I'll be giving a talk at the architecture seminar next Tuesday (3/7),
and I'd be thrilled to get some fellow PL folks involved in the
discussion. The talk is in 1221 CS at 4PM. Title and abstract are
below, or see the architecture seminar page:
http://www.cs.wisc.edu/arch/uwarch/seminar/seminar_s06.html
thanks,
wb
Title: Statically identifying implicit module-level parallelism
Abstract:
Currently available multi-core microprocessors offer hardware
support for concurrent workloads, but writing parallel programs is
still as hard as ever. My thesis research is concerned with
exploiting the modular structure of Java programs in order to
automatically find independent tasks. In this talk, I describe a
simple type-based analysis that statically identifies a
conservative estimate of the memory locations that are modified by
a given program fragment. I then show how to improve the precision
of this analysis by extending Java's type system with compiler-
supplied annotations that encode information about potential
aliasing relationships. A parallelizing compiler or JVM can
leverage the results of this analysis to determine whether or not
the side effects of two program fragments are independent -- and,
thus, whether those two fragments may be safely executed
concurrently. The program decomposition technique I propose does
not rely on any specialized hardware support and is complementary
to proposed work involving speculative execution and transactional
memory.
This is joint work with Charles Fischer.
Attachment:
smime.p7s
Description: S/MIME cryptographic signature
|