Author:
Paznikov A.A.,Kurnosov M.G.
Abstract
Leveraging graph partitioning techniques is a fundamental framework. This research paper presents an innovative method accompanied by heuristic algorithms designed for the mapping of parallel MPI-programs for hierarchical HPC systems. The proposed approach aims to enhance performance by distributing highly interactive processes to processor cores interconnected through high-speed communication channels. Notably, our method takes into meticulous account all hierarchical layers within the communication network of the HPC systems. Additionally, we offer empirical insights derived from the mapping of MPI programs extracted from the SPEC MPI and NAS Parallel Benchmarks collections into a geographically distributed multi-cluster HPC system.
Reference18 articles.
1. Kurnosov M., Paznikov A., Efficiency analysis of decentralized grid scheduling with job migration and replication, Proc. of the 7th Int. Conf on Ubiquitous Information Management and Communication, 1–7 (2013)
2. Broquedis F. et al., hwloc: a Generic Framework for Managing Hardware Affinities in HPC Applications, Proc. Int. Conference Parallel, Distributed and Network-Based Processing (PDP2010), 180–186 (2010)
3. Khoroshevsky V.G., Kurnosov M.G., Mapping Parallel Programs into Hierarchical Distributed Computer Systems ICSOFT, 123–128 (2009)