Dtree: Dynamic Task Scheduling at Petascale

Abstract

Irregular applications are challenging to scale on supercomputers due to the difficulty of balancing load across large numbers of nodes. This challenge is exacerbated by the increasing heterogeneity of modern supercomputers in which nodes often contain multiple processors and coprocessors operating at different speeds, and with differing core and thread counts. We present Dtree, a dynamic task scheduler designed to address this challenge. Dtree shows close to optimal results for a class of HPC applications, improving time-to-solution by achieving near-perfect load balance while consuming negligible resources. We demonstrate Dtree’s effectiveness on up to 77,824 heterogeneous cores of the TACC Stampede supercomputer with two different petascale HPC applications: ParaBLe, which performs large-scale Bayesian network structure learning, and GTFock, which implements Fock matrix construction, an essential and expensive step in quantum chemistry codes. For ParaBLe, we show improved performance while eliminating the complexity of managing heterogeneity. For GTFock, we match the most recently published performance without using any application-specific optimizations for data access patterns (such as the task distribution design for communication reduction) that enabled that performance. We also show that Dtree can distribute from tens of thousands to hundreds of millions of irregular tasks across up to 1024 nodes with minimal overhead, while balancing load to within 2 % of optimal.

X. Liu—During this research, Xing Liu was affiliated with Georgia Institute of Technology.