PonIC: Using stratosphere to speed up pig analytics

Abstract

Pig, a high-level dataflow system built on top of Hadoop MapReduce, has greatly facilitated the implementation of data-intensive applications. Pig successfully manages to conceal Hadoop's one input and two-stage inflexible pipeline limitations, by translating scripts into MapReduce jobs. However, these limitations are still present in the backend, often resulting in inefficient execution. Stratosphere, a data-parallel computing framework consisting of PACT, an extension to the MapReduce programming model and the Nephele execution engine, overcomes several limitations of Hadoop MapReduce. In this paper, we argue that Pig can highly benefit from using Stratosphere as the backend system and gain performance, without any loss of expressiveness. We have ported Pig on top of Stratosphere and we present a process for translating Pig Latin scripts into PACT programs. Our evaluation shows that Pig Latin scripts can execute on our prototype up to 8 times faster for a certain class of applications. © 2013 Springer-Verlag.