Toward Dynamic Communication Runtime for Scalable Irregular Parallel Computing
Abstract: The irregular computing model is being utilized by a number of emerging applications in domains such as chemistry, bioinfomatics, and data analytics. Unlike the well-studied bulk synchronous parallel (BSP) model, the irregular model is usually characterized by unpredictable and fine-grained data access. As the most widely used communication model over distributed memory systems, MPI defines the nonblocking message passing as well as the one-sided communication for the semantic support of irregular and asynchronous data access. The runtime implementation, however, still cannot take over such data movement efficiently.
In this talk, we analyze the crucial communication issues existing in irregular computing and present the techniques for enabling a dynamic communication environment. Through the case study of a large quantum chemistry application, we discover that the performance bottlenecks in communication are usually caused by static resource assignment for communication progress and computation. As the MPI-level solution, we first present a flexible and adaptive progress resource management for complex multiphase applications. We then further exploit a low-level multiprocess execution model that enables VAS sharing with privatized variable sets. This model allows MPI to share arbitrary information among MPI tasks, thus enabling fully dynamic resource sharing and scheduling.
Bio: Min Si is an Enrico Fermi postdoctoral scholar in the Mathematics and Computer Science Division. She received her Ph.D. and M.S. in computer science from the University of Tokyo in 2016 and 2012, respectively. Min's research interests include irregular and dynamic communication models and runtime systems in high-performance computing.