Downloads: 108 | Views: 168

Research Paper | Computer Science & Engineering | India | Volume 4 Issue 7, July 2015

Parallel Data Shuffling for Hadoop Acceleration with Network Levitated Merge and RDMA for Interconnectivity

Kishorkumar Shinde | Venkatesan N.

Abstract: Performance is measure issue in todays hadoop framework. The execution time required for Map reduce model is depends on multiple factors. Shuffling and merging in map reduce requires much amount of time. Proper implementation of shuffling and merging improves the performance of overall system. With this Serialization, multiple interconnect issues are also covered in this paper. Serialization keeps reduce phase to wait, repetitive merges requires multiple disk access and lack of portability for different interconnections. Repetitive merges can be reduced by network levitated merge algorithm, Serialization issue is overcome by parallelization. RDMA is used to for multiple interconnects. A non Hadoop and non java machine can also use the hadoop features. If we use pipelining to avoid serialization some sort of serialization is there in shuffle and merge phase. In pipelining output file is shuffled and merged before providing it to reduce task. Instead of pipelined shuffling, parallel shuffling is proposed. This reduces the number of disk accesses resulting in improved performance.

Keywords: Hadoop, Network levitated merge, MapReduce, Big- data, RDMA

Edition: Volume 4 Issue 7, July 2015,

Pages: 1096 - 1101