Iterative sparse matrix-vector multiplication on in-memory cluster computing accelerated by GPUs for big data
Authors:
- 24 October 2016
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1454-1460
Abstract
Iterative SpMV (ISpMV) is a key operation in many graph-based data mining algorithms and machine learning algorithms. Along with the development of big data, the matrices can be so large, perhaps billion-scale, that the SpMV can not be implemented in a single computer. Therefore, it is a challenging issue to implement and optimize SpMV for large-scale data sets. In this paper, we used an in-memory heterogeneous CPU-GPU cluster computing platforms (IMHCPs) to efficiently solve billion-scale SpMV problem. A dedicated and efficient hierarchy partitioning strategy for sparse matrices and the vector is proposed. The partitioning strategy contains partitioning sparse matrices among workers in the cluster and among GPUs in one worker. More, the performance of the IMHCPs-based SpMV is evaluated from the aspects of computation efficiency and scalability.Keywords
This publication has 17 references indexed in Scilit:
- GraphX2013
- HEigen: Spectral Analysis for Billion-Scale GraphsIEEE Transactions on Knowledge and Data Engineering, 2012
- Accelerating MapReduce on a coupled CPU-GPU architecture2012
- Multi-GPU MapReduce on GPU Clusters2011
- Mars: Accelerating MapReduce with Graphics ProcessorsIEEE Transactions on Parallel and Distributed Systems, 2010
- PEGASUS: A Peta-Scale Graph Mining System Implementation and Observations2009
- Implementing sparse matrix-vector multiplication on throughput-oriented processors2009
- Mars2008
- Spectral grouping using the nystrom methodIEEE Transactions on Pattern Analysis and Machine Intelligence, 2004
- Sparse matrix solvers on the GPUACM Transactions on Graphics, 2003