TWI786564B - Task scheduling method and apparatus, storage media and computer equipment - Google Patents

Abstract

Examples of the present disclosure provide a task scheduling method and apparatus. According to the method, information of a task may be obtained, where the information of the task includes at least one of a task type and resource requirement information of the task; a target node can be assigned to each of a plurality of subtasks included in the task according to the information of the task.

Description

Task scheduling method and device, storage medium and computer equipment

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

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

The present disclosure provides a task scheduling scheme.

According to the first aspect of an embodiment of the present disclosure, there is provided a task scheduling method, the method comprising: obtaining task information, the task information including at least one of the task type and resource requirement information of the task; according to the task information, respectively Assigning target nodes to multiple subtasks included in the task.

In some embodiments, the resource requirement information of the task is the resource requirement information of a plurality of subtasks included in the task; the method further includes: currently available resources in the distributed system satisfy the plurality of subtasks of the task In the case of the total resource corresponding to the resource requirement information of , assign corresponding target nodes to multiple subtasks of the task respectively.

In some embodiments, the method further includes: determining a target node for a plurality of subtasks of the task based on resource requirement information of the plurality of subtasks; the currently available resources in the distributed system satisfy the task In the case of the total resource corresponding to the resource requirement information of multiple subtasks, assigning corresponding target nodes to the multiple subtasks of the task respectively includes: after successfully determining the corresponding target for each subtask in the multiple subtasks In the case of a node, assign corresponding target nodes to multiple subtasks of the task.

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

In an optional example, the resource requirement information of the plurality of subtasks includes resource requirement information of each subtask in the plurality of subtasks, wherein the resource requirement information may include required resource types and resources of each type. The number of resources, or further include other information.

In an optional example, the current resource status information of the distributed system may include current status information of multiple nodes in the distributed system, wherein, when The previous state information is used to indicate whether the resource is available, the type and quantity of the available resource, the load condition, the topological connection information, or further include other information.

In some embodiments, after each determination of the target node assigned to a subtask, 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, in the event that the corresponding target node for at least one subtask of the plurality of subtasks fails to be determined, it is determined that currently available resources of the distributed system do not meet the requirements for each of the plurality of subtasks. resources required by each subtask.

In some embodiments, the sequentially determined order of the target nodes of the plurality of subtasks is obtained according to at least one of priorities of subtasks and dependencies between subtasks.

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

Wherein, the preselected node sets of different subtasks may be the same or different. In some examples, multiple subtasks may have the same set of preselected nodes.

In some embodiments, the plurality of subtasks are determined based on the current resource status information of the distributed system and the resource requirement information of the plurality of subtasks. The preselected node set for each subtask in the task includes: determining the preselected node set for each subtask based on the current resource status information of the distributed system and the resource requirement information of each subtask.

The determination of the pre-selected node sets of multiple subtasks may be independent without interdependence, for example, may be executed in parallel or in any sequence. For example, the preselected node sets of multiple subtasks are determined based on the same current resource state information of the distributed system, that is, the current resource state information of the distributed system is not updated during the preselection process. For example, the determination of the pre-selected node set of each subtask is only related to its own resource requirement information, but has nothing to do with the resource requirement information of other subtasks.

In some embodiments, the determining the preselected node set for each of the multiple subtasks includes: sequentially determining the preselected node set for each of the multiple subtasks in a specific order.

In one example, the preselected nodes in the preselected node set of the first subtask among the multiple subtasks are selected from the preselected node set of the second subtask among the multiple subtasks, wherein the second subtask The order of is located before the first subtask.

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

In some embodiments, the selecting the target node of each subtask from the preselected node set of each subtask in the plurality of subtasks includes: sequentially selecting the target node of each subtask in the plurality of subtasks in a specific order Select the selected node in the pre-selected Describe the target node of each subtask.

In an optional example, the target node of the second subtask among the plurality of subtasks is preferentially selected as the target node of the first subtask among the plurality of subtasks, wherein the order of the second subtask is located in the Before the first subtask.

In some embodiments, the selecting the target node of each subtask from the preselected node set of each subtask in the plurality of subtasks includes: based on the current resource state information of the distributed system, from the Determining the target node of the first subtask from the preselected node set of the first subtask among the plurality of subtasks; updating the current resource state information of the distributed system based on the resource requirement information of the first subtask, and based on The updated current resource status information of the distributed system determines the target node of the second subtask from the preselected node set of the second subtask in the plurality of subtasks.

In an optional example, based on the current state information of each preselected node in the first subtask's preselected node set, the target node of the first subtask is determined from the first subtask's preselected node set, and Based on the required resource information of the first subtask, current state information of the target node of the first subtask is updated.

In some embodiments, the selecting the target node of each subtask from the preselected node set of each subtask in the plurality of subtasks includes: determining the preselected node of each subtask according to the task type of the task The score of the preselected nodes included in the set; based on the score of the preselected nodes included in the preselected node set of each subtask, select the target node of each subtask from the preselected node set of each subtask.

In an optional example, based on the task type of the task, a score determination strategy for each preselected node in the preselected node set is determined.

In some embodiments, assigning target nodes to multiple subtasks included in the task respectively according to the information of the task includes: determining the task type of the task; The multiple subtasks of are allocated to corresponding target nodes.

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

In some embodiments, the assigning corresponding target nodes to the multiple subtasks of the task includes: dividing the multiple subtasks of the task into at least one group, wherein each group includes the multiple subtasks at least one subtask in the at least one group; each group in the at least one group is assigned a corresponding target node, wherein each subtask in the same group is assigned to the same target node.

In some embodiments, the dividing the multiple subtasks of the task into at least one group includes: dividing the multiple subtasks of the task into at least one group according to the 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, according to the resource requirements of the plurality of subtasks of the task information, determine the total number of nodes required by the multiple subtasks, and divide the multiple subtasks of the task into at least one group according to the total number of nodes.

In some embodiments, the multiple subtasks of the task are divided into At least one grouping includes: when the task type of the task is computationally intensive, each subtask in the plurality of subtasks is regarded as a grouping.

In some embodiments, respectively assigning corresponding target nodes to each group in the at least one group includes: respectively determining a preselected node set for each group in the at least one group, and sequentially selecting from the at least one group The target node of the group is selected from the preselected node set of each group.

In some embodiments, the assigning corresponding target nodes to multiple subtasks of the task based on the task type of the task includes: determining for the multiple subtasks of the task based on the task type of the task Target node: assigning target nodes to the multiple subtasks included in the task when the corresponding target nodes are all successfully determined for the multiple subtasks of the task.

In some embodiments, the method further includes: in a case that determining the corresponding target node for at least one of the subtasks is unsuccessful, performing delayed allocation on multiple subtasks of the task.

In some embodiments, the method further includes: when the current available resources in the distributed system do not meet the total resources corresponding to the resource requirement information of the multiple subtasks of the task, for the multiple subtasks included in the task Tasks are assigned with delay.

In some embodiments, the method further includes: after assigning corresponding target nodes to the multiple subtasks of the task, synchronously scheduling the multiple subtasks of the task.

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

According to a second aspect of an embodiment of the present disclosure, a task scheduling device is provided, The device includes: an acquisition module for acquiring task information, the task information including at least one of the task type and resource requirement information of the task; a first allocation module for information, respectively assigning target nodes to multiple subtasks included in the task.

According to a third aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method described in any embodiment is implemented.

According to a fourth aspect of the embodiments of the present disclosure, there is provided a computer device, including a memory, a processor, and a computer program stored in the memory and operable on the processor. When the processor executes the program, any The method described in the examples.

