Check out the new USENIX Web site. next up previous
Next: Acknowledgements Up: //TRACE: Parallel trace replay Previous: Related work


Conclusion

This paper presents a technique for accurately extracting and replaying I/O traces from parallel applications. By selectively delaying I/O while tracing an application, computation time and inter-node dependencies can be discovered and approximated in trace annotations. Unlike previous approaches to trace collection and replay, such approximation allows a replayer to closely mimic the behavior of a parallel application. Across the applications and storage systems evaluated in this study, the average replay error is below 6%.



Michael Mesnier 2006-12-22