SMARTS:
Accelerating Microarchitecture Simulation via Rigorous Statistical Sampling
Thursday May 8, 2003
Hamerschlag Hall D-210
2:00pm
Tom Wenisch
Carnegie Mellon University
Computer architects rely on software simulation to evaluate and validate
the functionality and performance of new hardware designs. As computer
architecture designs have become more complex, the throughput of simulation
models has dropped dramatically: modern cycle-accurate processor simulators
are approximately five orders of magnitude slower than the modeled hardware,
and RTL-based models are six or more orders of magnitude slower. Because
of prohibitive simulation time, researchers often rely on performance
estimates from drastically abbreviated instruction execution streams,
which fail to correctly capture global variations in program behavior
and performance.
In this talk, I present the SMARTS framework for applying statistical
sampling theory to the problem of accelerating processor simulation while
achieving accurate results. I present conclusions about simulation sampling
derived from the theory, the practical problems associated with direct
application of these conclusions, and how SMARTS solves these problems.
I demonstrate a SMARTS-enabled simulation infrastructure which can estimate
performance and energy consumption with an average error of 0.6% and
a 35 to 60 x speedup over the baseline cycle-accurate simulation model.
Thomas Wenisch is a second year PhD student in computer architecture,
working with Prof. Babak Falsafi. Tom spends his days developing the
SimFlex simulation infrastructure, a full-system multiprocessor simulator
based on the SMARTS framework and Virtutech's Simics simulation engine.
|