CN112395052A - Container-based cluster resource management method and system for mixed load - Google Patents
Container-based cluster resource management method and system for mixed load Download PDFInfo
- Publication number
- CN112395052A CN112395052A CN202011412657.0A CN202011412657A CN112395052A CN 112395052 A CN112395052 A CN 112395052A CN 202011412657 A CN202011412657 A CN 202011412657A CN 112395052 A CN112395052 A CN 112395052A
- Authority
- CN
- China
- Prior art keywords
- task
- resource
- cluster
- straggler
- resources
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000007726 management method Methods 0.000 title claims abstract description 62
- 238000012544 monitoring process Methods 0.000 claims abstract description 74
- 238000011084 recovery Methods 0.000 claims abstract description 66
- 238000012545 processing Methods 0.000 claims abstract description 55
- 238000012216 screening Methods 0.000 claims abstract description 14
- 238000000034 method Methods 0.000 claims description 31
- 230000008569 process Effects 0.000 claims description 10
- 230000009286 beneficial effect Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000000725 suspension Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
- G06F9/45558—Hypervisor-specific management and integration aspects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/44—Arrangements for executing specific programs
- G06F9/455—Emulation; Interpretation; Software simulation, e.g. virtualisation or emulation of application or operating system execution engines
- G06F9/45533—Hypervisors; Virtual machine monitors
- G06F9/45558—Hypervisor-specific management and integration aspects
- G06F2009/4557—Distribution of virtual machine instances; Migration and load balancing
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a mixed load-oriented container-based cluster resource management method, which belongs to the field of cloud computing cluster resource management and scheduling and comprises the following steps: task state collection step, comprising: acquiring state information of running tasks on each working node, combining the task state information and the belonging operation information into task monitoring data, and storing the task monitoring data into a time sequence; task screening, comprising: identifying a Straggler task and a resource redundancy task according to the time sequence; a decision generation step, comprising: judging whether the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, if so, establishing a resource expansion strategy for the Straggler task, and establishing a resource recovery strategy for the preempted task; otherwise, after enough resources are obtained through resource recovery and/or preemption, a resource extension strategy is made for the Straggler task. The invention effectively ensures the service quality of the delay sensitive application in the mixed load deployment environment under the condition of reducing the performance loss of the batch processing tasks as much as possible.
Description
Technical Field
The invention belongs to the field of cloud computing cluster resource management and scheduling, and particularly relates to a mixed load-oriented container-based cluster resource management method and system.
Background
With the wide use of emerging technologies such as the internet of things and artificial intelligence, the diversity of data center applications is greatly increased in order to meet the service requirements of users. If the reserved resources or dedicated clusters are used to ensure the service quality of the application, the resource utilization rate of the data center is low, and the operation and maintenance cost of the data center is increased. Therefore, the data center begins to deploy various loads in a same cluster by using the characteristics of different resource requirements and the like of the workload, so that the cluster resources are shared by various loads, and the utilization rate of the cluster resources is improved. Research shows that batch processing type load needs a large amount of resources such as CPU, memory and the like without delay requirement, so that the utilization rate of cluster resources can be obviously improved by batch processing load. For example, Alibara deploys batch jobs with delay-sensitive, user-oriented Web services in the same cluster; microsoft must perform mixed deployment on online search service and batch processing operation, wherein the online search task is a delay sensitive task and has the characteristics of less required resources, low delay, high throughput and the like; google mixes enterprise-level Jobs (Production Tier Jobs) with high reliability, stringent service level requirements with batch Jobs with low priority no requirements. Multiple loads share cluster resources, and meanwhile, potential performance interference such as resource competition on software stacks of different levels exists, and performance loss of loads with different service quality requirements exists in different degrees. Since batch processing operation has no requirement for performance such as completion time, and the user-oriented delay-sensitive service has strict requirement for response delay, performance interference between tasks will greatly increase the completion time of the delay-sensitive task, and seriously affect the service quality of the application. Therefore, how to guarantee the service quality of the user-oriented delay-sensitive application in the mixed load deployment environment becomes a challenge faced by the existing cluster management system.
A job is a logical instance of an application or load in a cluster, and a job is typically co-processed by one or more tasks. As a basic execution unit of the job, the life cycle of the task mainly comprises two stages of task scheduling and task running. In an actual production environment, a user may request more resources for a job and may wish to process the job more quickly. However, only part of the allocated resources are used in the running process of the task, and a large amount of resources which are not used by the task are allocated, and are called redundant resources. In a large-scale cluster, particularly when a large amount of batch processing jobs exist, the proportion of redundant resources in the cluster is greatly increased, so that the distributable resources of the cluster are deficient, and the scheduling delay of a new task is increased. Although resource preemption based on killing and container suspension can ensure that the resource demand of delay-sensitive tasks reduces task scheduling delay, these methods do not solve the problem of low utilization of cluster resources and also cause severe performance loss of preempted tasks.
Moreover, the dependent relationship exists among all tasks belonging to the same operation. If a task in the job processing process is slow in processing progress or fails to run, a subsequent task which depends on the task cannot start processing, and the task is called a Straggler. It was found that 80% of the Straggler run times were 2-2.5 times slower than the normal tasks and that 10% of the Straggler ran 10 times slower. Straggler can slow down the process of job processing, causing the service to generate a large tail delay, reducing the quality of service of the application (especially delay sensitive services) and even failing to meet the service level target of the user. Straggler is therefore one of the important factors affecting the performance of delay-sensitive applications. Although the problem of the Straggler can be solved by creating the task copy at present, the process progress of the task is lost by creating the copy, and meanwhile, complete task resources need to be distributed, so that the resource overhead of the task is increased.
Generally, the existing hybrid load-oriented cluster resource management method cannot well guarantee the service quality of the user-oriented delay-sensitive application in the hybrid load deployment environment, and greatly loses the processing performance of batch processing jobs.
Disclosure of Invention
Aiming at the defects and improvement requirements of the prior art, the invention provides a cluster resource management method and system based on a container and oriented to mixed load, aiming at effectively ensuring the service quality of delay sensitive application in a mixed load deployment environment under the condition of reducing the performance loss of batch processing tasks as much as possible.
