Run-time configuration prefetching to reduce the overhead of dynamically reconfiguration

Abstract

Reconfigurable computing is a promising approach with both flexibility and efficiency for high performance computing. However, the overhead of run-time reconfiguration affects the performance severely. In the paper, we develop a simple and effective method to hide the latency of configuration by configuration prefetching at runtime. A simulation platform based on Simics is developed for evaluation. The experimental results show that the predictive accuracy is rather high, the hit rate of reconfigurable processing unit is increased by 24.6%~53.7% when reconfigurable resource is not adequate. I. INTRODUCTIONReconfigurable computing (RC) has been proven to be an effective manner of high performance computing. However, the overhead of run-time reconfiguration affects the performance severely, and becomes one of the main challenges of RC.Configuration prefetching (CP) [2] is proposed to reduce the cost of run-time reconfiguration, but the policy of CP and the strategy combining with scheduling are still challenging.In many previous works, CP is considered to optimize scheduling. Scheduling heuristic at runtime and hybrid design/run-time prefetch heuristic are proposed. Also parallel configuration models are designed with prefetch. However, run-time heuristic scheduling generates an excessive runtime penalty and the hybrid one need design-time preprocesses which is not convenient in practice.This paper proposes a run-time configuration prefetch scheduling to reduce the overhead of dynamically reconfiguration. Without the support of compilers or programmers, predictor can predict the next tasks run-time. According to the predictor, the scheduler can reduce or even eliminate the overhead of run-time reconfiguration and replacements of RPUs will be more reasonable.