Computer Systems Colloquium
"Efficient Parallelisation by Unifying Loop and Data Transformations"
3.30 pm, Thursday 3 February
Room 2511, James Clerk Maxwell Building
Abstract
This talk describes recent results in a long standing interest of mine, namely
how can we reconcile control based approaches to program restructuring
versus data orientated approaches. It will describe work in minimising
parallelisation overhead on distributed shared memory machines, such as
the SGi Origin 2000, by the combination of non-singular loop and data
transformations.
By treating loop and data transformations in a novel unified manner,
we show that conflicting requirements on a loop
transformation may be resolved by using a data transformation and
vice versa. We develop optimisation criteria for locality,
synchronisation and communication, and show that neither loop nor data
transformations can be solely used for efficient parallelisation. This
leads to the development of a novel global optimisation heuristic which
is applied to 3 SPEC kernels where it is shown to outperform
techniques solely based on loop or data transformations and gives
significant improvement over an existing state-of-the-art commercial
auto-paralleliser.
Last modified: Thu Jan 27 10:23:04 GMT 2000
Please contact our
webadmin with any comments or changes.
Unless explicitly stated otherwise, all material is
copyright © The University of Edinburgh.
