A Coarse-grained Task Scheduling Algorithm For Dynamic Task Flow

The continuous upgrading and technological innovation of cloud computing industry has made users put forward higher requirements for cloud computing.As the core algorithm in cloud computing,task scheduling affects the user experience and the service efficiency of the cloud platform.However,with the constantly increase of the number and scale of tasks,and the increasing proportion of energy consumption in data centers,reasonable and efficient task scheduling algorithms have become the key and difficult issues in cloud computing research.The general task scheduling algorithms lack detailed analysis of the actual data center load characteristics,and rarely divide tasks and resources,which increases the scope of task selection resources.Therefore,based on the characteristics of the actual data center task arrival and the features of the task itself,we propose a coarse-grained task scheduling algorithm for dynamic task flow,the dynamic task flow is constructed from the tasks arriving in each time period,and the simulation platform is used to compare and analyze the proposed algorithm.The main research work is as follows:(1)Load statistics and analysis of data center cluster trace data.According to the Google cluster trace data,the data center load is statistically analyzed,including machine statistics: various machine event distribution,daily machine quantity distribution,machine clustering analysis;job statistics: daily job quantity distribution,task number distribution in job,job waiting and running time,task clustering analysis.The results of data center load statistics and analysis can be used as the basis and support for the data center task scheduling.(2)The task load prediction model and the virtual machine opening strategy based on the time division of tasks are proposed.According to the data center load analysis result,the tasks arrive every day are divided into hours,and the number of tasks in each hour of each day is obtained.Then,the improved LSTM model based on periodic mean penalty is used to predict the number of tasks in each hour for the next day,so as to construct a dynamic task flow.The time is divided according to a specific time window and the number of tasks in each time period is counted,and the virtual machine is turned on as needed to reduce the energy consumption.The experimental results show that the LSTM prediction model based on the periodic mean penalty has a higher accuracy than the original model in predicting the distribution of tasks in a day,and after time division and turn on the virtual machine on demand,it can effectively reduce unnecessary energy consumption and improve the utilization of virtual machines.(3)A coarse-grained task scheduling algorithm with equivalence class partitioning is designed.After a specific number of virtual machines are turned on,a concrete task scheduling algorithm is required.According to the heterogeneity of tasks and resources in the cloud environment,tasks and resources are quantified in term of their attributes,the task and resource models are established.Then,the equivalence class partitioning is used to classify the diverse tasks and resources,and each group of task is allocated to the resource group with the matching execution ability according to its average instruction length,so that the task assignment within the group is reasonable and the range of resource selection for a single task within the group is reduced.Then,in the internal scheduling process of each task group,greedy strategy is adopted to improve the overall scheduling performance.Finally,experiments are carried out to compare the proposed algorithm with the round robin scheduling algorithm,the pure greedy scheduling algorithm,the ant colony algorithm and the genetic algorithm.The total task execution time,task completion time,virtual machine load and utilization are selected as the evaluation indexes.Experimental results show that the proposed algorithm is superior to all other algorithms in terms of total task execution time.In terms of task completion time and virtual machine utilization,it is better than round robin algorithm and genetic algorithm,and it is better than greedy algorithm and ant colony algorithm in terms of virtual machine load.