Layer Based Partition for Matrix Multiplication on Heterogeneous Processor Platforms

• DOI: 10.48550/arxiv.1812.06329
• Publication Date: 2018-01-01
• Authors: Yang Liu; Shi Li; Junwei Zhang; Thomas G. Robertazzi
• Keywords: Partition problem; Speedup

While many approaches have been proposed to analyze the problem of matrix multiplication parallel computing, few them address on heterogeneous processor platforms. It still remains an open question platforms find optimal schedule that balances load within set while minimizing amount communication. A great studies are based rectangular partition, whereas optimality partition as basis has not well justified. In this paper, we propose a new method schedules with mixed co-design goal total communication volume and completion time. We first present schema our layer (LBP) method. Subsequently, demonstrate approach guarantees minimal volume, which is smaller than what can reach. further task time, network topologies taken into account. solve in both single-neighbor case multi-neighbor case. cases, equality LBP, simulation shows reduced by 75% from lower bound partition. formulate LBP Mixed Integer Programming problem, reduce 81% through simulation. To summarize, promising perspective tackling problems