Performance Bug Analysis and Detection for Distributed Storage and Computing Systems

This article systematically studies 99 distributed performance bugs from five widely deployed storage and computing systems (Cassandra, HBase, HDFS, Hadoop MapReduce ZooKeeper). We present the TaxPerf database, which collectively organizes analysis results as over 400 classification labels 2,500 lines of bug re-description. is classified into six categories (and 18 subcategories) by their root causes; resource, blocking, synchronization, optimization, configuration, logic. can be used a benchmark for debug tool designs. Although it impractical to automatically detect all in TaxPerf, we find that an important category blocking effectively solved tools. analyze cascading nature design automatic detection called PCatch , (i) performs program identify code regions whose execution time potentially increase dramatically with workload size; (ii) adapts traditional happens-before model reason about software resource contention dependency relationship; (iii) uses dynamic tracking whether slowdown propagation contained one job. Evaluation shows accurately representative observing system executions under small-scale workloads.