Scheduling stream programs with improving arithmetic unit usage on NoC-based VLIW multi-core architectures

Abstract

Stream programming model has received a lot of interest due to its naturally-exposed task, data and pipeline parallelism. Many researches concentrated on scheduling stream programs on multi-core systems. However, few of them consider the arithmetic unit utilization, which is a vital factor to determine the performance of multi-core systems. This paper focuses on scheduling stream programs on NoC-based VLIW multi-core architectures, aiming at improving the performance through increasing the arithmetic unit utilization. Three phases are proposed for the scheme. First, the stream program is replicated into multiple threads for providing enough parallel kernels. Second, parallel kernels are grouped and operators of each kernel group are scheduled together for high arithmetic unit utilization. Third, a hierarchical integer linear programming (ILP)-based methodology is proposed to map kernel groups onto each core for optimizing the maximum workload. A set of benchmarks are exploited for evaluation. Experimental results show that, compared with two other existing scheduling schemes, our proposed scheme can significantly improve the performance of the multi-core processor.