Login
Research at the Institute of Data Analysis and Visualization
Jump to:
Multi-GPU Volume Visualization using MapReduce

Jeff A. Stuart, Cheng-Kai Chen, Kwan-Liu Ma, and John D. Owens


Abstract

Image We present a novel MapReduce library, GPMR, that leverages the power of GPU-based clusters for large-scale computing. To better utilize GPU resources, we modify the MapReduce paradigm by combining large amounts of individual map and reduce items into chunks, using partial reductions, and using accumulation. We use persistent map and reduce tasks and stress individual aspects of GPMR by porting a suite of MapReduce benchmarks (Word Occurrence, Sparse Integer Occurrence, Linear Regression, Matrix Multiplication, and K-Means Clustering) to GPMR. We execute these benchmarks using a cluster of sixty-four NVIDIA GPUs among sixteen compute nodes and achieve desirable speedup and efficiency for all benchmarks. We compare our implementation to the current best GPU MapReduce library (which runs only on a single GPU) and a highly-optimized multi-core implemention of MapReduce to show the power of optimized multi-GPU MapReduce. We demonstrate how typical MapReduce applications are easily modified to fit into GPMR and thus effectively leverage a cluster of GPUs. We highlight how the ratio of computation between different phases of the pipeline, as well as the ratio between communication and computation, affects results from GPMR. We conclude with an exposition on the types of MapReduce tasks well-suited to GPMR, and why some tasks require more modification than others to work well with the GPU.