On data skewness, stragglers, and MapReduce progress indicators
This work addresses the need for reliable progress indicators in MapReduce systems, which is crucial for programmers managing large-scale data processing, though it is incremental as it builds on existing progress analysis methods.
The paper tackles the problem of predicting MapReduce job performance by designing accurate progress indicators, showing that existing methods are harmed by data skewness and stragglers, and introduces NearestFit, which improves accuracy with minimal overhead in real-world benchmarks.
We tackle the problem of predicting the performance of MapReduce applications, designing accurate progress indicators that keep programmers informed on the percentage of completed computation time during the execution of a job. Through extensive experiments, we show that state-of-the-art progress indicators (including the one provided by Hadoop) can be seriously harmed by data skewness, load unbalancing, and straggling tasks. This is mainly due to their implicit assumption that the running time depends linearly on the input size. We thus design a novel profile-guided progress indicator, called NearestFit, that operates without the linear hypothesis assumption and exploits a careful combination of nearest neighbor regression and statistical curve fitting techniques. Our theoretical progress model requires fine-grained profile data, that can be very difficult to manage in practice. To overcome this issue, we resort to computing accurate approximations for some of the quantities used in our model through space- and time-efficient data streaming algorithms. We implemented NearestFit on top of Hadoop 2.6.0. An extensive empirical assessment over the Amazon EC2 platform on a variety of real-world benchmarks shows that NearestFit is practical w.r.t. space and time overheads and that its accuracy is generally very good, even in scenarios where competitors incur non-negligible errors and wide prediction fluctuations. Overall, NearestFit significantly improves the current state-of-art on progress analysis for MapReduce.