An Accurate and Fast Trace-aware Performance Estimation Model For Prioritized MPSoC Bus With Multiple Interfering Bus-Masters

Abstract

Accurate and fast performance estimation methods for modern and future multi-core systems are the focal point of much research due to the complexity associated with such architectures. The communication architecture of such systems has a huge impact on the performance and power of the whole system. Architects need to explore many design possibilities by using performance estimation techniques at early stages of design to make design decisions earlier in the design cycle. While software developers need to develop and test applications for the target architecture and gather performance measurements as early in the design cycle as possible. Full system simulation techniques provide accurate performance values but are extremely time consuming. Static analysis techniques are fast but cannot capture the dynamic behavior associated with shared resource contention and arbitration. Moreover, synthetic traffic patterns have been used to analyze the communication architecture however, such patterns are not realistic enough. We propose a statistical based model to predict the dynamic cost of bus arbitration on the performance of a bus architecture. The proposed model uses workload trace of the actual applications and benchmarks to capture the real application traffic behavior. Statistics on the traffic patterns are collected and input to the analytical model which calculates performance values for the communication architecture under consideration. By knowing the performance measures, designers can avoid over and under-design of the communication architecture. This paper builds up on a previously developed performance estimation model. The previous work modeled single and burst bus-transfers, however only one interfering bus master at a time for each blocked bus request was considered. The proposed, improved accuracy model considers multiple interfering masters for each blocked request hence improving the estimation accuracy especially for traffic intensive applications and many PE architectures. Experiments are performed for two different architectures i.e., 4 processing elements connected via a shared bus and 8 processing elements connected via a shared bus. Results show no significant difference in accuracy compared to previously developed model, for low traffic applications SPARSE and ROBOT however notable accuracy improvement for traffic intensive applications. Maximum estimation error is reduced from 1.75% to 0.6% for FPPPP and from maximum 13.91% to 8.8% for FFT on the 4PE architecture. On the 8PE architecture, maximum estimation error is reduced from 11.8% to 2.7% for the FPPP benchmark. Moreover simulation speed-up for the proposed technique over simulation method is reported.