TurboSmarts:
They've gone to plaid!
Tuesday September 7, 2004
Hamerschlag Hall 1112
4:00 pm
Thomas Wenisch & Roland Wunderlich
Carnegie Mellon University
Prior work here at Carnegie Mellon, titled Smarts, proposed accelerating
microarchitecture simulation by extracting a small but representative
fraction of a benchmark's execution trace through statistical sampling.
To minimize error and increase confidence in performance estimates, Smarts
advocates detailed microarchitecture simulation of a large number (e.g.,
10,000) of brief (e.g., 1000-instruction) execution windows. Although
this approach requires only minutes of detailed simulation, warming requirements
increase total turnaround times to hours.
In this talk, we review our prior work on Smarts (last presented to
CALCM on 5/8/2003) and introduce TurboSmarts, a new framework that stores
warmed state in a library of reusable checkpoints. The fundamental insight
enabling TurboSmarts is that only a tiny fraction of warmed state is
necessary to measure brief execution windows. By identifying and storing
a minimal subset of model state in checkpoints, TurboSmarts can replace
slow functional warming with rapid loading of checkpoints, reducing simulation
runtime to minutes. We demonstrate that TurboSmarts can accurately estimate
the performance of an 8-way out-of-order superscalar processor running
SPEC CPU2000 in 91 seconds per benchmark on average with a 12 GB checkpoint
library for the entire benchmark suite.
Thomas Wenisch is a fourth year PhD student in computer architecture,
working with Prof. Babak Falsafi. Tom makes databases run faster on multiprocessors.
Roland Wunderlich is a fourth year PhD student in computer architecture,
working with Prof. James Hoe. Roland makes digital signal processing
run faster on VLIW.
|