Computer Studies Bldg. Room 703
Ph.D. Thesis Proposal
With the rapid adoption of multi-core, multi-processor systems,
applications are increasingly run in parallel using a shared
memory hierarchy. In traditional profiling, memory performance
of a single application is mostly measured by reuse distance, which
is defined for pairs of memory accesses to the same data. In a
shared environment, a reuse distance in one application is affected
by the footprint of other concurrent programs. Informally, the
footprint is the number of distinct data accessed in an execution
window. Complete measurement of footprints requires measuring
all possible execution windows, the number of which is quadratic to
the length of the execution.
In my thesis research, I will first develop efficient techniques
for measuring all-window footprint and then study the difference between
reuse distance and footprint especially their effect on shared cache
performance. Through the study, I will build footprint-based models
of cache sharing. Some footprint oriented improvement techniques
using the models will also be proposed. These techniques will be
explored to reduce application footprint and improve
its shared-cache performance on today's multicore systems.