In an embodiment of the present disclosure, task information can be obtained, and the task information includes at least one of the task type and resource request information of the task, wherein the resource request information of the task is a plurality of subtasks included in the task resource requirements information for . Allocation of target nodes can be performed according to the task type or resource requirement information of the task. For example, when the currently available resources in the distributed system satisfy the total resources corresponding to the resource requirement information of the multiple subtasks of the task, assign corresponding target nodes to the multiple subtasks of the task respectively. In this way, the resource requirements of the entire task can be obtained, resource allocation is implemented at the granularity of the entire task, and the resource utilization rate of the cluster is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

201: Obtain task information, where the task information includes at least one of the task type and resource requirement information of the task

202: According to the information of the task, assign target nodes to multiple subtasks of the task respectively

701: Get module

702: The first allocation module

801: Processor

802: random access memory

803: Network interface

804: Non-volatile memory

FIG. 1 is a schematic diagram of a decentralized system of an embodiment of the present disclosure.

Fig. 2 is a flowchart of a task scheduling method according to an embodiment of the present disclosure.

FIG. 3A is a schematic diagram of resource changes in the process of simulating allocation of nodes according to an embodiment of the present disclosure.

FIG. 3B is a flow diagram of a simulated allocation node process of an embodiment of the disclosure.

Fig. 4 is a schematic diagram of a traditional task scheduling process.

Fig. 5 is a schematic diagram of a task scheduling process according to an embodiment of the present disclosure.

FIG. 6 is a schematic diagram of scheduling logic of an embodiment of the disclosure.

Fig. 7 is a schematic structural diagram of a task scheduling device according to an embodiment of the present disclosure.

FIG. 8 is a schematic structural diagram of a computer device of an apparatus according to an embodiment of the present disclosure.

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numerals in different drawings indicate the same or similar elements. The implementations described in the following exemplary examples do not represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatuses and methods consistent with aspects of the present disclosure as detailed in the appended claims.

The terminology used in the present disclosure is for the purpose of describing particular embodiments only, and is not intended to limit the present disclosure. As used in this disclosure and appended claims The singular forms "a", "said" and "the" are also intended to include the plural unless the context clearly dictates otherwise. It should also be understood that the term "and/or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items. In addition, the term "at least one" herein means any one of a plurality or any combination of at least two of a plurality.

It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various pieces of information, these pieces of information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of the present disclosure, first information may also be called second information, and similarly, second information may also be called first information. Depending on the context, the word "if" as used herein may be interpreted as "at" or "when" or "in response to a determination."

At present, more and more tasks are being processed by decentralized systems, and such tasks are called decentralized tasks. These distributed tasks include GPU (Graphics Processing Unit, graphics processing unit), DSP (Digital Signal Processor, digital signal processor), FPGA (Field Programmable Gate Array, field programmable gate array) and other high-performance processing resources Tasks with demanding requirements, e.g. deep learning tasks. As shown in Figure 1, it is a distributed system of some embodiments, the distributed system includes one or more clusters, each cluster includes one or more servers, and each server can be regarded as a node (in the figure Each node includes at least one of CPU (Central Processing Unit, central processing unit), GPU, random access memory, disk, network interface and other resources. After a subtask is assigned to a node, it can be executed by resources on that node, each A node can execute one or more subtasks. The subtasks performed by each node in the same cluster can be the same or different. In a distributed system, both the task itself and the resource requirements of the task are completely different from traditional tasks. Therefore, it is necessary to adjust the task scheduling method in the distributed system accordingly, so as to improve the efficiency of the distributed system. Rational use of cluster resources, thereby improving task performance and cluster resource utilization.

When allocating resources for tasks in a decentralized system, different allocation methods may have an impact on the execution efficiency of tasks. For example, in some cases, frequent communication is required between multiple subtasks in a task, therefore, multiple subtasks in this task are suitable to be assigned to nodes with lower communication costs; in other cases, a task The multiple subtasks in will generate a relatively large amount of calculation during the execution process. Therefore, multiple subtasks in this task are suitable for allocation to nodes with more computing resources.

Moreover, most of the tasks in the current distributed system have the following characteristics: a task often includes multiple subtasks, the content and required resources of each subtask may be completely the same or different, and frequent subtasks may need to be performed. communication. In addition, in some cases, multiple subtasks in a task need to be scheduled for execution together, otherwise the task cannot be executed normally; in other cases, the execution order of each subtask will have a certain impact on the execution of the entire task , for example, a slave (slave) subtask needs to be executed after its subordinate master (master) subtask is executed.

There are currently some schedulers (for example, kube-batch) that batch tasks in a decentralized system according to the following: the scheduler tries to batch a task in order Each subtask in the task is scheduled one by one, and after the scheduling attempt of each subtask of the task is completed, it is judged whether the entire task has met the user's expected runnable state. This scheduling strategy is called gang scheduling (group scheduling).

However, the current implementation of gang scheduling essentially schedules at the granularity of subtasks, and the scheduler does not obtain the resource requirements of the entire task. As a result, the nodes allocated for subtasks may not be the best choice, making task operation efficiency and The resource utilization of the cluster is relatively low. Therefore, when cluster resources are tight or task types are complex, a large number of scheduling failures will occur.

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

Step 201: Obtain task information, where the task information includes at least one of task type and resource requirement information of the task.

Step 202: Assign target nodes to multiple subtasks of the task according to information of the task.

The method in the embodiments of the present disclosure may be executed by any electronic device, such as a terminal device or a cloud server, etc., and in some embodiments, the method may be executed by a processor or a scheduler, wherein, optionally, the processor Or the scheduler may run on a cloud platform such as Kubernetes, and specifically may be set on a container orchestration engine, but the embodiments of the present disclosure are not limited thereto. In step 201, information of one or more tasks can be obtained.

The multiple subtasks included in the task may be all subtasks of the task, or some subtasks in the task, such as subtasks with large resource requirements Tasks, or subtasks with higher priority, or subtasks that have dependencies with each other, subtasks that have similar types or quantities of resource requirements, and so on. Among them, subtasks with mutual dependencies refer to multiple subtasks with mutual dependence or master-slave relationship. For example, the execution of subtask A can only start after the execution of subtask B is completed. For example, subtasks with the same type or with Multiple subtasks of at least one specific type, etc. The task may be a neural network training task, an inference task, or other types of deep learning tasks. In some optional embodiments, each subtask in the same task may be performed by the same or different nodes in the same cluster in the distributed system. Where the decentralized system includes multiple clusters, individual subtasks of the same task may be performed by nodes of the same or different clusters. Optionally, the task may be assigned to a certain cluster first, and then the target nodes for multiple subtasks of the task are determined from the cluster. In this way, by allocating multiple subtasks of a task to the same cluster for execution, it is beneficial to reduce communication overhead and improve task execution efficiency.

The task information may be obtained based on user input or settings, or based on task analysis, and so on. The task information may include task type and/or resource requirement information of the task. Wherein, the resource requirement information of the task is specifically the resource requirement information of multiple subtasks included in the task. In some embodiments, in addition to the task type and resource requirement information, the task information may also include the priority information of the task, the priority information of multiple subtasks included in the task, and the dependencies between subtasks One or more of information, user provided information, historical information, calculation amount information, communication amount information, etc.

Wherein, the resource requirement information of the subtask may include executing the subtask The resource type and/or resource quantity required by the service may include but not limited to at least one of resources such as CPU, GPU, DSP, FPGA, random access memory, disk, and network interface.

The priority information of the task may include, but not limited to, the waiting time of the task to be scheduled and/or the total resources required by the plurality of subtasks in the task. The waiting time to be scheduled refers to the time between the time when the scheduler receives a task and the time when the task is scheduled, and the task with a longer waiting time to be scheduled and the total resources required can be more The priority of the task is set higher, so as to prevent a task from waiting for scheduling for a long time.

The priority information of the multiple subtasks included in the task is used to determine the priorities of the multiple subtasks included in the task. In some embodiments, in a scheduling cycle or a resource allocation process, the subtasks can be scheduled at least based on the priority information of each subtask.

The dependency information between the subtasks is used to determine the dependency or master-slave relationship between the multiple subtasks included in the task. If subtask B can only start to execute after the execution of subtask A is completed, then subtask B depends on Subtask A, or subtask A is the master task, and subtask B is the slave task of subtask A.

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

The historical information may include historical scheduling information and/or historical execution conditions and the like. The task type of the task in the historical scheduling process can be determined according to the historical information.

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

In some embodiments, in step 202, assigning target nodes to multiple subtasks included in the task respectively according to the information of the task includes: the currently available resources in the distributed system satisfy the multiple subtasks of the task In the case of the total resource corresponding to the resource requirement information of the task, corresponding target nodes are assigned to the multiple subtasks of the task respectively. The total resources corresponding to the resource requirement information of multiple subtasks of the task may refer to the total resources required by the multiple subtasks of the task. Optionally, when the multiple subtasks are all subtasks of the task, the total resource Resource is also the total resource required for the task in question. Optionally, if the currently available resources of the distributed system meet the multiple subtasks or the total resources required by the task, it indicates that the currently available resources in the decentralized system are sufficient to support the running of multiple subtasks. Optionally, if the currently available resources of the decentralized system do not meet the total resources required by multiple subtasks or tasks, it indicates that the currently available resources in the decentralized system can only support the operation of a part of the multiple subtasks at most, and cannot Supports running of all subtasks.

In the embodiment of the present disclosure, only when the currently available resources in the distributed system satisfy the total resources required by the multiple subtasks of the task, the corresponding target nodes are respectively assigned to the multiple subtasks of the task. In this way, the task can be used as the unit of resource allocation, and the multiple subtasks contained in the task can be allocated to nodes together, so as to avoid the failure of some subtasks in the task to be successful when the current available resources in the decentralized system are not enough Allocating resources, resulting in the fact that only some subtasks in the task occupy the entire resources in the distributed system, which is conducive to improving resource allocation. Reasonable use.

In some embodiments, target nodes can be assigned to multiple subtasks based on current resource status information of the distributed system and resource requirement information of the multiple subtasks. Wherein, each subtask can be assigned a target node that can meet its resource requirements, and the target nodes of different subtasks can be the same or different.

In some embodiments, based on the current resource state information of the distributed system and the resource requirement information of the multiple subtasks, target nodes for the multiple subtasks can be determined, and each of the multiple subtasks can be successfully When subtasks determine corresponding target nodes, assign corresponding target nodes to multiple subtasks of the task. The above-mentioned process of determining target nodes for the multiple subtasks is a simulated allocation process. The simulated allocation is a virtual resource allocation process, not an actual resource allocation process. The simulated allocation process is used to determine whether multiple subtasks are currently Can be assigned to nodes in a decentralized system. By performing simulated allocation of nodes for multiple subtasks of the task, it can be determined whether the distributed system currently meets the resources required by the multiple subtasks.

Wherein, the current resource state information of the distributed system is used to characterize the resources currently available in the distributed system, which may include at least one of the quantity, type and distribution of the currently available resources of the distributed system, or may include The current status of multiple nodes in the node, such as whether or not there are resources available, the type and quantity of available resources, and so on. The resource requirement information of multiple subtasks may include resource requirement information of each subtask in the multiple subtasks, such as at least one of required resource type and resource quantity of a certain resource type, such as required GPU, DSP, CPU, etc. One or more, the embodiments of the present disclosure are not limited thereto.

In some embodiments, based on the resource requirement information of multiple subtasks and the current resource status information of the distributed system, it can be determined whether there is a target node that meets the requirements of multiple subtasks in the distributed system, or, multiple subtasks can be sequentially determined in a specific order. The target node for each subtask in the task. For example, if each of the multiple subtasks can find a target node in the decentralized system that meets its needs, it can be determined that the currently available resources of the decentralized system satisfy the total resources required by the multiple subtasks. For another example, if at least one subtask in the multiple subtasks cannot find a target node that meets its requirements in the decentralized system, it may be determined that the currently available resources of the decentralized system do not meet the total resources required by the multiple subtasks.

Different from the traditional single subtask node determination, in the embodiment of the present disclosure, the target nodes of multiple subtasks are sequentially determined in a specific order, and the determination of the target nodes of the subsequent subtasks may be influenced by the previous subtasks. The impact of the target node determined by the task. In some embodiments, after the corresponding target node is determined for one of the subtasks each time, the current resource state information of the distributed system is updated, for example, the current state information of the determined target node is updated, and then based on the updated current The resource status information determines the target node corresponding to the next subtask, so as to avoid resource conflicts.

It should be noted that the update of the current resource status information mentioned here is used for the above simulation allocation or determination process of the target node. It is a virtual rather than a real update. After the allocation or determination of the target node, the virtual change of the currently available resources in the decentralized system, due to The actual allocation of subtasks has not yet taken place, so the currently available resources in the decentralized system have not actually changed. For example, after determining the target node of a subtask, it can be considered that the target node is virtually assigned to the subtask, and some resources on the target node corresponding to the resources required by the subtask are virtually bound to the subtask , so that the virtually bound resource can no longer be assigned to other subtasks.

The above specific order may refer to a preset order. Or, in some examples, the specific order can be determined according to the priorities of the plurality of subtasks, for example, first determine the corresponding target nodes for subtasks with higher priority, and then determine the corresponding target nodes for subtasks with lower priority Determine the corresponding target node, where the priority can be determined based on one or more factors such as the subtask type, the dependency relationship between the subtask and other subtasks, for example, for two subtasks that have a dependency or a master-slave relationship, The priority of the master (master) subtask is higher than that of the slave (slave) subtask, but the embodiments of the present disclosure are not limited thereto. Or, in some other examples, the specific order may be determined according to the dependency or master-slave relationship of the subtasks. Since the slave subtask cannot start to run without the master subtask, it is beneficial to improve the operation efficiency of the entire task by assigning the corresponding target node to the master subtask first. Alternatively, the specific order may also be determined based on other factors, which is not limited in this embodiment of the present disclosure.

In some embodiments, if there is a target node in the distributed system that satisfies the resources required by each subtask in the plurality of subtasks, then it is determined that the currently available resources in the distributed system satisfy the plurality of subtasks of the task The total resources corresponding to the resource requirement information of the task; otherwise, it is determined that the currently available resources in the distributed system do not satisfy the total resources corresponding to the resource requirement information of the plurality of subtasks of the task.

As shown in Figure 3A and 3B, assuming a task includes subtask 1, subtask 2, and subtask 3 with priority from high to low, it can be based on the current available resource information of the distributed system and the resource demand information of subtask 1 , to simulate the distribution of nodes for subtask 1, obtain the target node of subtask 1, and then update the current resource status information in the distributed system; then according to the updated distributed system’s current available resource information and resource demand information of subtask 2, Simulate the allocation of nodes for subtask 2, obtain the target node of subtask 2, and then update the current resource status information in the distributed system; then, according to the updated distributed system's current available resource information and resource demand information of subtask 3, Simulate the distribution of nodes for subtask 3, and get the target node of subtask 3.

If the simulated assignment of nodes for subtask 1, subtask 2, and subtask 3 is successful, then it is determined that the currently available resources of the distributed system meet the requirements of each subtask in the plurality of subtasks, that is, satisfy the requirements of a plurality of subtasks total resources. If the simulated allocation of nodes for subtask 1, subtask 2, or subtask 3 fails, for example, no target node that satisfies its required resources can be found in the distributed system, then it is determined that the currently available resources of the distributed system do not meet the required The total resources required by the multiple subtasks described above.

When it is determined based on the simulated allocation that the currently available resources of the distributed system meet the resources required by each of the multiple subtasks, the target corresponding to each subtask determined in the simulated allocation may be A node is assigned to each of the subtasks.

In some embodiments, target nodes can be determined for each subtask based on the current state information of each node in the decentralized system, for example, based on each The current status information of the nodes, sort or score each node, and assign nodes with higher ranking or higher scores to multiple subtasks in a specific order.

In some embodiments, when determining the target nodes for the multiple subtasks of the task, the target nodes respectively corresponding to the multiple subtasks may be determined through two processes of preselection and optimization.

Among them, in the pre-selection process, based on the resource demand information of multiple sub-tasks, at least one pre-selected node in the distributed system that can meet its needs is initially selected for each sub-task in the multiple sub-tasks, or eliminated in the distributed system. the node it needs. For example, based on the current state information of multiple nodes in the distributed system and the resource requirement information of the multiple subtasks, a preselected node set for each subtask in the multiple subtasks may be determined, wherein each subtask's The set of preselected nodes includes at least one preselected node. The preselected node sets of different subtasks may include the same or different nodes.

In the optimization process, the target node of each subtask is selected from the preselected node set of each subtask. For example, a most suitable node may be selected from the pre-selected node set of the subtask as the target node of the subtask. In some optional examples, it can be based on the current status information of each pre-selected node included in the pre-selected node set of the subtask, such as the current load situation, available resource type, available resource quantity, available resource distribution, and topological connections between nodes The target node is determined by at least one of a relationship and a topological connection relationship with other nodes. The available resource types include but are not limited to at least any of the following: communication interface, disk storage space, random access memory, CPU, GPU, and the like. The number of available resources may be the number of CPUs, Disk remaining capacity, random access memory remaining capacity, etc. The distribution of available resources is also referred to as resource fragmentation, that is, the distribution position of currently available resources, for example, how many disks the remaining disk capacity is distributed on. The topology of each node in the distributed system may be, for example, a bus topology, a star topology, a ring topology, a tree topology, and the like.

In some embodiments, in the preselection process, the preselected node set for each subtask can be determined based on the current resource state information of the distributed system and the resource requirement information of each subtask, that is, the multiplicity of the task The preselection process of each subtask in the subtasks may not interfere with each other, and the selection of the preselected node set of each subtask is only related to the resource requirement information of the subtask, and does not depend on the resource requirement information of other subtasks. Correspondingly, the determination of the preselected node sets of multiple subtasks can be performed in parallel or in any sequence. In this way, the determination of the pre-selected nodes can be performed quickly, thereby improving the overall efficiency of the simulation assignment.

For example, assuming that the task includes subtask 1, subtask 2 and subtask 3 with priority from high to low, then in the preselection process, the following operations can be performed in parallel: according to the resource requirement information of subtask 1 and the current distributed system The resource status information determines the preselected node set of subtask 1, determines the preselected node set of subtask 2 according to the resource requirement information of subtask 2 and the current resource status information of the distributed system, and determines the preselected node set of subtask 2 according to the resource requirement information of subtask 3 and the distributed system The current resource state information of the system determines the set of preselected nodes for subtask 3, assuming that the set of preselected nodes corresponding to subtask 1 includes {node 1, node 2, node 3}, and the set of preselected nodes corresponding to subtask 2 includes {node 2, Node 5, Node 6, Node 7}, the preselected node set corresponding to subtask 3 includes {Node 6 and Node 7}.

In some embodiments, during the preselection process, the preselected node sets of multiple subtasks may be sequentially selected in a specific order. In an optional example, the preselected node set of the following subtask is determined based on the preselected node set of the previous subtask. For example, the preselected nodes of the following subtasks are selected from the set of preselected nodes of the preceding subtasks, wherein the number of the preceding subtasks can be one or more than two, for example, the preceding subtasks can be the All subtasks, the preceding subtask and the following subtask may be adjacent to each other or separated by at least one subtask, but the embodiments of the present disclosure are not limited thereto.

In some embodiments, the resource requirement information of multiple subtasks can also be considered comprehensively to determine the preselected node sets for multiple subtasks, for example, select a common preselected node set for the multiple subtasks, wherein the shared The preselected nodes included in the preselected node set can meet the resource requirements of at least two subtasks in the plurality of subtasks.

In some embodiments, in the optimization process, the target nodes of multiple subtasks may be determined in parallel, or the target nodes of multiple subtasks may be comprehensively determined by combining the candidate node sets of multiple subtasks. For example, target nodes of at least a part of the subtasks are determined by intersecting the candidate node sets of at least a part of the subtasks among the multiple subtasks, but the embodiments of the present disclosure are not limited thereto.

In some embodiments, the target nodes corresponding to the multiple subtasks may be sequentially determined from the preselected node sets of the multiple subtasks in a specific order, for example, based on the priorities of the multiple subtasks. Wherein, in an optional example, each time a target node corresponding to a subtask is determined, the current state information of the target node is updated. example For example, based on the resource requirement information corresponding to the subtask, the current state information of the target node is updated. In this way, during the determination of the target node for the subsequent subtask, the part of the target node corresponding to the resource required by the subtask will not be available, thereby ensuring the smooth execution of multiple subtasks. 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, the following subtasks prefer to select the target node of the previous subtask as its own target node, unless the target node is not included in the candidate node set of the subsequent subtask or other key factors of the target node do not meet the requirements of the subsequent subtasks. requirements, but the embodiments of the present disclosure are not limited thereto.

In the case of synchronous resource allocation for multiple subtasks, the above preselection and optimization process may be a simulated allocation process. In the case of asynchronous resource allocation for multiple subtasks, the above preselection and optimization process may also be a real node allocation process. Similarly, the update of the current resource status information involved in the above process may also refer to the real update process performed due to changes in the currently available resources in the distributed system after the node allocation of a subtask is performed.

In the embodiment of the present disclosure, the target node may be selected from the pre-selected node set based on a certain strategy. For example, the target node is selected based on current resource status information of the distributed system or current status information of each pre-selected node and resource requirement information of multiple subtasks. In some embodiments, the score of the preselected nodes contained in the preselected node set of each subtask may be determined, for example, based on the current state information of each preselected node in the subtask preselected node set and the resource requirement information of the subtask, Determine the score of each pre-selected node, and based on the pre-selected node set of each subtask Select the target node of each subtask from the preselected node set of each subtask. For example, the preselected node with the highest score may be selected from at least one preselected node included in the preselected node set as the target node. Alternatively, target node selection is performed based on scores and other factors.

Following the previous example, take determining the target nodes of multiple subtasks sequentially based on task priority as an example. First, you can determine the scores of each preselected node in the preselected node set of subtask 1. Assume that the scores of nodes 1, 2 and 3 are The values are 80, 90 and 70 respectively, then it is determined that node 2 is the target node of subtask 1, and the current state information of node 2 is updated. Then, determine the score of each preselected node in the preselected node set of subtask 2, wherein the score of node 2 is determined based on the updated current state information of node 2, assuming node 2, node 5, node 6 and node The scores of 7 are 60, 80, 75 and 70 respectively, then determine node 5 as the target node of subtask 2, and update the current state information of node 5. Finally, determine the score of each preselected node in the preselected node set of subtask 3, assuming that the scores of nodes 6 and 7 are 70 and 60 respectively, then determine node 6 as the target node of subtask 3.

In some embodiments, the task type of the task may be determined based on the information of the task acquired in step 201, and target nodes may be assigned to multiple subtasks based on the task type of the task. Wherein, the task types of tasks may be divided based on actual conditions. For deep learning tasks such as training and reasoning of deep learning models such as neural networks, tasks can optionally be divided into communication-intensive tasks and computation-intensive tasks. Among them, the communication-intensive task refers to the task with frequent communication between sub-tasks in the process of task processing, and the calculation-intensive task refers to the task that has been processed Tasks with a relatively large amount of computation in the process.

