On The Metric Of Cloud Resources And The Corresponding Scheduling

Modern large-scale datacenters(DCs)consist of thousands of heterogeneous servers and the fine-grained heterogeneity of the hardware configurations among the machines becomes distinctive.This heterogeneity shows the differences among machines with respect to the processor's architecture,model,clock speed and caches,size and type of the memory,I/O speed of the disks,and so forth.However,current task-scheduling polices of today's cluster management systems are based on the coarse-grained metrics of resources such as the CPU cores and the size of memory.These policies ignore the fine-grained heterogeneity between different servers and do not consider the genuine computing powers of the heterogeneous machines.Thus,it cannot achieve the opti-mization of task scheduling and slow down the runtime of tasks.To solve above prob-lems,we propose SMHC,a synthetic metric for heterogenous cloud resouces,which has two main functions:1.SMHC encompasses five fine-grained hardware resources(CPU model,CPU speed,L3 cache of CPU,speed of memory,and I/O speed of disk).These five fine-grained hardware resources are able to approximately relect the computing power of a machine.2.By a mathematical model of linear combination,SMHC evaluates comprehen-sively the speed-up effects for big data tasks derived from both coarse-grained and fine-grained heterogeneous hardware resources.3.Based on SMHC,task-scheduling policies which compare the genuine com-puting power among heterogenous machines could be designed on cluster management system in order to effectively utilize the computing power of heterogeneous machines and accelerate the executions of tasks.In order to evaluate the effectiveness of SMHC,we propose a task-scheduling pol-icy based on the comparison of SMHC factors and implement it on Mesos,a popular dis-tributed system of resouces management.This task-scheduling policy based on SMHC compares the speed-up factors from different heteregenous machines,resulting in prior-itizing the servers of higher computing power to execute the big data programs.The ex-periment results show that compared with the original task-scheduling policy of Mesos based on coarse-grained resource metric,the task-scheduling policy with SMHC de-creases the runtime of Spark tasks from Hibench by 16.7%on average,because it could utilize the genuine computing power of fine-grained heterogenous machines to acceler-ate the executations of big data programs.