To achieve the above object, according to an aspect of the present invention, there is provided a method for managing mixed-load container-based cluster resources, where a containerized operating environment is deployed in a cluster, the method including: a task state collection step, a task screening step and a decision generation step which are executed in the management node;
the task state collection step comprises the following steps: acquiring state information and operation information of tasks on each working node, combining the state information of the tasks and the operation information of the operation to which the tasks belong into task monitoring data, and storing the task monitoring data and the task monitoring data into a pre-established time sequence;
the task screening step comprises the following steps: identifying a Straggler task in the delay sensitive tasks and a resource redundancy task in the batch processing tasks according to the monitoring data in the time sequence;
the decision generating step comprises: judging whether the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, if so, establishing a resource expansion strategy for the Straggler task; otherwise, resource expansion strategies are formulated for the Straggler task after the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion through a resource recovery and/or resource preemption mode.
The invention deploys a containerized operating environment in the cluster, so that all tasks in the cluster are operated in the container, thereby calling an API of a container management program to accurately monitor the operating state of the tasks in real time, accurately identifying the Straggler task and the resource redundancy task on the basis, and recovering the resources through a resource recovery and/or resource preemption mode if necessary, thereby ensuring that the resource requirements of sensitive delay type tasks such as the Straggler task and a new task are met, and further effectively ensuring the service quality of delay sensitive applications in a mixed load deployment environment.
Further, in the task screening step, the mode of identifying the Stragger task according to the monitoring data in the time sequence comprises the following steps:
(S1) for each delay-sensitive task, obtaining the task from the time sequence at the last four monitoring time t1、t2、t3、t4Progress of treatment p1、p2、p3、p4And respectively taking the average processing speeds of the estimation task in two different time periods as follows: v. of1＝(p2-p1)/(t2-t1)，v2＝(p4-p3)/(t4-t3) And therefore the estimated task completion time is: t ═ 1-p4)/v2+t4；
(S2) if v2>v1If the processing speed is increased, the task is determined not to be the Straggler task, and the process proceeds to step (S4); otherwise, go to step (S3);
(S3) if T<TSLOIf the estimated task completion time can meet the user service level target, judging that the task is not a Stragger task, and turning to the step (S4); otherwise, the estimated task completion time cannot meet the user service level target, the task is judged to be a Straggler task, and the step (S4) is carried out;
(S4) recognizing the end;
wherein, t1<t2<t3<t4；TSLOIs the longest task runtime that can meet the user service level objectives.
According to the method, the processing speed of the tasks in different time periods and the completion time of the tasks are estimated by sampling in the time sequence in which the task state information is stored, the tasks which cannot meet the user service level target in the estimated task completion time are identified as Straggler tasks according to the gradual reduction of the task processing speed, and the identification basis is consistent with the running characteristics of the Straggler tasks, so that the Straggler tasks in the cluster can be accurately identified, and a foundation is provided for ensuring that the resource requirements of the Straggler tasks in the cluster are met.
Further, in the task screening step, the resource redundancy task is identified according to the monitoring data in the time sequence, and the method comprises the following steps:
(T1) for each batch task, the actual resource utilization of the computing task is: u shapec＝Max{Ri|i∈S}/Ra；
(T2) if Uc<UECRUIf the actual resource utilization rate of the task is smaller than the expected cluster resource utilization rate, the task is judged to be a resource redundancy task; otherwise, judging that the task is not a resource redundancy task;
wherein S represents a timestamp set corresponding to task-related state information in a time sequence, RiIndicating the resources actually used by the task at time i, RaResource, U, representing task allocationECRURepresenting the desired cluster resource utilization.
The method and the device acquire the resource usage amount of the task at each moment according to the time sequence stored with the task state information, estimate the ratio of the maximum resource usage amount to the resource allocated by the task as the actual resource utilization rate of the task, and identify the task with the estimated resource utilization rate less than the expected cluster resource utilization rate as the resource redundancy task, wherein the identification basis can accurately identify the redundancy resource in the cluster.
Further, the decision generating step comprises the steps of:
(1) obtaining free resources R in a clusterIDLEAnd calculating the sum R of the resource requirements needed by the Straggler task and the newly arrived delay sensitive task in the clusterNE；
(2) If R isIDLE>RNEIf the cluster idle resources can meet the task resource requirements, the step (3) is carried out; otherwise, indicating that the cluster idle resources can not meet the task resource requirements, and updating the allocable resources in the cluster to be RM＝RIDLEAnd go to step (5);
(3) judging whether the unextended Straggler task still exists in the cluster, if so, turning to the step (4); otherwise, turning to the step (10);
(4) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource expansion requirement of the Stragler task, if so, making a resource expansion strategy for the Stragler task, and turning to the step (3); otherwise, selecting a working node of which the allocable resource can meet the resource expansion requirement of the Straggler task, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (3);
(5) judging whether a resource redundancy task which is not subjected to resource recovery still exists in the cluster, if so, turning to the step (6); otherwise, turning to the step (7);
(6) selecting a resource redundancy task and making a resource recovery strategy for the resource redundancy task, and updating the distributable resource of the cluster to be RM＝RM+RrIf updated RM>RNEIf the distributable resources can meet the task resource requirements after the resources are recycled, the step (8) is carried out; otherwise, indicating that the distributable resources still can not meet the task resource requirements after the resources are recovered, and turning to the step (5);
Rrthe amount of resources that can be reclaimed for a resource redundancy task;
(7) selecting a batch processing task and making a resource preemption strategy for the batch processing task, and updating the distributable resource of the cluster to be RM＝RM+RpIf updated RM>RNEIf the allocable resources can meet the task resource requirements after the resources are preempted, the step (8) is carried out; otherwise, indicating that the allocable resources still can not meet the task resource requirements after the resources are preempted, and turning to the step (7);
Rpthe amount of resources that can be preempted for the batch processing task;
(8) judging whether the unextended Straggler task still exists in the cluster, if so, turning to the step (9); otherwise, turning to the step (10);
(9) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource requirement of the Stragler task expansion, if so, making a resource expansion strategy for the Stragler task, and turning to the step (8); otherwise, selecting a working node of which the allocable resource can meet the resource requirement of the Straggler task expansion, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (8);
(10) the operation is ended.
