CSE Seminar: Communication for HPC, by Dan Ibanez
Title: Communication for HPC
Speaker: Dan Ibanez
When: 4-4:45 PM, 13th Sept.
Where: CII 3206
The Parallel Control Utility provides an simple user interface for communication on massively parallel computers that handles complex communication patterns with optimal performance even for multi-threaded programs. Simple yet novel methods including non-blocking collective algorithms and inter-thread communication primitives enable new capabilities in parallel applications. First, an algorithm is provided which handles a collective communication where the pattern can be any general graph, also called phased communication. This interface is completely non-blocking, maximizing computation and communication overlap. Second, an implementation based on abstract primitives allows applications to instantly be ported to a hybrid MPI/thread environment, and to progressively take advantage of shared memory optimizations.
This presentation will describe the algorithm for phased communication and compare it to similar state-of-the art algorithms. We will also present the non-blocking collective framework which encompasses algorithms such as reductions, broadcasts, and scans. Finally, we present the implementation of point-to-point communication primitives which allow all the above algorithms to function as inter-thread communication on the latest manycore architectures, which provides a clear way forward for adoption of emerging hybrid distributed and shared memory programming. The usefulness of these systems will be demonstrated by case studies selected from unstructured mesh procedures at scale.