simulation: Simulating a $2M commercial server on a $2K PC?
Tuesday December 2, 2003
Hamerschlag Hall D-210
Carnegie Mellon University
These days, more than enough people in the computer architecture community
seem to believe that SPEC2000 CPU benchmarks are doomed. Your unrealistic
2X speed up from GCC simulation is not very convincing. Industry people
simply do not care. Instead, they will be crazy if your computer system
shows 10 percent improvement under database and web server simulation.
It is time that we computer architects need to present the performance
of new designs through simulations of commercial workloads, which will
require full system simulation including operating system and I/O. It
will be even better if we can simulate a $2M server system on our $2K
Pentium4 desktop. It sounds too good. What are major issues behind full
In this talk, I will discuss several issues behind full system simulation
such as benchmark tuning, non-deterministic measurement and simulation
validation. I will also give you brief summary of related works and our
current approach at Carnegie Mellon.
Jangwoo Kim is a Ph.D candidate in the Electrical and Computer Engineering
Department at Carnegie Mellon. He received his B.S. degree in Electrical
and Computer Engineering and M.S degree in Computer Science, both from
Cornell University. His current research interests are in fault tolerant
microprocessor, fault tolerant distributed computer system, full system
simulation and interconnection network. His advisor is Babak Falsafi.