MapReduce programs and systems are large-scale, highly distributed and parallel, consisting of many interdependent Map and Reduce tasks executing simultaneously on potentially large numbers of cluster nodes. They typically process large datasets and run for long durations. Thus, diagnosing failures in MapReduce programs is challenging due to their scale. This renders traditional time-based Service-Level Objectives ineffective. Hence, even detecting whether a MapReduce program is suffering from a performance problem is difficult. Tools for debugging and profiling traditional programs are not suitable for MapReduce programs, as they generate too much information at the scale of MapReduce programs, do not fully expose the distributed interdependencies, and do not expose information at the MapReduce level of abstraction. Hadoop, the open-source implementation of MapReduce, natively generates logs that record the system's execution, with low overheads. From these logs, we can extract state-machine views of Hadoop's execution, and we can synthesize these views to create a single unified, causal, distributed control-flow and data-flow view of MapReduce program behavior. This state-machine view enables us to diagnose problems in MapReduce systems. We can also generate visualizations of MapReduce programs in combinations of the time, space, and volume dimensions of their behavior that can aid users in reasoning about and debugging performance problems. We evaluate our diagnosis algorithm based on these state-machine views on synthetically injected faults on Hadoop clusters on Amazon's EC2 infrastructure. Several examples illustrate how our visualization tools were used to optimize application performance on the production M45 Hadoop cluster.

113 pages

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by [email protected]