Title: Qthreads: Portable, Lightweight, Locality-aware Threading

Speaker: Kyle Wheeler, Sandia National Laboratories

Date/Time: Thursday, April 16, 2009, 10:30 am

Location: CSRI/279

Brief Abstract: Locality and data layout are critical to the performance of threaded parallel programs, but are not built into existing threading models. Standard threading interfaces presume, incorrectly, that data is equally accessible from all threads. Meanwhile, commodity processors are adopting features to enable greater levels of multi-threaded programming, but provide unique interfaces and memory layouts, hindering development for them and comparisons between them. A portable abstraction that provides basic lightweight thread control and hardware synchronization primitives as well as a sense of data locality is needed. The qthread abstraction enables development of large-scale multithreaded applications on commodity and experimental parallel architectures. The combination of a powerful portable locality and threading interface with adaptive distributed data structures provides reliable performance across multiple parallel architectures.

CSRI POC: Sue Kelly, (505) 845-9770