Specifically, the task type of the task may be determined based on the overall situation of multiple subtasks of the task. As an optional example, the task type of the task may be determined based on the calculation amount and communication amount of multiple subtasks of the task. For example, when the calculation amount is greater than the preset calculation amount threshold, the task is considered to be a calculation-intensive task; for another example, when the communication amount is greater than the preset communication amount threshold, the task is considered to be a communication-intensive task . For another example, if both the calculation amount and the communication amount exceed corresponding preset thresholds, it may be determined as a communication-intensive task. As another optional example, the task type may also be determined according to historical scheduling experience. For a specific task, if the scheduler determines that the task is a computation-intensive task in the historical scheduling process, it determines that the task is a computational-intensive task; if the scheduler determines that the task is a communication-intensive task in the historical scheduling process type task, it is determined that the task is a communication-intensive task. Optionally, the task type of the task may also be determined based on at least one historical execution of the task, for example, if in the latest one or more historical executions, the time occupied by the calculation of the task or more computing resources, determine the task as a computing-intensive task; for another example, if the communication of the task takes up more time or communication resources during one or more recent historical executions, then determine the task as The task was determined to be communication intensive. Alternatively, the task type of the task may also be determined based on other methods. Or, the task type of the task is determined based on the information provided by the user, for example, the task type information provided by the user, or the calculation amount and/or communication amount information of the task provided by the user, and so on. The embodiment of the present disclosure does not limit the specific implementation of determining the task type.

In the embodiment of the present disclosure, the target nodes of multiple subtasks may be determined based on the task type, so that the task as a whole can be executed more efficiently. For example, for a communication-intensive task, multiple subtasks in the task can be allocated to a small number of target nodes as much as possible, thereby reducing communication overhead; for example, for a calculation-intensive task, only the target node The resources on the network can meet the requirements of multiple subtasks of the task, or choose nodes with better computing performance first, or give priority to the use of fragmented resources of nodes, or try to optimize communication overhead on the premise of ensuring that computing resources meet user needs ,wait.

In some optional examples, the above preselection and/or optimization process may be performed according to the task type of the task.

In some embodiments, the above preselection process is performed based on the task type of the task. For example, for a communication-intensive task, it is obviously impossible to satisfy more subtasks in multiple subtasks, such as the available resource type and/or the small number of available resources. The node elimination required by the task, for example, the number of nodes contained in the pre-selected node set is controlled within a certain range, among which, the nodes with more resource types and/or numbers can be preferentially selected, or the overall resource type and multiple subtasks Nodes with matching resource requirements are used as pre-selected nodes. In some optional examples, a preselected node set may also be determined for part or all of the subtasks of the communication-intensive task, instead of determining a separate preselected node set for each subtask, but the embodiments of the present disclosure do not limited to this.

In other embodiments, the above preferred process is performed based on the task type. For example, based on the task type, a strategy for selecting a target node is determined. For example, in the process of determining the target node of a subtask, it is preferred to determine target node. For example, a task includes subtask 1 and subtask 2, and the target nodes of subtask 1 include node 1 and node 2, then node 1 and node 2 may preferably be selected as target nodes for subtask 2. In some optional examples, the score of each preselected node included in the subtask preselected node set may be determined according to the current state information of each node included in the subtask preselected node set and the task type of the task. In some other embodiments, the score of each preselected node included in the subtask preselected node set may also be determined according to other methods, which will not be repeated here.

In one example, for communication-intensive tasks, since it is expected to reduce the number of target nodes as much as possible to reduce the communication cost between sub-tasks, therefore, for the sub-tasks of communication-intensive tasks, the scores of nodes in the resource distribution concentration are generally high Nodes with scattered resource distribution. For example, the available disk capacity of the preselected node 1 of a subtask is distributed on two disks, and the available disk capacity of the preselected node 2 is distributed on one disk. Therefore, the preselected node 2 of this subtask is A score of 1 is high. In another example, for computationally intensive tasks, nodes with more computational power generally have higher scores than nodes with less computational power. For example, if the processor of the preselected node 1 of a subtask includes 1 processor core, and the processor of the preselected node 2 of the subtask includes 2 processor cores, then the score of the preselected node 2 of the subtask is higher than that of the preselected node 1. high value.

The above process is described by taking subtasks as resource allocation objects as an example. In some embodiments, multiple subtasks of the task can also be divided into at least one grouping, wherein each grouping includes at least one subtask in the multiple subtasks; Each group is assigned a corresponding target node, Wherein, each subtask in the same group is assigned to the same target node.

The number of groups may be the same as the number of nodes required by the task, that is, when multiple subtasks of the task need to be allocated to N target nodes, the multiple subtasks of the task are divided into N groups. Wherein, the number of subtasks in each group may be the same or different. For example, randomly assign multiple subtasks, or divide subtasks with similar priorities into a group, or divide subtasks with or without dependencies into a group, or divide subtasks with large differences in resource requirements Divide into one group, or divide subtasks with the same type of required resources into one group, etc., and the embodiment of the present disclosure does not limit the specific grouping manner. In this way, the determination of the target node in units of groups can improve the efficiency of the resource allocation process without significantly affecting the efficiency of task execution.

In some embodiments, the multiple subtasks of the task may be divided into at least one group according to the task type of the task. For example, for a communication-intensive task, multiple subtasks of the task can be allocated to the least number of target nodes, thereby reducing the communication cost between subtasks, for example, dividing all subtasks in multiple subtasks into the same Group so that all subtasks execute on the same target node. In some optional examples, the total number of nodes required by the multiple subtasks of the task may be determined according to the resource requirement information of the multiple subtasks of the task; according to the total number of nodes, the multiple subtasks of the task Tasks are divided into at least one group. For another example, for a computation-intensive task, each of the multiple subtasks of the task may be regarded as a group, that is, the target node is determined in units of subtasks.

For example, assuming that a task includes 4 subtasks, each subtask requires a GPU, and the number of available GPUs per node in the distributed system is 2, then at least 2 nodes are required to execute the 4 subtasks in the task. Therefore, if the task is a communication-intensive task, the 4 subtasks can be equally divided into 2 groups, wherein the grouping can be performed randomly, or according to the resource requirement information of the subtasks, and so on. If the task is a computationally intensive task, the 4 subtasks can be divided into 4 groups, and each subtask is regarded as a group.

In the case of grouping, optionally, the above determination of the target node for the subtasks may also be applicable to the determination of the target node for the grouping, which will be described in detail below.

For example, in some embodiments, in the case that each group in at least one group of the task has a corresponding target node, it is determined that the currently available resources in the decentralized system meet the requirements of multiple subtasks of the task of total resources. In some other embodiments, in the case that at least one group of the task does not have a corresponding target node, it is determined that currently available resources in the distributed system do not meet the total resources required by the multiple subtasks of the task.

For another example, based on the current resource status information of the distributed system and the resource requirement information of each group in the at least one group, the target node can be determined for each group in the at least one group in a specific order . The resource requirement information of a group may be determined based on the resource requirement information of at least one subtask included in the group, for example, the resource required by the group may be the sum of resources of the same type required by all the subtasks included in the group. Wherein, the specific order is optional The location may be determined according to at least one of the priorities and dependencies of the multiple groups or other factors. The priority of the grouping may be determined based on the priorities of two or more subtasks included in the grouping, for example, determining the priority of the grouping as the highest priority of the two or more subtasks, or determining the priority of the grouping as the two or more subtasks The average priority of the above subtasks, etc. The dependencies between groups can be determined based on the dependencies between subtasks included in different groups. For example, assuming that one subtask in group 1 depends on one subtask in group 2, then it is determined that group 1 depends on group 2, Etc., the embodiments of the present disclosure are not limited thereto.

In some embodiments, based on the current resource state information of the distributed system and the resource demand information of each group, the preselected node set of each group can be determined, and the preselected node set of each group can be sequentially selected according to a certain order. The target node of the group is determined in the node set.

In the embodiment of the present disclosure, first determine the target node for the entire task, for example, perform the above preselection and optimization operations, and finally obtain the information of the target node corresponding to the entire task, and then split the task into subtask levels to do a binding , that is, to bind the target node in units of subtasks.

