Private Cloud Tuning for Efficient Inter-Server Simulation Execution

The rise of cloud popularity, availability, and ever-increasing computing power allows for novel research applications in simulation science. Moreover, with unmatched accessibility and private cloud modifications, scientists could tailor their computing environments to meet the individual project requirements. However, hardware virtualization introduces substantial overheads with each new interface in the communication stack. Also, traditional switching networks, due to their inherent latencies, strongly limit cloud execution efficiency. Therefore, to answer these issues, one must first recognize different cloud communication layers and then address them accordingly. In this paper, we present the enhancement of the weakest cloud simulation layer, the Inter-Server communication. For this purpose, we have selected a micromagnetic solver called Magpar and renewed its obsolete parallelization libraries with the latest MPI-3 standard. