PROTOFLEX: FPGA-Accelerated Instrumentation
Tuesday September 23, 2008
Hamerschlag Hall D-210
4:30 pm
Michael Papamichael
Carnegie Mellon University
The use of FPGAs has proven to be very effective at overcoming the
throughput limitations of existing full-system functional multiprocessor
simulators. A key advantage of FPGA-accelerated simulation over conventional
software simulation is the support for low overhead instrumentation, which
can be carried out by FPGA-resident instrumentation components operating in
parallel. Fast, flexible functional instrumentation is useful for various
hardware research activities, such as trace generation, workload warming and
characterization, development of software debugging/monitoring tools and
carrying out architectural studies. In the context of performance studies
using simulation sampling techniques (e.g. SMARTS), fast FPGA-based
instrumentation can vastly reduce simulation turn-around time.
This presentation is focused on two examples of FPGA-accelerated
instrumentation components, which leverage our existing FPGA-based
full-system functional simulation infrastructure and can be used in the
scope of fast functional workload warming. The first instrumentation
component functionally models a two-level Piranha-like chip multiprocessor
cache hierarchy with private L1 I&D caches and a shared non-inclusive L2,
while the second instrumentation component simulates the branch predictors
of the target multiprocessor system. The presentation includes
virtualization techniques that significantly simplified the implementation
and concludes with a brief demo of a web-based interface for real-time
display of system and cache statistics.
Michael K. Papamichael is a second-year Ph.D. student in the Computer
Science Department at Carnegie Mellon University, working with Prof. James
C. Hoe. His research interests are in the area of computer architecture,
with emphasis on FPGA-accelerated simulation and instrumentation for
multiprocessor architectures.
|