Trace Replay with Approximate Causal Events
Tuesday February 6, 2007
Hamerschlag Hall D-210
Intel Research & Carnegie Mellon University
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.
However, one well-known problem with trace replay is the lack of appropriate feedback
between storage response times and the arrival rate of requests. Information
regarding such feedback is rarely present in I/O traces, leaving replayers with
little guidance as to the proper replay rate.
This talk presents //TRACE - a new approach to extracting and replaying the I/O of
parallel applications. //TRACE contains a tracing engine that automatically discovers
inter-node data dependencies and inter-I/O compute times for each node (process) in
an application. This information is reflected in per-node annotated I/O traces,
allowing a parallel replayer to closely mimic the behavior of a traced application.
Mike Mesnier is a research scientist at Intel and also a 5th year doctoral candidate
in ECE's Parallel Data Lab. Prior to joining Intel in 1998, Mike received his Master
of Computer Science from the University of Illinois (UIUC), and prior to that was as
a researcher at Argonne National Laboratory.
Mike's interests include storage systems, distributed and parallel systems,
high-performance computing, and machine learning for autonomics. His current projects
include storage device modeling (thesis work), I/O tracing and replay, and