A Power Characterization and Management of GPU Graph Traversal


Published at the Fourth Workshop on Architectures and Systems for Big Data (ASBD 2014), June, 2014 (acceptance rate: 6/13 ≈ 46%)


Adam McLaughlin, Indrani Paul, Joseph L. Greathouse, Srilatha Manne, Sudhakar Yalamanchili


Graph analysis is a fundamental building block in numerous computing domains. Recent research has looked into harnessing GPUs to achieve necessary throughput goals. However, comparatively little attention has been paid to improving the power-constrained performance of these applications.

Through firmware changes on a state-of-the-art commodity GPU, we characterize the power consumption of Breadth-First Search (BFS) as a function of the structural properties of the graph. We choose to study this algorithm since graph traversals are used as a building block for many other graph analysis applications. Based on this characterization, we propose and evaluate a power management algorithm to maximize power cap efficiency, or the performance under a fixed power cap. Across a range of benchmark graphs, we demonstrate power cap efficiency improvements averaging 15.56% on a state-of-the-art GPU.




PPTX | PPT | PDF Copyright © ACM 2014. This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in ASBD 2014.