In some embodiments, when the currently available resources in the decentralized system do not meet the total resources required by the plurality of subtasks of the task, or when determining the corresponding target for at least one subtask of the plurality of subtasks In the case that the node is unsuccessful, delay allocation may be performed on multiple subtasks included in the task. The delayed allocation may be to allocate multiple subtasks included in the task in the next allocation period. In the next allocation cycle, the currently available resources in the decentralized system remain Only when the source satisfies the total resources required by the multiple subtasks of the task can the corresponding target nodes be assigned to the multiple subtasks of the task. If in the next allocation cycle, the currently available resources in the distributed system still do not satisfy the total resources required by the multiple subtasks of the task, continue to perform delayed allocation.

In some embodiments, after each target node corresponding to the current subtask or group is determined, a virtual assignment between the current subtask or group and the corresponding target node can be performed, or a virtual assignment between the current subtask or the current group and the corresponding target node can be performed. Virtual binding of resources. However, when it is determined that the current available resources in the distributed system do not meet the total resources required by the multiple subtasks of the task, for example, when a corresponding target node cannot be found for a certain subtask, the above virtual An allocation or virtual binding is considered invalid, so that its corresponding target node and its resources can be allocated to other tasks.

In some embodiments, after assigning corresponding target nodes to the multiple subtasks of the task, the multiple subtasks of the task may be scheduled synchronously.

It should be noted that, in the embodiments of the present disclosure, allocation refers to distributing subtasks to corresponding target nodes, and scheduling refers to executing subtasks by corresponding target nodes after subtasks are distributed to corresponding target nodes.

As shown in Figure 4, in the traditional scheduling method, in order to realize gang scheduling, the scheduler will process multiple subtasks in the task in order, and when a subtask can be scheduled, the subtask will be actually allocated immediately (not referring to The simulated allocation described in the embodiment of the present disclosure) is assigned to the corresponding target node, and the resources on the target node are occupied The source schedules the subtask. For a task, the currently available resources of each subtask in the task are dynamically changing and gradually decreasing when allocated. This scheduling method is called task by task scheduling. Moreover, when scheduling subtasks, the scheduling situation and resource requirements of multiple subtasks in the same task are not considered as a whole, resulting in each assignment being an independent behavior. However, in actual situations, there are often some associations between multiple subtasks in the same task. For example, some tasks to be scheduled are often accompanied by a large amount of communication requirements. If some subtasks of a task are assigned to the communication overhead On multiple nodes, it will significantly affect the operating efficiency of the entire task.

In the embodiment of the present disclosure, subtask allocation and scheduling are performed at the granularity of tasks. As shown in Figure 5, firstly, all tasks are obtained according to the priority. For each task, the entire task is considered as a whole. Only when the current available resources in the distributed system meet the total resources required by the multiple subtasks of the task, The corresponding target nodes are assigned to the multiple subtasks of the task, and then the multiple subtasks are scheduled synchronously, that is, resource allocation and scheduling are performed in units of tasks.

Different from the previous task-by-task scheduling method according to the requirements of sub-tasks, the embodiment of the present disclosure considers the entire task when selecting nodes for the task. By using the entire task as the smallest unit of scheduling, it is possible to make better use of each node in the task. Based on the interrelationship and demand of subtasks, a more suitable scheduling node is selected from the cluster for the entire task, thereby improving the operation efficiency of the task and the resource utilization efficiency of the cluster.

As shown in FIG. 6 , it is a schematic diagram of the scheduling logic of some embodiments of the present disclosure. In order to reduce the scheduling complexity, the following assumptions can be made for each task to be scheduled: (1) Each task has multiple subtasks; (2) Each subtask is homogeneous in terms of resource requirements, that is, the resource requirements are the same; (3) Each task submitted by each user specifies each subtask when submitting The demand for resources, and the number of subtasks and the resource demand of each subtask remain unchanged during the entire task processing process.

Firstly, a pre-selection process is performed to eliminate nodes that cannot meet the user's tasks. Specifically, it includes the following steps:

(1) For each subtask of the task, based on the current resource status information of the distributed system and the resource demand information of each subtask, the preselected node set of the subtask can be determined respectively, and the preselected node set of the subtask The node must be unable to support the subtask to run on it. Here, a task may be called a job, and a subtask may be called a task. The pre-selected node sets of different subtasks may be the same or different.

In some optional examples, the preselected nodes of the following subtasks can be determined from the set of preselected nodes of the previous subtasks, wherein the order of the subtasks can optionally be determined based on the priority of the subtasks. For example, assuming that the distributed system includes nodes 1 to 5, and a task includes subtask 1, subtask 2 and subtask 3, you can first determine the preselected nodes for subtask 1, assuming node 1, node 2 and node 3; Then, determine the preselected nodes of subtask 2 from nodes 1, 2 and 3, assuming node 2 and node 3; then determine the preselected nodes of subtask 3 from nodes 2 and 3, assuming node 3. Through this way of determining the set of pre-selected nodes, for a certain amount Tasks with high communication requirements, especially communication-intensive tasks, can improve the efficiency of determining the set of pre-selected nodes, thereby improving the efficiency of task allocation.

(2) Verify each preselected node in the preselected node set, and determine whether the total resources on the preselected node meet the total resource requirements of all subtasks allocated to the preselected node. If yes, perform the optimization process, otherwise, perform the preselection process again.

Wherein, this step is optional, and the verification of the pre-selected nodes can improve the reliability of the entire task assignment. Before performing the pre-selection process, the node set division can also be performed first, that is, each node is grouped, and the basis of the grouping can be the cluster to which the node belongs, that is, the nodes on different clusters are divided into different groups. By splitting the node set, different tasks can be assigned only to specific grouped nodes.

Then, an optimization process is performed to select the target node most suitable for the user's task. Specifically, the following steps may be included:

(1) Obtain the calculation amount and communication amount of the task during the training process through the information provided by the user in the user task and the historical experience of the scheduler, and judge whether the user's task is calculation-intensive or communication-intensive according to the calculation amount and communication amount .

In some optional examples, the determination of the task type may also be performed during or before the above pre-selection process, and the pre-selected node set of the sub-task is determined based on the task type.

In some optional examples, based on the task type, multiple subtask target nodes in the optimization process are determined. If it is a communication-intensive task, place (i.e. assign) the task to the minimum number of physical nodes (i.e. target nodes). The number of assigned physical nodes is determined by the available resources of the nodes and the resources required by the task. , If this placement constraint cannot be satisfied, allocation is delayed for multiple subtasks in the task. If it is a computing-intensive task, relax this placement constraint. Even if the minimum physical nodes cannot be satisfied, as long as a node with sufficient resources can be selected to meet the task requirements, the allocation will be performed; otherwise, the allocation will be delayed.

(2) Score and sort the cluster nodes according to the free resources from large to small, and get the score of the mapping from the pre-selected node set to the task, where the score of the pre-selected nodes can be based on the node load and node resources in the distributed system cluster Debris to determine. Node load refers to the current total amount of available resources and resource usage of each node in the cluster at the current moment. Resource fragmentation refers to the distribution of resources on nodes.

(3) For all the nodes selected in the previous step, the tasks of the tasks are placed sequentially from high scores to low scores until they cannot be placed or all tasks of the task are placed.

(4) Task status check. If the task is still free and the task is not placed, it means that the cluster resources are not enough and the allocation fails. Then release the resources occupied by the task and delay the allocation. If the job has no free tasks, and the task type is communication-intensive, determine whether the minimum node requirements are met. If it is satisfied, the allocation is successful, if it is not satisfied, the allocation fails and the resource is released, and the allocation is delayed.

The method of the embodiment of the present disclosure can be applied to a task orchestration system or a cloud platform, and the task orchestration system can be implemented based on a platform such as Kubernetes.