The resource expansion strategy is formulated for the Straggler task through the mode, specifically, when the cluster idle resources can meet the resource requirements of a new task and the Straggler task, the resource expansion strategy is formulated for the Straggler task directly, when the cluster idle resources can not meet the resource requirements of the new task and the Straggler task, the cluster allocable resources are increased in a mode of preferentially recovering resources from a resource redundancy task and preempting the resources of a batch processing task through the mode, the cluster allocable resources can meet the resource requirements of the new task and the Straggler task finally, then the resource expansion strategy is formulated for the Straggler task, the resource requirements of the Straggler task expansion are guaranteed to be met, and the resources of the batch processing task are prevented from being preempted as much as possible.
Further, in the step (3) of the decision generating step, if there is no unexpanded Straggler task in the cluster, before the step (10), the method further includes the following steps:
(W1) computing the remaining free resources in the cluster as RL＝RIDLE-RNEAnd calculating the current idle resource occupation ratio of the cluster as eta-RL/RTIf 1- η < UECRUIf the cluster resource utilization rate does not reach the expected cluster resource utilization rate, the step (W2) is carried out; otherwise, go to step (W4);
(W2) judging whether the task whose resource is preempted and whose resource needed for recovery is less than the rest of the idle resource in the cluster exists in the cluster, if yes, turning to the step (W3); otherwise, go to step (W4);
(W3) selecting a preempted resource, recovering a task with the required resource less than the rest idle resources in the cluster, judging whether the allocable resource on the working node where the task is located can meet the resource requirement of task recovery, if so, making a resource recovery strategy for the task, and turning to the step (W2); otherwise, selecting a working node of which the allocable resource can meet the resource requirement for the task recovery, allocating the working node to the task, making a resource recovery strategy for the task, and turning to the step (W2);
(W4) the operation ends;
wherein R isTRepresents the total resource amount, U, of the clusterECRURepresenting the desired cluster resource utilization.
After resource expansion strategies are established for all the Straggler tasks, under the condition that the utilization rate of cluster resources is low, the remaining idle resources are further distributed to the tasks of the preempted resources to recover the tasks, so that the invention can reduce the performance loss caused by the resource preemption to batch processing operation while ensuring the service quality of the delay sensitive tasks.
Furthermore, the mixed load container-oriented cluster resource management method further comprises a decision execution step executed in the working node;
the decision execution step comprises:
after receiving the resource recovery strategy from the management node, updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy;
after receiving the resource preemption strategy from the management node, updating the resource limit of the corresponding task container instance to the minimum value, and pausing the operation of the container;
after receiving a resource expansion strategy or a resource recovery strategy from a management node, judging whether a work node is reallocated for a task, if so, migrating a corresponding task container instance to the newly allocated work node, and then updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; and if the task container instance is in the suspended state, the operation of the container instance is resumed after the resource limit of the container instance is updated.
Further, in the task state collection step, after the state information of the task and the job information of the job to which the task belongs are merged into the task monitoring data, the mode of storing the task monitoring data in the time sequence includes:
acquiring monitoring time t of earliest task monitoring data in time sequence0If the monitoring time t and the monitoring time t of the current task monitoring data0Time interval t-t of0<D, storing the current task monitoring data into a time sequence; if t-t0If the time sequence is more than or equal to D, deleting the earliest task monitoring data in the time sequence and storing the current task monitoring data in the time sequence;
wherein D is a preset time interval.
The invention only maintains the task monitoring data of the latest period of time in the time sequence, thereby being beneficial to quickly identifying the task category and improving the identification accuracy.
Further, the cluster resource management method based on the container and oriented to the mixed load further comprises a task monitoring step executed in the working node;
the task monitoring step comprises:
acquiring all task container examples in a running state on a working node, acquiring state information of a task in real time, and reporting the state information to a management node; the state information includes the running time, processing progress and resource usage of the task.
According to another aspect of the invention, a mixed load container-based cluster resource management system is provided, wherein a containerized operating environment is deployed in a cluster, and the system comprises a task state collector, a task filter and a decision generator which are positioned on a management node;
the task state collector is used for acquiring state information and operation information of tasks on each working node, combining the state information of the tasks and the operation information of the operations to which the tasks belong into task monitoring data and storing the task monitoring data in a pre-established time sequence;
the task screener is used for identifying a Straggler task in the delay sensitive tasks and a resource redundancy task in the batch processing tasks according to the monitoring data in the time sequence;
and the decision generator is used for making a resource expansion strategy for the Straggler task when the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, and making the resource expansion strategy for the Straggler task after the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion through a resource recovery and/or resource preemption mode when the allocable resources in the cluster can not meet the resource requirements of the new task and the Straggler task expansion.
Furthermore, the cluster resource management system facing the mixed load and based on the container also comprises a decision executor positioned at the working node;
the decision executor is used for updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy after receiving the resource recovery strategy from the management node;
the decision executor is also used for updating the resource limit of the corresponding task container instance to the minimum value after receiving the resource preemption strategy from the management node, and pausing the operation of the container;
the decision executor is also used for judging whether the task is redistributed with the working nodes or not after receiving the resource expansion strategy or the resource recovery strategy from the management node, if so, the corresponding task container instance is migrated to the newly distributed working nodes, and then the resource limit of the corresponding task container instance is updated to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; and if the task container instance is in the suspended state, the operation of the container instance is resumed after the resource limit of the container instance is updated.
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
(1) according to the method, the task is containerized, so that the monitoring on the task is converted into the monitoring on the running state of the container, and the resource updating on the task is converted into the resource updating on the container instance, so that the running state information of the task can be accurately obtained in real time, and the expense of stopping running of the previous task and the risk of losing the task processing progress can be avoided; on the basis, the invention identifies the Straggler task, eliminates the Straggler through operations such as resource expansion and the like, reduces the influence on the operation completion time, improves the service tail delay and ensures the service quality of the application.
(2) By identifying the resource redundancy tasks and dynamically recovering the operation redundancy resources in the cluster in the process of the Straggler task expansion, the invention can ensure the resource requirements of new tasks and expanded Straggler tasks in the cluster, effectively reduce the scheduling delay of delay sensitive tasks, improve the service quality of delay sensitive applications, and simultaneously reduce the performance influence of resource preemption on batch processing operation.
(3) The invention can recover the task of the preempted resource as far as possible when the Straggler task is expanded and the utilization rate of the system resource is low by timely recovering the redundant resource of the task in the cluster and allocating the recovered resource to a new task or the Straggler task, thereby effectively improving the utilization rate of the cluster resource and effectively avoiding the performance loss of the task of the preempted resource.
