Communication-Aware Thread Scheduling for MPI

Event Sponsor: 
Mathematics and Computer Science Division Seminar
Start Date: 
Aug 11 2016 - 10:30am
Building/Room: 
Building 240/Room 4301
Location: 
Argonne National Laboratory
Speaker(s): 
Hoang-Vu Dang
Speaker(s) Title: 
University of Illinois at Urbana-Champaign
Host: 
Sangmin Seo and Pavan Balaji

We present in this work the issue of thread scheduling in the MPI implementations. We show that the excessive thread interactions inside the runtime are hindering the effectiveness of the hybrid MPI+Threads programming model. We show that by managing the thread scheduler using the information from the communication layer, one can achieve tremendous performance benefit. Our techniques, implemented atop MPICH 3.2, achieve upto 20-fold improvement for a set of thorough MPI+Threads performance benchmarks and kernels.

Bio:
Hoang-Vu Dang is a Ph.D. candidate at University of Illinois at Urbana-Champaign (UIUC) under the supervision of Prof. Marc Snir, and is a research aide in the Programming Models & Runtime Systems group at Argonne National Lab for a summer project. He currently studies the interactions between thread schedulers and communication runtimes to devise the necessary interpolations that are required for the performance of hybrid parallel applications. Before coming to UIUC, he has involved in several High Performance Computing projects at research centers in Singapore and Germany. His publications and experiences span on designing algorithms and runtime systems with a focus on parallel computations.