Those skilled in the art can understand that in the above method of specific implementation, the writing order of each step does not mean a strict execution order and constitutes any limitation on the implementation process. The specific execution order of each step should be based on its function and possible The inner logic is OK.

As shown in FIG. 7 , the embodiment of the present disclosure also provides a task scheduling device, the device includes: an acquisition module 701, configured to acquire task information, the task information includes the task type and resource requirement information of the task at least one of; the first assignment module 702, configured to assign target nodes to multiple subtasks included in the task according to the information of the task.

In some embodiments, the resource requirement information of the task is the resource requirement information of a plurality of subtasks included in the task; the first allocation module 702 also satisfies the multiplicity of the task with currently available resources in the distributed system In the case of the total resources corresponding to the resource requirement information of subtasks, the corresponding target nodes are assigned to the multiple subtasks of the task respectively.

In some embodiments, the device further includes: a first determining module, configured to determine targets for the multiple subtasks based on the current resource state information of the distributed system and the resource requirement information of the multiple subtasks Node; the first assignment module is used for: in the case of successfully determining the corresponding target node for each subtask of the plurality of subtasks, respectively assigning corresponding target nodes to the plurality of subtasks of the task.

In some embodiments, the device further includes: a first determining module, configured to determine targets for the multiple subtasks based on the current resource state information of the distributed system and the resource requirement information of the multiple subtasks node; the first allocation module is used for: in the case of successfully determining the corresponding target node for each subtask in the multiple subtasks, respectively assigning the corresponding target node to the multiple subtasks of the task point.

In some embodiments, the first determining module is configured to: sequentially determine the target nodes corresponding to multiple subtasks in a specific order based on the current resource status information of the distributed system and the resource requirement information of the multiple subtasks .

In an optional example, the resource requirement information of the plurality of subtasks includes resource requirement information of each subtask in the plurality of subtasks, wherein the resource requirement information may include required resource types and resources of each type. The number of resources, or further include other information.

In an optional example, the current resource status information of the distributed system may include current status information of multiple nodes in the distributed system, where the current status information is used to indicate whether resources are available, the type of available resources and At least one of quantity, load condition, topological connection information, or further include other information.

In some embodiments, the device further includes: an update module, configured to update the current resource state information of the distributed system after each determination of the target node assigned to a subtask, and based on the updated current resource state The information determines the target node corresponding to the next subtask.

In some embodiments, the sequentially determined order of the target nodes of the plurality of subtasks is obtained according to at least one of priorities of subtasks and dependencies between subtasks.

In some embodiments, the first determining module includes: a first determining unit, configured to, based on the current resource state information of the distributed system and the resource requirement information of the plurality of subtasks, sequentially assign the plurality of subtasks The task determines the target node; the update order The element is used for updating the current resource status information of the distributed system based on the resource requirement information of the subtask after each target node of a subtask is determined.

In some embodiments, the apparatus further includes: a second determination module, configured to determine the distributed The currently available resources of the system do not satisfy the resources required by each subtask in the plurality of subtasks.

In some embodiments, the first determining unit includes: a determining subunit, configured to determine each of the multiple subtasks based on the current resource status information of the distributed system and the resource requirement information of the multiple subtasks. A preselected node set of subtasks; a selection subunit, configured to select the target node of each subtask from the preselected node set of each subtask in the plurality of subtasks.

Wherein, the preselected node sets of different subtasks may be the same or different. In some examples, multiple subtasks may have the same set of preselected nodes.

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

The determination of the pre-selected node sets of multiple subtasks may be independent without interdependence, for example, may be executed in parallel or in any sequence. For example, the preselected node sets of multiple subtasks are determined based on the same current resource state information of the distributed system, that is, the current resource state information of the distributed system is not updated during the preselection process. For example, the determination of the pre-selected node set of each subtask is only related to its own resource demand information, and not related to the resource requirements of other subtasks. Demand information is irrelevant.

In some embodiments, the determining subunit is configured to: sequentially determine the preselected node set of each of the multiple subtasks in a specific order, wherein the preselected node set of the first subtask of the multiple subtasks The preselected nodes are selected from the preselected node set of the second subtask among the plurality of subtasks, wherein the order of the second subtask is before the first subtask.

In some embodiments, the target node of each subtask is sequentially selected from the preselected nodes of each subtask in the plurality of subtasks according to a specific order, wherein the target node of the second subtask among the plurality of subtasks is preferentially selected The node serves as the target node of the first subtask among the plurality of subtasks, wherein the order of the second subtask is before the first subtask.

In some embodiments, the selection subunit is configured to: determine the first subtask from the pre-selected node set of the first subtask among the plurality of subtasks based on the current resource status information of the distributed system. The target node: based on the resource requirement information of the first subtask, updating the current resource state information of the distributed system, and based on the updated current resource state information of the distributed system, from the plurality of subtasks The target node of the second subtask is determined from the preselected node set of the second subtask.

In an optional example, based on the current state information of each preselected node in the first subtask's preselected node set, the target node of the first subtask is determined from the first subtask's preselected node set, and Based on the required resource information of the first subtask, current state information of the target node of the first subtask is updated.

In some embodiments, the selection subunit is configured to: determine the score of the preselected nodes contained in the preselected node set of each subtask according to the task type of the task; The score of the preselected node is to select the target node of each subtask from the preselected node set of each subtask.

In an optional example, based on the task type of the task, a score determination strategy for each preselected node in the preselected node set is determined.

In some embodiments, the apparatus further includes: a third determination module configured to determine a task type of the task; wherein, the target nodes corresponding to the plurality of subtasks are determined based on the task type of the task.

In some embodiments, the first assigning module includes: a grouping unit, configured 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 A task; an allocating unit, configured to allocate a corresponding target node to each of the at least one group, wherein each subtask in the same group is allocated to the same target node.

In some embodiments, the grouping unit is configured to: divide multiple subtasks 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 computation-intensive or communication-intensive.

In some embodiments, the grouping unit is configured to: when the task type of the task is communication-intensive, according to the resource requirement information of the multiple subtasks of the task, determine the resources required by the multiple subtasks. total number of nodes, and according to the total The number of nodes, dividing multiple subtasks of the task into at least one group.

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

In some embodiments, a preselected node set is respectively determined for each of the at least one group, and a target node of the group is sequentially selected from the preselected node set of each of the at least one group.

In some embodiments, the first allocation module further determines target nodes for multiple subtasks of the task based on the task type of the task; when the corresponding target nodes are successfully determined for multiple subtasks of the task In the case of , assign target nodes to multiple subtasks included in the task.

In some embodiments, the apparatus further includes: a delay module, configured to, in the case that the corresponding target node is determined to be unsuccessful for at least one subtask in the subtask, all the subtasks of the task Make a delayed allocation.

In some embodiments, when the currently available resources in the distributed system do not meet the total resources corresponding to the resource requirement information of the multiple subtasks of the task, the delay module will Delayed distribution is performed.

In some embodiments, the device further includes: a scheduling module, configured to perform synchronous scheduling on multiple subtasks of the task after assigning corresponding target nodes to the multiple subtasks of the task.

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

In some embodiments, the functions of the device provided by the embodiments of the present disclosure or the modules contained therein can be used to execute the methods described in the method embodiments above, which specifically For implementation, reference may be made to the description of the method embodiments above, and details are not repeated here for the sake of brevity.

The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, That is, it can be located in one place, or it can be distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution in this manual. It can be understood and implemented by those skilled in the art without creative effort.

Embodiments of the apparatus in this specification can be applied to computer equipment, such as servers or terminal equipment. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it reads the corresponding computer program instructions in the non-volatile memory into the random access memory through the processor where the file is processed, and then reads them from the random access memory. Read from memory to run in processor. From the perspective of hardware, as shown in Figure 8, it is a hardware structure diagram of the computer equipment where the device of this specification is located, except for the processor 801 shown in Figure 8, the random access memory 802, the network interface 803, and In addition to the non-volatile memory 804, the server or the electronic device where the device in the embodiment is located may also include other hardware generally according to the actual function of the computer device, so details will not be repeated here.

