Shadow: Exploiting the Power of Choice for Efficient Shuffling in MapReduce

Abstract

How to reduce the costly cross-rack data transferring is challenging in improving the performance of MapReduce platforms. Previous schemes mainly exploit the data locality in the Map phase to reduce the cross-rack communications. However, the Map locality based schemes may lead to highly skewed distribution of Map tasks across racks in the platform, resulting in serious load imbalance among different cross-rack links during Shuffling. Recent research results show that the slow Shuffling is the root cause of the MapReduce performance degradation. Very limited work has been done for speeding up the Shuffle phase. A notable scheme leverages the principle of the power of choice to balance the network loads on different cross-rack links during Shuffling for a specific type of sampling applications, where processing a random subset of the large-scale data collection is sufficient to derive the final result. The scheme launches a few additional tasks to offer more choices for task selection during Shuffling. However, such a scheme is designed for sampling applications and not applicable to general applications, where all the input data instead of a random subset is processed. In this work, we observe that with high Map locality, the network is mainly saturated in Shuffling but relatively free in the Map phase. A little sacrifice in Map locality may greatly accelerate Shuffling. Based on this, we propose a novel scheme called Shadow for Shuffle-constrained general applications, which strikes a trade-off between Map locality and Shuffling load balance. Specifically, Shadow iteratively chooses an original Map task from the most heavily loaded rack and creates a duplicated task for it on the most lightly loaded rack. During processing, Shadow makes a choice between an original task and its replica by efficiently pre-estimating the job execution time. We conduct extensive experiments to evaluate the Shadow design. Results show that Shadow greatly reduces the cross-rack skewness by 36.6% and the job execution time by 26% compared to existing schemes.