Probabilistic Performance Modelling when using Partial Reconfiguration to Accelerate Streaming Applications with Non-Deterministic Task Scheduling

Bruno da Silva, An Braeken, Abdellah Touhafi

Many streaming applications composed of multiple tasks self- adapt their tasks’ execution at runtime as response to the processed data. This type of application promises a better solution to context switches at the cost of a non-deterministic task scheduling. Partial reconfigura- tion is a unique feature of FPGAs that not only offers a higher resource reuse but also performance improvements when properly applied. In this paper, a probabilistic approach is used to estimate the acceleration of streaming applications with unknown task schedule thanks to the appli- cation of partial reconfiguration. This novel approach provides insights in the feasible acceleration when partially reconfiguring regions of the FPGA are partially reconfigured in order to exploit the available re- sources by processing multiple tasks in parallel. Moreover, the impact of how different strategies or heuristics affect to the final performance is included in this analysis. As a result, not only an estimation of the achievable acceleration is obtained, but also a guide at the design stage when searching for the highest performance.