Drawings
Fig. 1 is a schematic diagram of a hybrid load container-oriented cluster resource management method according to an embodiment of the present invention;
FIG. 2 is a flowchart of task state collection steps provided by an embodiment of the present invention;
FIG. 3 is a flowchart of task screening steps provided by an embodiment of the present invention;
FIG. 4 is a flowchart of decision making steps provided by an embodiment of the present invention;
FIG. 5 is a flowchart of decision-making steps provided by an embodiment of the present invention;
FIG. 6 is a flowchart of task monitoring steps provided by an embodiment of the present invention;
fig. 7 is a schematic diagram of a hybrid load container-oriented cluster resource management system according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
In the present application, the terms "first," "second," and the like (if any) in the description and the drawings are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order.
The invention provides a mixed load container-based cluster resource management method and system, aiming at the technical problem that the existing cluster resource management method can not effectively ensure the service quality of delay-sensitive application in a mixed load deployment environment, and the overall thought of the method is as follows: the method comprises the steps of containerizing tasks to accurately acquire running state information of the tasks in real time and avoid the expense of stopping running of the tasks and the loss of task processing progress, on the basis, identifying the Stragger tasks and the resource redundancy tasks in a cluster, directly making corresponding resource expansion strategies for the identified Stragger tasks when the cluster allocable resources are sufficient, preferentially recovering the resources of the resource redundancy tasks when the cluster allocable resources are insufficient, obtaining enough resources through a resource preemption mode if necessary, and then making corresponding resource expansion strategies for the identified Stragger tasks, so as to ensure the service quality of delay sensitive tasks, improve the utilization rate of the cluster resources and reduce the performance loss caused by resource preemption on batch processing operation.
Before explaining the technical scheme of the invention in detail, the related technical terms in the cluster are explained as follows:
management Node (Master Node): in order to achieve high reliability of a cluster, a plurality of management nodes are usually operated in the cluster, wherein one management node normally works and the other management nodes are ready at any time.
Working Node (Worker Node): the system mainly comprises a virtual machine or a physical machine which mainly provides task calculation or data storage capacity in a cluster, wherein all working nodes in the cluster are uniformly managed by a management node;
job (Job): a logical instance of an application or load in a cluster;
task (Task): a minimum unit of job execution, one job typically being co-processed by one or more tasks;
cluster manager (resource manager, RM): a process program for managing the whole cluster resource, and scheduling and managing the operation in the cluster;
node manager (NodeManager, NM): the system runs on the working nodes and manages the resources of the working nodes and the tasks running on the working nodes.
The following are examples.
Example 1:
a cluster resource management method based on a container and oriented to mixed load is characterized in that a containerized operation ring is deployed in a cluster; as shown in fig. 1, the method includes: a task state collection step, a task screening step and a decision generation step which are executed in the management node;
as shown in fig. 2, the task state collection step includes:
acquiring state information and operation information of tasks on each working node; optionally, in this embodiment, the state information of the task includes: the task running time, the task processing progress, the resources used by the task and the like, and the job information comprises: all tasks of the job and the dependency relationship among the tasks; the related information can be directly obtained from the cluster management;
merging the state information of the task and the operation information of the operation to which the task belongs into task monitoring data and storing the task monitoring data into a pre-established time sequence; optionally, in this embodiment, the time series formed by the monitoring data is specifically stored in the time series database;
as a preferred embodiment, in the task state collection step of this embodiment, a method of merging the state information of the task and the job information of the job to which the task belongs into the task monitoring data and storing the task monitoring data in the time series includes:
acquiring monitoring time t of earliest task monitoring data in time sequence0If the monitoring time t and the monitoring time t of the current task monitoring data0Time interval t-t of0<D, storing the current task monitoring data into a time sequence; if t-t0If the time sequence is more than or equal to D, deleting the earliest task monitoring data in the time sequence and storing the current task monitoring data in the time sequence;
wherein D is a preset time interval, and in practical application, the setting can be carried out by referring to the characteristics of a time sequence database storage space, load running time and the like; in the embodiment, only the task monitoring data of the latest period of time is maintained in the time sequence, which is beneficial to rapidly identifying the task category and improving the accuracy of identification;
the task screening step comprises the following steps:
identifying a Straggler task in the delay sensitive tasks and a resource redundancy task in the batch processing tasks according to the monitoring data in the time sequence;
as shown in fig. 2, in the task screening step of this embodiment, the manner of identifying the Straggler task according to the monitoring data in the time series includes the following steps:
(S1) for each delay-sensitive task, obtaining the task from the time sequence at the last four monitoring time t1、t2、t3、t4Progress of treatment p1、p2、p3、p4And respectively taking the average processing speeds of the estimation task in two different time periods as follows: v. of1＝(p2-p1)/(t2-t1)，v2＝(p4-p3)/(t4-t3) And therefore the estimated task completion time is: t ═ 1-p4)/v2+t4Due to t4Is the time closest to the current time, and the corresponding processing speed is closest to the current actual processing speed according to t4The processing time of the speed estimation task at the moment can obtain a relatively accurate estimation result;
(S2) ifv2>v1If the processing speed is increased and the task can be completed quickly, the task is determined not to be the Stragger task, and the process proceeds to step (S4); otherwise, go to step (S3);
(S3) if T<TSLOIf the estimated task completion time can meet the user service level target, judging that the task is not a Stragger task, and turning to the step (S4); otherwise, the estimated task completion time cannot meet the user service level target, the task is judged to be a Straggler task, and the step (S4) is carried out;
(S4) recognizing the end;
wherein, t1<t2<t3<t4；TSLOIs the longest task running time that can meet the user service level objective; in the embodiment, by estimating the processing speeds of the tasks in different time periods and the completion time of the tasks, the tasks which cannot meet the user service level target according to the estimated task completion time are identified as the Stragler tasks according to the running characteristics of the Stragler tasks, so that the Stragler tasks in the cluster can be accurately identified, and a basis is provided for ensuring that the resource requirements of the Stragler tasks in the cluster are met;
as shown in fig. 3, in the task screening step of this embodiment, the resource redundancy task is identified according to the monitoring data in the time series, and the method includes the following steps:
(T1) for each batch task, the actual resource utilization of the computing task is: u shapec＝Max{Ri|i∈S}/Ra；
(T2) if Uc<UECRUIf the actual resource utilization rate of the task is smaller than the expected cluster resource utilization rate, the task is judged to be a resource redundancy task; otherwise, judging that the task is not a resource redundancy task;
