Fast and Generic Concurrent Message-Passing

Student: Hoang-Vu Dang (University of Illinois)
Advisor: Marc Snir (University of Illinois)

Abstract: Communication hardware and software have a significant impact on the performance of clusters and supercomputers. Message-passing model and the Message-Passing Interface (MPI) is a widely used model of communications in the High-Performance
Computing (HPC) community. However, MPI has recently faced new challenges due to the emergence of many-core architecture and of programming models with dynamic task parallelism, assuming a large number of concurrent threads. These applications come from important classes of applications such as graph and data analytics.

In this thesis, we studied MPI under the new assumptions. We identified several factors in the standard which were inherently problematic for scalability and performance. Next, we analyzed graph, threading and data-flow frameworks to understand the communication. We then proposed a communication engine (LCI) targetting these frameworks. Our thesis benefits MPI by developing several patches in production MPI. Furthermore, LCI presents a simple and ground-up design which benefits various frameworks of study.

Summary: pdf
Thesis Canvas: pdf



Back to Doctoral Showcase Archive Listing