Workshop: Workshop on High Performance Fabrics for AI and HPC
Authors: Akshita Vyas, Rakesh Kumar Yadav, Yogeshwar Sonawane, and Sanjay Wandhekar (Centre for Development of Advanced Computing)
Abstract: As HPC systems scale, the physical connectivity of
the underlying network plays an increasingly significant role in
application performance. The effectiveness of a parallel
application is strongly influenced by how well process placement
reflects the network topology.
This paper presents a methodology designed to enable
topology-aware process placement on Trinetra-A, a high-speed
switchless 3D torus network developed by C-DAC, highlighting
the key considerations and challenges involved in the process. Our
approach involves application communication analysis and a
process mapping framework that leverages existing process
mapping tools to implement topology-aware process placement.
We evaluate the methodology using MPI benchmarks and a
real application, achieving peak latency improvements of 89% for
MPI_Alltoall, 68% for MPI_Alltoallv and 8% for MPI_Allgather,
along with a 68% improvement in heFFTe application
performance. These results demonstrate the effectiveness of
optimized placement on Trinetra-A and provide a reference that
can be readily adopted by HPC network developers.
Back to Workshop on High Performance Fabrics for AI and HPC Archive Listing
Back to Full Workshop Archive Listing