wherein S represents a timestamp set corresponding to task-related state information in a time sequence, RiIndicating the resources actually used by the task at time i, RaResource, U, representing task allocationECRUPresentation periodThe expected cluster resource utilization rate; in the embodiment, the resource usage amount of the task at each moment is obtained according to the time sequence in which the task state information is stored, the ratio between the maximum resource usage amount and the resource allocated by the task is estimated as the actual resource utilization rate of the task, and the task with the estimated resource utilization rate smaller than the expected cluster resource utilization rate is identified as the resource redundancy task, so that the redundancy resource in the cluster can be accurately identified according to the identification basis;
the decision generating step comprises: judging whether the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, if so, establishing a resource expansion strategy for the Straggler task; otherwise, making a resource expansion strategy for the Straggler task after the allocable resources in the cluster can meet the resource requirements of a new task and the Straggler task expansion through a resource recovery and/or resource preemption mode;
as shown in fig. 4, the decision generating step of the present embodiment includes the following steps:
(1) obtaining free resources R in a clusterIDLEAnd calculating the sum R of the resource requirements needed by the Straggler task and the newly arrived delay sensitive task in the clusterNE；
Free resources R in a clusterIDLECan be directly obtained from the cluster manager; sum of resource requirements R required by Straggler task and newly arrived delay sensitive task in clusterNEIt can be calculated as follows:
computing total resources required for newly arriving delay sensitive tasks in a cluster
Counting total resources required by the Straggler task extension
Based on this, R is calculatedNE＝RN+RE；
Where j is the new task number of the delay sensitive service in the cluster, RjJ is more than or equal to 0 and less than or equal to n which is the resource needed by the jth new task and is the targetThe number of new tasks of the delay sensitive service in the former cluster, k is the serial number of the Straggler task in the cluster, RkExpanding needed resources for the kth Straggler task, wherein k is more than or equal to 0 and less than or equal to m, and m is the number of the Straggler tasks in the current cluster;
(2) if R isIDLE>RNEIf the cluster idle resources can meet the task resource requirements, the step (3) is carried out; otherwise, indicating that the cluster idle resources can not meet the task resource requirements, and updating the allocable resources in the cluster to be RM＝RIDLEAnd go to step (5);
(3) judging whether the unextended Straggler task still exists in the cluster, if so, turning to the step (4); otherwise, turning to the step (10);
(4) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource expansion requirement of the Stragler task, if so, making a resource expansion strategy for the Stragler task, and turning to the step (3); otherwise, selecting a working node of which the allocable resource can meet the resource expansion requirement of the Straggler task, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (3);
the resource expansion strategy is used for allocating extra resources for the Straggler task to enable the Straggler task to recover the normal processing speed;
(5) judging whether a resource redundancy task which is not subjected to resource recovery still exists in the cluster, if so, turning to the step (6); otherwise, turning to the step (7);
when allocable resources in the cluster are not enough to meet the resource requirements of the newly arrived delay sensitive task and the Straggler task, preferentially recovering redundant resources of the resource redundant task, and after the redundant resource recovery is finished, if the allocable resources are still not enough to meet the resource requirements of the newly arrived delay sensitive task and the Straggler task, acquiring enough resources in a mode of preempting resources of other batch processing tasks;
(6) selecting a resource redundancy task and making a resource recovery strategy for the resource redundancy task, and updating the distributable resource of the cluster to be RM＝RM+RrIf updated RM>RNEIf the distributable resources can meet the task resource requirements after the resources are recycled, the step (8) is carried out; otherwise, indicating that the distributable resources still can not meet the task resource requirements after the resources are recovered, and turning to the step (5);
Rrthe amount of resources that can be reclaimed for a resource redundancy task;
the resource recovery strategy is used for recovering the redundant resources of the resource redundancy task;
(7) selecting a batch processing task and making a resource preemption strategy for the batch processing task, and updating the distributable resource of the cluster to be RM＝RM+RpIf updated RM>RNEIf the allocable resources can meet the task resource requirements after the resources are preempted, the step (8) is carried out; otherwise, indicating that the allocable resources still can not meet the task resource requirements after the resources are preempted, and turning to the step (7);
Rpthe amount of resources that can be preempted for the batch processing task;
the resource preemption strategy is used for allocating minimum resources to the batch processing tasks, pausing the operation of the task container instance, and withdrawing and allocating the rest resources to other tasks;
(8) judging whether the unextended Straggler task still exists in the cluster, if so, turning to the step (9); otherwise, turning to the step (10);
(9) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource requirement of the Stragler task expansion, if so, making a resource expansion strategy for the Stragler task, and turning to the step (8); otherwise, selecting a working node of which the allocable resource can meet the resource requirement of the Straggler task expansion, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (8);
(10) finishing the operation;
by the method, redundant resources in the tasks can be recovered according to needs, and the resource requirements of the delay sensitive tasks are met;
scheduling a newly arrived delay sensitive task in the cluster by a self scheduler of the cluster management system;
in order to further reduce the performance loss caused by resource preemption on the batch processing job and improve the utilization rate of the cluster resources, as shown in fig. 4, in this embodiment, in step (3) of the decision generation step, if there is no unexpanded Straggler task in the cluster, before the step (10), the following steps are further included:
(W1) computing the remaining free resources in the cluster as RL＝RIDLE-RNEAnd calculating the current idle resource occupation ratio of the cluster as eta-RL/RTIf 1- η < UECRUIf the cluster resource utilization rate does not reach the expected cluster resource utilization rate, the step (W2) is carried out; otherwise, go to step (W4);
(W2) judging whether the task whose resource is preempted and whose resource needed for recovery is less than the rest of the idle resource in the cluster exists in the cluster, if yes, turning to the step (W3); otherwise, go to step (W4);
(W3) selecting a preempted resource, recovering a task with the required resource less than the rest idle resources in the cluster, judging whether the allocable resource on the working node where the task is located can meet the resource requirement of task recovery, if so, making a resource recovery strategy for the task, and turning to the step (W2); otherwise, selecting a working node of which the allocable resource can meet the resource requirement for the task recovery, allocating the working node to the task, making a resource recovery strategy for the task, and turning to the step (W2);
the resource recovery strategy is used for recovering the resource limitation of the task of the preempted resource to the setting before the resource is preempted and recovering the operation of the task container instance;
(W4) the operation ends;
wherein R isTRepresents the total resource amount, U, of the clusterECRURepresenting the desired cluster resource utilization.
