Multithreaded MPI communication is increasingly adopted by hybrid MPI applications for its suitability in fine-grained tasking environments and to better utilize network resources. Unfortunately, current MPI implementations still suffer from thread synchronization overheads and hinders the adoption of multithreaded communication models in production environments.  In this talk, we will revisit multithreaded MPI communication challenges and potential improvements from a *progress* perspective.  We first make the observation that MPI communication progress is tightly related to the progress properties (such as fairness and blocking properties) of the underlying thread safety implementation. We follow with an in-depth look at pure blocking lock methods and the effect of altering lock progress properties on MPI communication. We then introduce several variations from this model that offer new progress properties and improvement opportunities. We conclude with the remaining challenges and unexploited capabilities that are being investigated to further improve communication performance of hybrid MPI applications.

Biography:
Dr. Abdelhalim Amer holds a Postdoctoral Researcher appointment at Argonne National Laboratory. His research falls under the parallel and distributed computing landscape. More specifically, he focuses on parallel runtime systems and programming models targeting applications in massively parallel environments. This includes tackling scalability challenges in shared-memory as well as distributed-memory parallel environments. Dr. Amer is a co-PI of two projects and authored several papers in this research area and presented numerous talks in conferences and institutions.  Dr. Amer graduated from Tokyo Institute of Technology in 2015 where he is the recipient of the Seiichi Tejima Doctoral Dissertation Award for his Ph.D work. He has served as a chair and as a technical program committee in numerous conferences and workshops and as a technical referee for several international journals. Furthermore, Dr.  Amer is a professional member of IEEE and ACM, and a member of SIGHPC.