SRC13. Mapping Applications on Irregular Allocations
Student: Seonmyeong Bak (University of Illinois)
Supervisor: Laxmikant V. Kale (University of Illinois)
Abstract: Mapping applications on clusters becomes more difficult as the number of nodes become larger. Supercomputers assign allocations with irregular shapes to users to maximize the utilization of resources, and it is much more difficult to map applications on these irregular allocations.
We extended Rubik, a python based framework to map applications on irregular allocations with a few lines of python code. Rubik was originally designed for regular allocations, so we added features to handle allocations with irregular structure and unavailable nodes and two mapping algorithms such as row-ordering and recursive splitting.
We evaluate our work with two widely used HPC applications on Blue Waters: MILC and Qbox. We reduced execution time by 32.5% in MILC, and by 36.3% in Qbox, and communication time by 60% in MILC and 56% in Qbox.
Two-page extended abstract: pdf