score-paperdvi-ProblemDiagnosisinLarge-Scale：在大型的得分paperdvi问题诊断.pdfVIP

Problem Diagnosis in Large-Scale Computing Environments Alexander V. Mirgorodskiy∗ Naoya Maruyama† Barton P. Miller‡ VMware, Inc. Department of Mathematical Computer Sciences Department and Computing Sciences University of Wisconsin Tokyo Institute of Technology Abstract or processes can be significantly harder than se- quential execution. In a concurrent system, events We describe a new approach for locating the causes can happen in different order in different runs. Fur- of anomalies in distributed systems. Our target en- thermore, concurrent systems suffer from bugs not vironment is a distributed application that contains present in sequential software. Bugs such as dead- multiple identical processes performing similar ac- locks and race conditions are notorious for being tivities. We use a new, lightweight form of dynamic hard to track down. Finally, the non-interactive instrumentation to collect function-level traces from nature of many system components complicates er- each process. If the application fails, the traces ror detection. A silent failure of a process on one of are automatically compared to each other. We find the hosts may remain unnoticed until a later time. anomalies by identifying processes that stopped ear- To simplify the tasks of problem detection and lier than the rest (sign of a fail-stop problem) or root cause analysis in large di