Generally speaking, in this embodiment, a containerized operating environment is deployed in a cluster, on this basis, a Straggler task and a resource redundancy task are accurately identified, and if necessary, resources are also recovered by a resource recovery and/or resource preemption manner, so as to ensure that resource requirements of sensitive delay tasks such as the Straggler task and a new task are met, thereby effectively ensuring the service quality of delay sensitive applications in a mixed load deployment environment, and reducing the performance loss of batch processing operations.
Based on the resource adjustment policy, that is, the resource preemption policy, the resource recovery policy, the resource extension policy, or the resource recovery policy, which is made for each task in the management node, this embodiment further includes: a decision execution step executed in the working node;
as shown in fig. 5, the decision-making step includes:
after receiving the resource recovery strategy from the management node, updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy;
after receiving the resource preemption strategy from the management node, updating the resource limit of the corresponding task container instance to the minimum value, and pausing the operation of the container;
after receiving a resource expansion strategy or a resource recovery strategy from a management node, judging whether a work node is reallocated for a task, if so, migrating a corresponding task container instance to the newly allocated work node, and then updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; if the task container instance is in a suspended state, the container instance is resumed to run after the resource limit of the container instance is updated;
in this embodiment, the operations of updating the resource limit of the task container instance, suspending the running of the container, resuming the running of the container instance, and the like, may be completed by directly calling the corresponding API of the container management program.
This embodiment still includes: a task monitoring step executed in the working node;
as shown in fig. 6, the task monitoring step includes:
acquiring all task container instances in a running state on a working node;
acquiring state information of a task in real time and reporting the state information to a management node; the state information comprises information such as running time, processing progress and resource use condition of the task; the collection of state information may be accomplished by calling an API of the container manager.
Example 2:
a mixed load based container cluster resource management system is provided, wherein a containerized operating environment is deployed in a cluster, as shown in FIG. 7, the system comprises a task state collector, a task filter and a decision generator which are positioned at a management node;
the task state collector is used for acquiring state information and operation information of tasks on each working node, combining the state information of the tasks and the operation information of the operations to which the tasks belong into task monitoring data and storing the task monitoring data in a pre-established time sequence; optionally, as shown in fig. 7, in this embodiment, the task monitoring data is specifically stored in a time sequence database;
the task screener is used for identifying a Straggler task and a resource redundancy task according to the monitoring data in the time sequence;
the decision generator is used for making a resource expansion strategy for the Straggler task when the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, and making the resource expansion strategy for the Straggler task after the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion by means of resource recovery and/or resource preemption when the allocable resources in the cluster can not meet the resource requirements of the new task and the Straggler task expansion;
as shown in fig. 7, the present embodiment further includes: a decision executor located at the working node;
the decision executor is used for updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy after receiving the resource recovery strategy from the management node;
the decision executor is also used for updating the resource limit of the corresponding task container instance to the minimum value after receiving the resource preemption strategy from the management node, and pausing the operation of the container;
the decision executor is also used for judging whether the task is redistributed with the working nodes or not after receiving the resource expansion strategy or the resource recovery strategy from the management node, if so, the corresponding task container instance is migrated to the newly distributed working nodes, and then the resource limit of the corresponding task container instance is updated to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; if the task container instance is in a suspended state, the container instance is resumed to run after the resource limit of the container instance is updated;
as shown in fig. 7, the present embodiment further includes: a task monitor located at the work node;
the task monitor is used for acquiring all task container examples in a running state on the working node, acquiring the state information of the tasks in the task container examples in real time and reporting the state information to the management node; the state information comprises the running time, the processing progress and the resource use condition of the task;
in this embodiment, the detailed implementation of the task state collector, the task filter, the decision generator, the decision executor and the task monitor may refer to the description in the above method embodiments, and will not be repeated herein;
as shown in fig. 7, in this embodiment, the task state collector and the time sequence database in the management node are further divided into a job state information base, the decision generator and the task filter in the management node are divided into an adaptive decision generator, and the decision executor and the task monitor in the working node are divided into a node task manager.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A method for managing cluster resources based on a container and oriented to mixed load is characterized in that a containerized operating environment is deployed in the cluster, and the method comprises the following steps: a task state collection step, a task screening step and a decision generation step which are executed in the management node;
the task state collecting step comprises: acquiring state information and operation information of tasks on each working node, combining the state information of the tasks and the operation information of the operation to which the tasks belong into task monitoring data, and storing the task monitoring data and the task monitoring data into a pre-established time sequence;
the task screening step comprises the following steps: identifying a Straggler task in the delay sensitive tasks and a resource redundancy task in the batch processing tasks according to the monitoring data in the time sequence;
the decision generating step comprises: judging whether the allocable resources in the cluster can meet the resource requirements of the new task and the Straggler task expansion, if so, establishing a resource expansion strategy for the Straggler task; otherwise, making the distributable resources in the cluster capable of meeting the resource requirements of the new task and the Straggler task expansion by means of resource recovery and/or resource preemption, and then making a resource expansion strategy for the Straggler task.
2. The method for managing mixed load container-based cluster resources according to claim 1, wherein the task screening step comprises the following steps of identifying the Stragger task according to the monitoring data in the time sequence:
(S1) for each delay-sensitive task, obtaining the task' S last four monitoring time instants t from the time sequence1、t2、t3、t4Progress of treatment p1、p2、p3、p4And respectively taking the average processing speeds of the estimation task in two different time periods as follows: v. of1＝(p2-p1)/(t2-t1)，v2＝(p4-p3)/(t4-t3) And therefore the estimated task completion time is: t ═ 1-p4)/v2+t4；
(S2) if v2>v1If yes, judging that the task is not a Stragger task, and turning to the step (S4); otherwise, go to step (S3);
(S3) if T<TSLOIf the task is not the Stragger task, the process proceeds to step (S4); otherwise, judging the task as a Straggler task, and turning to the step (S4);
(S4) recognizing the end;
wherein, t1<t2<t3<t4；TSLOIs the longest task runtime that can meet the user service level objectives.
