Project: Shared Memory Communication on Manycore compute nodes

Team: Prof. Dr.-Ing. Rainer Keller

Research institution: Hochschule für Technik Stuttgart

Abstract: The current technology trend in High-Performance Computing leads to large installations employing not Multi- but Manycore nodes, with each node consisting of highly parallel architectures. While currently Graphics-Processing Units (GPUs) are currently much investigated, the traditional CPU architectures are catching up. We therefore may expect compute nodes of a few hundred compute cores plus compute power from additional threaded GPGPUs or MICs. The main parallel programming paradigm for coarse-grained parallelization however will remain using an evolving Message Passing Interface (MPI). A lot of work has been put into the Open MPI implementation from various research institutions: Indiana University (IU), Oak Ridge National Lab (ORNL), High Performance Computing Center Stuttgart (HLRS) and others, in develop an Open Source MPI implementation for largescale installations. This project will investigate the different communication strategies within a compute node (and possibly settings for inter-node communication depending on network availability). The intra-node communication strategies investigated: using different Linux-APIs for memory-mapping (mmap), synchronization via the kernel, and the newly added Linux-3.2 System calls process_vm_readv and process_vm_writev. Especially the latter calls may be very beneficial for splitting communication from synchronization on these many-core architectures, where memory-locality is a prime requirement for performance: using HPI’s Future SOC lab with its 40-core Intel and 64 core NUMA compute nodes, we will be able to analyze the sweet-spot to switch communication between the various shared-memory methods. We expect these findings to be of importance to other multi-process, communication-intensive applications, such as SAP’s Hana Database or File- System daemons. The results will be gathered in the final report and as poster.

Last modified 8 years ago Last modified on Apr 12, 2013 6:13:10 PM