ACM SRC: Mapping Applications on Irregular Allocations
International Conference for High Performance Computing, Networking, Storage and Analysis (SC) 2016
Publication Type: Poster
Repository URL: http://sc16.supercomputing.org/sc-archive/src_poster/src_poster_pages/spost152.html
Thumbnail
TextRef
Topology-aware mapping of applications on clusters and supercomputers becomes more difficult as the number of nodes in the systems increase and interconnection networks become more complex. To facilitate this process, Rubik [1] was proposed to generate mappings that lead to good performance on allocations that are symmetric, compact, and convex, e.g. allocations on most Blue Gene/Q systems. However, many supercomputers provide irregular, asymmetric, and disjoint allocations to users for better utilization of resources. It is significantly more difficult to map applications on these allocations because of the lack of structure in the allocations and unavailability of certain nodes which are reserved for other work, e.g. IO servers. In this poster, we extend Rubik to provide mapping support for these complex cases by using different heuristics to project virtual machine topologies onto real machine topologies. We evaluate our work using two widely used HPC applications, namely MILC and Qbox, on Blue Waters, a Cray XE supercomputer. We show that, for these communication intensive applications, MPI time is reduced by 60% in MILC and 56% in Qbox when running on 16,384 ranks using the proposed extension to Rubik.
People
Research Areas