3. The method for managing cluster resources based on mixed load container as claimed in claim 1, wherein in the task screening step, the redundant tasks of resources are identified according to the monitoring data in the time series, in a manner including the following steps:
(T1) for each batch task, the actual resource utilization of the computing task is: u shapec＝Max{Ri|i∈S}/Ra；
(T2) if Uc<UECRUIf yes, judging the task as a resource redundancy task; otherwise, judging that the task is not a resource redundancy task;
wherein S represents a timestamp set corresponding to the task related state information in the time sequence, RiIndicating the resources actually used by the task at time i, RaResource, U, representing task allocationECRURepresenting the desired cluster resource utilization.
4. The hybrid load container-oriented cluster resource management method according to any one of claims 1 to 3, wherein the decision generating step comprises the steps of:
(1) obtaining free resources R in a clusterIDLEAnd calculates the delay of the Straggler task and new arrival in the clusterSum of resource requirements R required by sensitive tasksNE；
(2) If R isIDLE>RNEThen, the step (3) is carried out; otherwise, updating the allocable resources in the cluster to be RM＝RIDLEAnd go to step (5);
(3) judging whether the unextended Straggler task still exists in the cluster, and if so, turning to the step (4); otherwise, turning to the step (10);
(4) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource expansion requirement of the Stragler task, if so, making a resource expansion strategy for the Stragler task, and turning to the step (3); otherwise, selecting a working node of which the allocable resource can meet the resource expansion requirement of the Straggler task, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (3);
(5) judging whether a resource redundancy task which is not subjected to resource recovery still exists in the cluster, if so, turning to the step (6); otherwise, turning to the step (7);
(6) selecting a resource redundancy task and making a resource recovery strategy for the resource redundancy task, and updating the distributable resource of the cluster to be RM＝RM+RrIf updated RM>RNEThen, the step (8) is carried out; otherwise, turning to the step (5);
Rrthe amount of resources that can be reclaimed for a resource redundancy task;
(7) selecting a batch processing task and making a resource preemption strategy for the batch processing task, and updating the allocable resource of the cluster to be RM＝RM+RpIf updated RM>RNEThen, the step (8) is carried out; otherwise, turning to the step (7);
Rpthe amount of resources that can be preempted for the batch processing task;
(8) judging whether the unextended Straggler task still exists in the cluster, if so, turning to the step (9); otherwise, turning to the step (10);
(9) selecting an unexpanded Stragler task, judging whether the allocable resources on the working node where the task is located can meet the resource requirement of the Stragler task expansion, if so, making a resource expansion strategy for the Stragler task, and turning to the step (8); otherwise, selecting a working node of which the allocable resource can meet the resource requirement of the Straggler task expansion, allocating the working node to the Straggler task, making a resource expansion strategy for the Straggler task, and turning to the step (8);
(10) the operation is ended.
5. The method for managing mixed load container-oriented cluster resources according to claim 4, wherein in the step (3) of generating the decision, if there is no unexpanded Straggler task in the cluster, before proceeding to the step (10), the method further comprises the steps of:
(W1) calculating the remaining free resources in the cluster as RL＝RIDLE-RNEAnd calculating the current idle resource occupation ratio of the cluster as eta-RL/RTIf 1- η < UECRUThen go to step (W2); otherwise, go to step (W4);
(W2) judging whether the cluster has the preempted resource and the task whose recovery required resource is less than the rest idle resource in the cluster, if yes, turning to the step (W3); otherwise, go to step (W4);
(W3) selecting a preempted resource, recovering a task with a resource less than the rest of idle resources in the cluster, judging whether the allocable resource on the working node where the task is located can meet the resource requirement for task recovery, if so, making a resource recovery strategy for the task, and turning to the step (W2); otherwise, selecting a working node of which the allocable resource can meet the resource requirement for the task recovery, allocating the working node to the task, making a resource recovery strategy for the task, and turning to the step (W2);
(W4) the operation ends;
wherein R isTRepresenting the total resource amount, U, of the clusterECRURepresenting the desired cluster resource utilization.
6. The hybrid load container-oriented cluster resource management method of claim 5, further comprising a decision execution step performed in a worker node;
the decision execution step includes:
after receiving a resource recovery strategy from a management node, updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy;
after receiving the resource preemption strategy from the management node, updating the resource limit of the corresponding task container instance to the minimum value, and pausing the operation of the container;
after receiving a resource expansion strategy or a resource recovery strategy from a management node, judging whether a work node is reallocated for a task, if so, migrating a corresponding task container instance to the newly allocated work node, and then updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; and if the task container instance is in the suspended state, the operation of the container instance is resumed after the resource limit of the container instance is updated.
7. The method for managing cluster resources based on a hybrid load container as claimed in claim 1, wherein in the task state collection step, after merging the state information of the task and the job information of the job to which the task belongs into task monitoring data, the way of storing the task monitoring data in the time sequence includes:
acquiring the monitoring time t of the earliest task monitoring data in the time sequence0If the monitoring time t of the current task monitoring data and the monitoring time t0Time interval t-t of0<D, storing the current task monitoring data into the time sequence; if t-t0If the time sequence is more than or equal to D, deleting the earliest task monitoring data in the time sequence and monitoring the current taskStoring the control data into the time sequence;
wherein D is a preset time interval.
8. The hybrid load container-oriented cluster resource management method of claim 1, further comprising a task monitoring step performed in a worker node;
the task monitoring step comprises:
acquiring all task container examples in a running state on a working node, acquiring state information of a task in real time, and reporting the state information to the management node; the state information includes the running time, the processing progress and the resource use condition of the task.
9. A mixed load based container cluster resource management system is characterized in that a containerized operating environment is deployed in a cluster, and the system comprises a task state collector, a task filter and a decision generator which are positioned on a management node;
the task state collector is used for acquiring state information and operation information of tasks on each working node, combining the state information of the tasks and the operation information of the operation to which the tasks belong into task monitoring data and storing the task monitoring data in a pre-established time sequence;
the task screener is used for identifying a Straggler task in the delay sensitive tasks and a resource redundancy task in the batch processing tasks according to the monitoring data in the time sequence;
the decision generator is used for making a resource expansion strategy for the Stragger task when the allocable resources in the cluster can meet the resource requirements of the new task and the Stragger task expansion, and making the resource expansion strategy for the Stragger task after the allocable resources in the cluster can meet the resource requirements of the new task and the Stragger task expansion through a resource recovery and/or resource preemption mode when the allocable resources in the cluster can not meet the resource requirements of the new task and the Stragger task expansion.
