Algorithm model and execution based on Petri Nets in an heterogeneous parallel computer
Abstract
Multicore - MultiGPU systems are frequently used in supercomputers design. The heterogeneity between both types of processors is a source of problems for the parallel programming: disparity in processing throughput and memory availability. While some problems are faster executed in a GPGPU, when its data size exceeds the memory available,
data partition must to be done in order to resolve, and become desirable to use both types of processors. In this paper we present a solution based on Petri Nets to model the algorithm and to guide the execution, balancing the load between the CPUs cores and GPGPUs. The matrix multiplication algorithm is used as testbed. Tests confirm the goodness
of the model and highlight the difficulties to address the problem.