Optimal Subset Mapping And Convergence Evaluation of Mapping Algorithms for Distributing Task Graphs on Multiprocessor SoC

Mapping an application on multiprocessor system-on-chip (MPSoC) is a crucial step in architecture exploration. The problem is to minimize optimization effort and application execution time. Applications are modeled as static acyclic task graphs which are mapped to an MPSoC. The analysis is based on extensive simulations with 300 node graphs from the standard graph set. We present a new algorithm, optimal subset mapping (OSM), that rapidly evaluates task distribution mapping space, and then compare it to simulated annealing (SA) and group migration (GM) algorithms. OSM was developed to make architecture exploration faster. Efficiency of OSM is 5.0 times and 2.4 times than that of GM and SA, respectively, when efficiency is measured as the application speedup divided by the number of iterations needed for optimization. This saves 81% and 62% in wall clock optimization time, respectively. However, this is a tradeoff because OSM reaches 96% and 89% application speedup compared to GM and SA. Results show that OSM and GM have opposite convergence behavior and SA comes between these two.