Diagnosing Performance Problems in Parallel File Systems

Abstract This work describes and compares two black-box approaches, using syscall statistics and OS-level perfor- mance metrics, to automatically diagnose different performance problems in parallel file systems. Both approaches rely on peer-comparison diagnosis to compare,statistical attributes of relevant metrics across servers in order to indict the culprit node. An observation-based checklist is developed to identify from the metrics affected the faulty resource, and it is demonstrated that this checklist applies commonly across stripe-based parallel file systems. These approaches are demonstrated for four realistic problems‐disk-hog, disk-busy, network-hog, and packet-loss‐injected into three different file-system benchmarks, dd, postmark, and IOzone, in PVFS and Lustre clusters. Contents