// TRACE

Overview

I/O traces play a critical role in storage systems evaluation. They are captured through a variety of mechanisms, analyzed to understand the characteristics and demands of different applications, and replayed against real and simulated storage systems to recreate representative workloads. Often, traces are much easier to work with than actual applications, particularly when the applications are complex to configure and run, or involve confidential data or algorithms.

//TRACE is a new approach for extracting and replaying traces of parallel applications. Its tracing engine (the causality engine) automatically discovers inter-node data dependencies and inter-request compute times for each node (process) in an application. It does so by selectively delaying I/O in order to expose data dependencies among the compute nodes. The learned dependency information is saved in per-node annotated I/O traces. Such annotation allows a parallel replayer to closely mimic the behavior of a traced application.

People

• Greg Ganger
• David O'Hallaron

• Gregg Economou
• Michael Stroucken

• James Hendricks
• Julio López
• Mike Mesnier
• Raja Sambasivan
• Matthew Wachs

Publications

• //TRACE: Parallel Trace Replay with Approximate Causal Events. Michael Mesnier, Matthew Wachs, Raja R. Sambasivan, Julio Lopez, James Hendricks, Gregory R. Ganger. Proceedings of the 5th USENIX Conference on File and Storage Technologies (FAST '07), February 13–16, 2007, San Jose, CA. Supercedes Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-06-108, September 2006.
  Abstract / PDF[ 187K]

Acknowledgements

We thank the members and companies of the PDL Consortium: American Power Conversion, Cisco Systems, EMC, Google, Hewlett-Packard Labs, Hitachi, IBM, Intel, LSI, Network Appliance, Oracle, Panasas, Seagate Technology, and Symantec for their interest, insights, feedback, and support.