Toward Scheduling I/O Request of Mapreduce Tasks Based on Markov Model

Abstract

In Cloud storage of multiple CPU cores, many Mapreduce applications may run in parallel on each compute node and collocate with local Disks storage. These Disks storage are shared by multiple applications that use full CPU power of the node. Each application tends to issue contiguous I/O requests in parallel to the same Disk; however if large number of Mapreduce tasks enters the I/O phase at the same time, the requests from the same task may be interrupted by the requests of other tasks. Then, the I/O nodes receive these requests as non-contiguous way under I/O contention. This interleaved access pattern causes performance degradation for Mapreduce application, this is particularly important when writing intermediate files by multiple tasks in parallel to the shared Disk storage. In order to overcome this problem, we have proposed approach for optimizing write access for Mapreduce application. The contributions of this paper are: 1 analyze the open issues on scheduling access request of Mapreduce workload; 2i?źpropose framework for scheduling and predicting I/O request of Mapreduce application; 3i?źdescribe each role of component that intervenes in the scheduling theses I/O request on Block-level of storage server to provide contiguous access.