84. Energy and Communication Efficient Partitioning for Large-Scale Finite Element Computations
Authors: Milinda Fernando (University of Utah)Dmitry Duplyakin (University of Colorado, Boulder)Hari Sundar (University of Utah)
Abstract: Load balancing and partitioning are critical when it comes to parallel computations. Generally partitioning involves equally dividing the work and data among the processors, reducing processor idle time and communication costs. As we march toward exascale machines, the cost of data movement and load-imbalances therein are a major bottleneck for achieving scalability. We propose an alternative Space Filling Curve (SFC)-based partitioning scheme where we allow some (user-specified) flexibility in the work assignment, so as to minimize the data-dependencies across partitions. Effectively, we show that the flexibility in SFC based partitioning schemes leads towards minimizing the communication load-imbalance (Hilbert- 4.9% and Morton- 12.18%) and overall energy consumption (Hilbert- 22.0% and Morton-5.0%) at the cost of a marginal increase in workload-imbalance. The traditional SFC-based partitioning can be recovered by setting the flexibility to zero.
Two-page extended abstract: pdf