Correspondingly, an embodiment of the present disclosure also provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the method described in any embodiment is implemented.

Correspondingly, an embodiment of the present disclosure also provides a computer device, including a memory, a processor, and a computer program stored in the memory and operable on the processor. When the processor executes the program, any embodiment is realized. the method described.

The present disclosure may take the form of a computer program product embodied on one or more storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) having program code embodied therein. Computer usable storage media includes permanent and non-permanent, removable and non-removable media, and may be implemented by any method or technology for information storage. Information may be computer readable commands, data structures, modules of programs, or other data. Examples of storage media for computers include, but are not limited to: phase change random access 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 random access memory technology, compact disc read-only memory (CD-ROM ), digital versatile disc (DVD) or other optical storage, magnetic tape cartridge, magnetic tape storage or other magnetic storage device, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

Other embodiments of the present disclosure will be readily apparent to those skilled in the art from consideration of the specification and practice of the specification disclosed herein. The present disclosure is intended to cover any modification, use or adaptation of the present disclosure. These modifications, uses or adaptations follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not disclosed in the present disclosure. . The specification and examples are to be considered exemplary only, with the true scope and spirit of the disclosure indicated by the following claims.

It should be understood that the present disclosure is not limited to the precise constructions which have been described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

The above descriptions are only preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present disclosure shall be included in the present disclosure within the scope of protection.

201: Obtain task information, where the task information includes at least one of the task type and resource requirement information of the task

202: According to the information of the task, assign target nodes to multiple subtasks of the task respectively

Claims (21)

A task scheduling method, the method is applied to a computer device, the computer device includes a processor and a memory, the method includes: obtaining information of a task by the processor, the task includes a plurality of subtasks, the information Including at least one of the task type of the task and resource requirement information, the resource requirement information of the task is the resource requirement information of the plurality of subtasks included in the task; the currently available resources in the distributed system meet the required In the case of total resources corresponding to the resource requirement information of the multiple subtasks, the processor assigns corresponding target nodes to the multiple subtasks respectively. The task scheduling method according to claim 1, further comprising: determining, by the processor, for the plurality of subtasks based on the current resource state information of the distributed system and the resource requirement information of the plurality of subtasks Target node: when the currently available resources in the distributed system satisfy the total resources corresponding to the resource demand information of the multiple subtasks, the processor assigns corresponding targets to the multiple subtasks respectively The node includes: if the corresponding target node is successfully determined for each subtask in the multiple subtasks, assigning the corresponding target node to the multiple subtasks by the processor. The task scheduling method according to claim 2, wherein the processor is based on the current resource state information of the distributed system and the plurality of child The resource requirement information of the task, determining the target node for the plurality of subtasks, comprising: by the processor based on the current resource status information of the distributed system and the resource requirement information of the plurality of subtasks , determining target nodes for the plurality of subtasks in turn; after each target node of a subtask is determined, the processor updates the current resource of the distributed system based on the resource requirement information of the subtask status information. The task scheduling method according to claim 2 or 3, wherein, in the case of failing to determine the corresponding target node for at least one of the plurality of subtasks, the processor determines the distributed The currently available resources of the system do not meet the resources required by each subtask in the plurality of subtasks. The task scheduling method according to any one of claims 2 to 3, wherein the processor is based on the current resource status information of the distributed system and the resource requirement information of the plurality of subtasks, Determining target nodes for the plurality of subtasks includes: determining, by the processor, one of the plurality of subtasks based on the current resource state information of the distributed system and the resource requirement information of the plurality of subtasks A preselected node set for each subtask; the processor selects a target node for each subtask from the preselected node set for each subtask. The task scheduling method according to claim 5, wherein the determining by the processor the preselected node set of each subtask in the plurality of subtasks includes: sequentially determining the set of nodes for each subtask in a specific order by the processor The preselected node set of subtasks, wherein the preselected node in the preselected node set of the first subtask of the plurality of subtasks is selected from the preselected node set of the second subtask of the plurality of subtasks , wherein the sequence of the second subtask is before the first subtask. The task scheduling method according to claim 5, wherein the selecting the target node of each subtask by the processor from the preselected node set of each subtask includes: by the processor Selecting the target node of each subtask from the preselected node set of each subtask in a specific order, wherein the target node of the second subtask among the plurality of subtasks is preferentially selected as the plurality of subtasks The target node of the first subtask in , wherein the order of the second subtask is before the first subtask. The task scheduling method according to claim 5, wherein the selecting the target node of each subtask by the processor from the preselected node set of each subtask includes: by the processor According to the task type of the task, determine the score of the preselected nodes included in the preselected node set of each subtask; Select the target node of each subtask from the set of preselected nodes of each subtask. The task scheduling method according to claim 2 or 3, wherein the order in which the plurality of subtasks sequentially determine the target node is determined based on at least one of the priority of the subtasks and the dependency relationship between the subtasks . The task scheduling method according to any one of claim items 1 to 3, further comprising: determining the task type of the task; by the processor based on the task type of the task, respectively The target node corresponding to the task assignment. The task scheduling method according to any one of claim items 1 to 3, wherein the assigning corresponding target nodes to the plurality of subtasks by the processor includes: assigning the plurality of subtasks by the processor The task is divided into at least one group, wherein each group includes at least one subtask in the plurality of subtasks; the processor assigns a corresponding target node to each group in the at least one group, wherein the same Each subtask in the group is assigned to the same target node. The task scheduling method according to claim 11, wherein said dividing the plurality of subtasks into at least one group by the processor includes: dividing by the processor according to the task type of the task The plurality of subtasks are divided into at least one group. The task scheduling method according to claim 10, wherein the task type of the task is computation-intensive or communication-intensive. The task scheduling method according to claim 11, wherein said dividing said plurality of subtasks into at least one group by said processor includes: when said task type of said task is communication-intensive , the processor determines the total number of nodes required by the multiple subtasks according to the resource requirement information of the multiple subtasks, and divides the multiple subtasks into at least one according to the total number of nodes grouping; and/or in a case where the task type of the task is computationally intensive, the processor treats each of the plurality of subtasks as a group. The task scheduling method according to claim 10, wherein said assigning corresponding target nodes to said plurality of subtasks by said processor based on said task type of said task comprises: said processor Based on the task type of the task, determine target nodes for the plurality of subtasks; when the corresponding target nodes are successfully determined for the plurality of subtasks, allocate the plurality of subtasks by the processor target node. The task scheduling method according to claim 15, further comprising: in the case of unsuccessful determination of the corresponding target node for at least one of the subtasks, the processor performs Delayed distribution. The task scheduling method according to any one of claims 1 to 3, further comprising: the currently available resources in the distributed system do not satisfy the total corresponding to the resource requirement information of the multiple subtasks In the case of resources, the processor performs delayed allocation on the multiple subtasks. The task scheduling method according to any one of claims 1 to 3, further comprising: after assigning corresponding target nodes to the multiple subtasks, the processor performs synchronous scheduling on the multiple subtasks. A task scheduling device, comprising: an acquisition module for acquiring task information, the task includes a plurality of subtasks, the information includes at least one of the task type and resource requirement information of the task, and the resource of the task The requirement information is the resource requirement information of the plurality of subtasks included in the task; the first allocation module is used for the currently available resources in the distributed system to satisfy the total corresponding to the resource requirement information of the plurality of subtasks In the case of resources, corresponding target nodes are assigned to the multiple subtasks respectively. A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the task scheduling method described in any one of claims 1 to 18 is implemented. A computer device, comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, the processor When the processor executes the program, the task scheduling method described in any one of claim items 1 to 18 is realized.

Images