: a scheduling module for allocating a target node corresponding to each of the plurality of subtasks of the task, and then executing synchronous scheduling on the plurality of subtasks of the task.

In some embodiments, the apparatus is applied to a task arranging system.

In some embodiments, a function provided by an apparatus provided in an embodiment of the present invention or a module included may be used to execute the method described in the method embodiment, for specific implementations, see the description of the method embodiment All I can do is not explain it further here for the sake of brevity.

The device examples described above are merely examples, modules described as separate members may or may not be physically separated, and members displayed as modules may or may not be physical units, that is, located in one place or , may be distributed across a plurality of network modules. Some or all of the modules may be selected according to actual requirements for implementing the objectives of the solutions of the present disclosure. Those of ordinary skill in the art can understand and implement them without creative work.

Embodiments of the apparatus of the present specification may be applied to a computer device such as a server or a terminal device. Embodiments of the apparatus may be implemented by software, hardware, or a combination thereof. Taking as an example implemented by software, a logical device that reads and executes, by a processor that processes a file on which the device resides, corresponding computer program instructions from a non-volatile storage component to memory and back to a process, is read and executed. will form From a hardware point of view, as shown in FIG. 8 , this is a hardware structural diagram of a computer device in which the apparatus of the present specification is located, and the processor 801 , the memory 802 , the network interface 803 and the In addition to the volatile storage component 804 , in an embodiment, a server or electronic device in which the device is located may further include other hardware according to the actual function of the computer device in question, which will not be described further.

Correspondingly, an embodiment of the present invention further provides a computer storage medium in which a computer program is stored, wherein the program implements the method according to any one embodiment when executed by a processor.

Correspondingly, an embodiment of the present invention further provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the program according to any one embodiment implement the method

The present invention may use the form of a computer program product implemented in one or more storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) including program code. Computer-readable media includes permanent and non-persistent, removable and non-removable media, and information storage may be realized by any method or technology. Information may be computer readable instructions, data structures, program modules, or other data. Examples of computer storage media include 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, CD-ROM, digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape may be configured to store information that can be accessed by computing devices, including, but not limited to, storage or other magnetic storage devices or any other non-transmission medium.

Other implementations of the disclosure will be apparent to those skilled in the art from consideration of the disclosure and practice of the disclosure herein. This disclosure is intended to follow the general principles of the present disclosure and is intended for any variation, usage, modification, or modification of the disclosure including ordinary skill in the relevant art or ordinary skill in the art for which flowcharts not in this disclosure are intended. It is intended to encompass adaptation. It is intended that the present disclosure and embodiments be regarded as illustrative only, with the true scope and spirit of the present disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the structure precisely described above and shown in the accompanying drawings, and that various modifications and changes may be made thereto without departing from its scope. The protection scope of the present disclosure is limited only by the appended claims.

The above description is only preferred examples of the present disclosure, and is not intended to limit the present disclosure, and any modifications, equivalents, improvements, etc. made within the spirit and principles of the present disclosure are within the scope of the present disclosure. should be included in

Claims (23)