10. The hybrid load container-oriented cluster resource management system of claim 9, further comprising a decision executor at a worker node;
the decision executor is used for updating the resource limit of the corresponding task container instance into the resource limit formulated in the resource recovery strategy after receiving the resource recovery strategy from the management node;
the decision executor is further used for updating the resource limit of the corresponding task container instance to a minimum value and pausing the operation of the container after receiving the resource preemption strategy from the management node;
the decision executor is further configured to determine whether a work node is reallocated for the task after receiving a resource extension policy or a resource recovery policy from the management node, and if so, update the resource restriction of the corresponding task container instance to the resource restriction formulated in the resource extension policy or the resource recovery policy after migrating the corresponding task container instance to the newly allocated work node; otherwise, directly updating the resource limit of the corresponding task container instance to the resource limit formulated in the resource expansion strategy or the resource recovery strategy; and if the task container instance is in the suspended state, the operation of the container instance is resumed after the resource limit of the container instance is updated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011412657.0A CN112395052B (en) | 2020-12-03 | 2020-12-03 | Cluster resource management method and system based on container for hybrid load |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011412657.0A CN112395052B (en) | 2020-12-03 | 2020-12-03 | Cluster resource management method and system based on container for hybrid load |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112395052A true CN112395052A (en) | 2021-02-23 |
| CN112395052B CN112395052B (en) | 2024-02-06 |
Family
ID=74604457
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011412657.0A Active CN112395052B (en) | 2020-12-03 | 2020-12-03 | Cluster resource management method and system based on container for hybrid load |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112395052B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113608875A (en) * | 2021-08-10 | 2021-11-05 | 天津大学 | High-throughput cloud computing resource recovery system |
| CN114077486A (en) * | 2021-11-22 | 2022-02-22 | 内蒙古大学 | MapReduce task scheduling method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110413391A (en) * | 2019-07-24 | 2019-11-05 | 上海交通大学 | Deep learning task service method for ensuring quality and system based on container cluster |
| WO2020206705A1 (en) * | 2019-04-10 | 2020-10-15 | 山东科技大学 | Cluster node load state prediction-based job scheduling method |
| CN111782355A (en) * | 2020-06-03 | 2020-10-16 | 上海交通大学 | Cloud computing task scheduling method and system based on mixed load |
-
2020
- 2020-12-03 CN CN202011412657.0A patent/CN112395052B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020206705A1 (en) * | 2019-04-10 | 2020-10-15 | 山东科技大学 | Cluster node load state prediction-based job scheduling method |
| CN110413391A (en) * | 2019-07-24 | 2019-11-05 | 上海交通大学 | Deep learning task service method for ensuring quality and system based on container cluster |
| CN111782355A (en) * | 2020-06-03 | 2020-10-16 | 上海交通大学 | Cloud computing task scheduling method and system based on mixed load |
Non-Patent Citations (1)
| Title |
|---|
| 杨忠: "面向Docker容器的动态负载集群伸缩研究", 舰船电子工程, vol. 38, no. 08 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113608875A (en) * | 2021-08-10 | 2021-11-05 | 天津大学 | High-throughput cloud computing resource recovery system |
| WO2023015787A1 (en) * | 2021-08-10 | 2023-02-16 | 天津大学 | High throughput cloud computing resource recovery system |
| CN113608875B (en) * | 2021-08-10 | 2023-09-12 | 天津大学 | High-throughput cloud computing resource recovery system |
| CN114077486A (en) * | 2021-11-22 | 2022-02-22 | 内蒙古大学 | MapReduce task scheduling method and system |
| CN114077486B (en) * | 2021-11-22 | 2024-03-29 | 内蒙古大学 | MapReduce task scheduling method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112395052B (en) | 2024-02-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108829494B (en) | Container cloud platform intelligent resource optimization method based on load prediction | |
| WO2021208546A1 (en) | Multi-dimensional resource scheduling method in kubernetes cluster architecture system | |
| CN103729248B (en) | A kind of method and apparatus of determination based on cache perception task to be migrated | |
| CN107832129B (en) | Dynamic task scheduling optimization method for distributed stream computing system | |
| CN111694633A (en) | Cluster node load balancing method and device and computer storage medium | |
| CN112395052B (en) | Cluster resource management method and system based on container for hybrid load | |
| CN110262897B (en) | Hadoop calculation task initial allocation method based on load prediction | |
| CN112328378A (en) | Task scheduling method, computer device and storage medium | |
| CN109783225B (en) | Tenant priority management method and system of multi-tenant big data platform | |
| CN113867959A (en) | Training task resource scheduling method, device, equipment and medium | |
| CN114661482B (en) | GPU (graphics processing Unit) computing power management method, medium, equipment and system | |
| WO2024021489A1 (en) | Task scheduling method and apparatus, and kubernetes scheduler | |
| CN113110914A (en) | Internet of things platform construction method based on micro-service architecture | |
| US20230037293A1 (en) | Systems and methods of hybrid centralized distributive scheduling on shared physical hosts | |
| CN105867998A (en) | Virtual machine cluster deployment algorithm | |
| CN116010064A (en) | DAG job scheduling and cluster management method, system and device | |
| CN113032102A (en) | Resource rescheduling method, device, equipment and medium | |
| CN113672391B (en) | Parallel computing task scheduling method and system based on Kubernetes | |
| CN111767145A (en) | Container scheduling system, method, device and equipment | |
| CN110084507B (en) | Scientific workflow scheduling optimization method based on hierarchical perception in cloud computing environment | |
| CN111404818A (en) | Routing protocol optimization method for general multi-core network processor | |
| CN114389955A (en) | Embedded platform heterogeneous resource pooling management method | |
| US20170024245A1 (en) | Workload-aware shared processing of map-reduce jobs | |
| CN114297002A (en) | Mass data backup method and system based on object storage | |
| CN111966481A (en) | Parallel computing management method and system suitable for multi-tenant scene |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |