Dr.Teng Ma , a 6th year PhD student in EECS Department at University of Tennessee, Knoxville, TN, gave a report on Kernel-assisted and Topology-aware MPI Collective Communication among Multicore or Many-core Clusters, on June 5, in the Conference Room 1048. This talk was invited and hosted by Associate Researcher Guangming Tan at intelligent information processing key laboratory.
Dr. Teng Ma introduced a kernel-assisted topology-aware collective framework, and how this framework orchestrates the collaboration between multiple layers of collective algorithms. The resulting scheme enables perfect overlap of intra-and inters- node communications.
Firstly, Dr. Teng Ma gave an introduction to project and research work of their laboratory—Innovative Computing Laboratory (ICL). Then he talked about issues in current MPI implementations: mismatch between static collective topologies and hardware topologies, lack of efficient intra-node message delivering approaches and scalability issue. After that, he talked about current existing methods and his method in optimization of collective operations, and then he presented his work --HierKNEM Collective: a kernel-assisted and topology-aware collective on distributed multicore clusters. At last, he presented the performance comparison of his work and related work; it shows significant improvement over scientific applications.
After the report, attendees discussed about the details of mechanisms to reduce communication overhead on multicores, and other relative issues.