Title: Application and Operating System Software Challenges in the Multicore Era Speaker: Doug Doerfler and Kevin Pedretti, and Ron Brightwell, Sandia National Laboratories Date/Time: Tuesday, April 8th, 2008, 11:30am – 1:00pm (talk starts at 12:00pm) Location: CSRI Building, Room 90 (Sandia NM) Brief Abstract: The going in position for programming new multicore architectures is MPI everywhere, i.e. one MPI task per core. It's a straight forward and for the most part pretty effective solution. It also allows a direct porting path for all MPI only applications. However, this approach may not be the most efficient as not all applications scale with the number of MPI tasks per socket. This can be due to task placement on the processor and the extra burden of MPI communications overhead on the already stressed out memory subsystem. This seminar will present recent research into techniques and optimizations at the OS and MPI layers to improve intranode performance on multicore architectures. CSRI POC: Heidi Thornquist, (505) 284-8426, Todd Coffey, (505) 845-0520, |