- Log In
Resources
Many-core architectures such as Intel Knights Landing (KNL) are seeing widespread adoption in current and next-generation supercomputing systems due to their power/performance ratio. However, this increased density of the compute nodes and the performance characteristics of the new architecture bring in a new set of challenges that must be tackled to extract the best performance. In this work, we present some of the advanced designs to tackle such challenges in the MVAPICH2 MPI library on the KNL architecture. In particular, we focus on the following aspects --- a) how MVAPICH2 achieves fast and scalable startup on the KNL+Omni-Path architecture, b) contention-aware, kernel-assisted designs for large-message intra-node collectives, and c) designs for scalable reduction operations on different message sizes. We also compare the proposed designs against other state-of-the-art MPI libraries such as Intel MPI and OpenMPI. Experimental evaluations show that the proposed designs offer significant improvements in terms of time to launch large-scale jobs, performance of intra-node and inter-node collectives, and performance of applications.
Event Name
IXPUG Annual Meeting 2016
Keywords
ixpug
Video Name
NA