Performance estimation in heterogeneous MPSoC based on elementary operation cost

摘要

Effective use of resources available on heterogeneous MPSoC platforms can only be achieved through careful resource allocation and scheduling. The diversity of processing and memory elements will manifest itself in the total time and resources required to perform a task or execute an application. Choosing the right platform element is the key and the first step is performance estimation. This paper tackles the issue of finding the most suitable processing element for each part of the software application through a novel approach - elementary operation cost. The cost of each elementary operation is experimentally determined through a set of carefully devised benchmarks and is used for estimating duration of complex functions found in common applications such as JPEG, AES etc. By raising the abstraction level on which the execution time is calculated from instruction to operation level, common problems in performance estimation such as pipelining and branch prediction can be avoided and estimation accuracy is improved. Demonstrated results show that the average error rate of estimated execution times for various benchmarks is around six percent compared to the actual execution times.