RSA Implementation on a Hetrogeneous Cluster Architecture using CPUs, GPUs, and MIC

There is a great degree of difficulty in addressing problems of significant computational complexity such as the Routing and Spectrum Assignment (RSA) problem on heterogeneous clusters comprising CPUs, CUDA GPUs and Intel Many Core (MIC) architecture. A scheduling approach is employed in the present study to address the RSA issue on a heterogeneous cluster. Such an approach relies on the speed of heterogeneous architectures for task planning. More specifically, the study seeks to use appropriate parallel computing paradigms (e.g. MPI, OpenMP, CUDA) on separate architectures to address the problem and afterwards refer to the speed of task solving to determine how many tasks each architecture should be allocated. Findings show that the proposed approach is more effective than pure CPU-based application in terms of overall execution time, as made apparent by relative comparison of the number of nodes possessed by each architecture. Furthermore, as very high CPU numbers can remove the effect of employing architectures of high speed, there is an indirect correlation between the speedup and the elevated number of CPUs. On the other hand, system use and, implicitly, the number of tasks that can be introduced in the system at the same time are affected by the use of a high number of CPUs in a single task to attain a notable speedup factor. Therefore, code amendment for the purpose of task allocation to the architectures is discussed by the study as well