A task scheduling method comprising:
obtaining information on a task including at least one of a task type and resource demand information of the task; and
Allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task
Task scheduling method, characterized in that. According to claim 1,
the resource demand information of the task is resource demand information of a plurality of sub-tasks included in the task;
Allocating a target node to each of a plurality of sub-tasks included in the task based on the information of the task comprises:
allocating a target node corresponding to each of the plurality of subtasks of the task when the currently available resources in the distributed system satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task
Task scheduling method, characterized in that. 3. The method of claim 2,
determining a target node for the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks;
allocating a target node corresponding to each of the plurality of subtasks of the task when the currently available resources in the distributed system satisfy total resources corresponding to resource demand information of the plurality of subtasks of the task;
allocating a target node corresponding to each of the plurality of subtasks of the task when a corresponding target node is successfully determined for each subtask of the plurality of subtasks;
Task scheduling method, characterized in that. 4. The method of claim 3,
Determining a target node for the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks includes:
determining a target node for the plurality of sub-tasks in order based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks; and
After determining the target node of each sub-task, updating the current resource state information of the distributed system based on the resource demand information of the one sub-task
Task scheduling method, characterized in that. 5. The method according to claim 3 or 4,
When a target node for at least one subtask among the plurality of subtasks is not successfully determined, it means that the currently available resources of the distributed system do not satisfy the resource demand of each subtask among the plurality of subtasks. to decide
Task scheduling method, characterized in that. 6. The method according to any one of claims 3 to 5,
Determining a target node for the plurality of sub-tasks based on the current resource state information of the distributed system and the resource demand information of the plurality of sub-tasks includes:
determining a preselection node set of each subtask among the plurality of subtasks based on the current resource state information of the distributed system and the resource demand information of the plurality of subtasks; and
selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks
Task scheduling method, characterized in that. 7. The method of claim 6,
Determining a preselection node set of each subtask among the plurality of subtasks comprises:
determining a preselection node set of each subtask among the plurality of subtasks sequentially according to a specific order, wherein a preselection node of a preselection node set of a first subtask among the plurality of subtasks is selected from the plurality of subtasks. It is selected from the pre-selection node set of the second sub-task among the sub-tasks, and the order of the second sub-task is located before the first sub-task.
Task scheduling method, characterized in that. 8. The method according to claim 6 or 7,
The step of selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks comprises:
selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks in order according to a specific order, wherein first, a target node of a second sub-task of the plurality of sub-tasks is selected as a target node of a first subtask among the plurality of subtasks, and the order of the second subtask is located before the first subtask
Task scheduling method, characterized in that. 9. The method according to any one of claims 6 to 8,
The step of selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task among the plurality of sub-tasks comprises:
determining a score of a preselection node included in a preselection node set of each subtask based on the task type of the task; and
selecting a target node of each sub-task from a set of pre-selection nodes of each sub-task based on the scores of pre-selection nodes included in the set of pre-selection nodes of each sub task
Task scheduling method, characterized in that. 10. The method according to any one of claims 3 to 9,
The order in which the plurality of sub-tasks sequentially determines the target node is determined based on at least one of a priority of the sub-task and a dependency relationship between the sub-tasks.
Task scheduling method, characterized in that. 11. The method according to any one of claims 1 to 10,
Allocating a target node to each of a plurality of sub-tasks included in the task based on the information of the task comprises:
determining a task type of the task; and
allocating a target node corresponding to each of a plurality of sub-tasks of the task based on the task type of the task
Task scheduling method, characterized in that. 12. The method according to any one of claims 1 to 11,
Allocating a target node corresponding to each of the plurality of sub-tasks of the task comprises:
dividing the plurality of subtasks of the task into at least one group, each group including at least one subtask of the plurality of subtasks; and
allocating a corresponding target node to each of the at least one group;
Allocating the same target node to each subtask in the same group
Task scheduling method, characterized in that. 13. The method of claim 12,
The step of dividing the plurality of sub-tasks into at least one group includes:
dividing a plurality of sub-tasks of the task into at least one group based on the task type of the task;
Task scheduling method, characterized in that. 14. The method according to any one of claims 11 to 13,
The task type of the task is a computation intensive type or a communication intensive type.
Task scheduling method, characterized in that. 15. The method according to any one of claims 12 to 14,
The step of dividing the plurality of sub-tasks into at least one group includes:
When the task type of the task is communication-intensive, the total number of nodes required for the plurality of sub-tasks is determined based on resource demand information of the plurality of sub-tasks of the task, and the number of total nodes of the task is determined based on the total number of nodes. dividing the plurality of sub-tasks into at least one group;
and/or
When the task type of the task is a computation intensive type, grouping each sub-task among the plurality of sub-tasks into one group;
Task scheduling method, characterized in that. 16. The method according to any one of claims 11 to 15,
Allocating a target node corresponding to each of the plurality of sub-tasks of the task based on the task type of the task includes:
determining a target node for a plurality of sub-tasks of the task based on the task type of the task; and
allocating target nodes for a plurality of sub-tasks included in the task when all corresponding target nodes for the plurality of sub-tasks of the task are successfully determined
Task scheduling method, characterized in that. 17. The method of claim 16,
If a corresponding target node is not successfully determined for at least one subtask among the subtasks, the method further comprising the step of executing delay allocation for all of the plurality of subtasks of the task
Task scheduling method, characterized in that. 18. The method according to any one of claims 2 to 17,
When the currently available resources in the distributed system do not satisfy the total resources corresponding to the resource demand information of the plurality of subtasks of the task, the method further includes the step of executing delayed allocation for all the subtasks included in the task. doing
Task scheduling method, characterized in that. 19. The method according to any one of claims 1 to 18,
Allocating a target node corresponding to each of the plurality of subtasks of the task, and then performing synchronous scheduling on the plurality of subtasks of the task
Task scheduling method, characterized in that. 20. The method according to any one of claims 1 to 19,
The method is applied to a task arranging system
Task scheduling method, characterized in that. an acquiring module for acquiring information on a task including at least one of a task type and resource demand information of the task; and
a first allocation module for allocating a target node to each of a plurality of sub-tasks included in the task based on the information on the task;
Task scheduling device, characterized in that. A computer-readable storage medium storing a computer program, comprising:
21. Implementing the method according to any one of claims 1 to 20 when the program is executed by a processor.
A computer-readable storage medium, characterized in that. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the computer program comprising:
21. Implementing the method according to any one of claims 1 to 20 when the processor executes the program.
Computer device, characterized in that.

Images