MapReduce has become popular and Reduce-side join is one of the most important application of MapReduce. Data skew, in which the data load assigned to each Reduce task fluctuates task by task, increases the MapReduce job completion time. This paper proposes a dynamic profiling and feedback framework that works on a MapReduce cluster. The framework allows programmers to build their own algorithm to address data skew on Reduce-side join based on their specific knowledge and/or requirements. This paper also proposes an estimation method which makes our framework adapt to a wide range of MapReduce cluster sizes. This paper presents two example algorithms to address data skew using the estimation method, and the experimental results shows up to 2.59 times speed-up of join completion time on a cluster with 50 servers and highly skewed input data.
|出版物ステータス||Published - 2013 12 1|
|イベント||2013 16th IEEE International Conference on Computational Science and Engineering, CSE 2013 - Sydney, NSW, Australia|
継続期間: 2013 12 3 → 2013 12 5
|Conference||2013 16th IEEE International Conference on Computational Science and Engineering, CSE 2013|
|期間||13/12/3 → 13/12/5|
ASJC Scopus subject areas
- Computer Science (miscellaneous)