Harnessing Chip-Multiprocessors with Concurrent Threaded Pipelines [abstract] (PDF)
John Giacomoni and Manish Vachharajani
University of Colorado Technical Report CU-CS-1024-07, January 2007.
Single-core performance increases have stalled. To increase available
cycles, microprocessor designers have shifted to chip-multiprocessor
(CMP) designs. Unfortunately, the additional processors provided by
CMPs may remain idle because most applications lack data-parallelism
and task-parallelism is unlikely to saturate future CMP designs. The
systems community needs to rethink how systems are structured to fully
utilize CMPs.
We propose that operating systems be adapted to harness CMP resources
by leveraging recent results in Concurrent Threaded Pipeline
(pipeline-parallel) organizations. This paper discusses potential
performance improvements of CTPs and the necessary OS support.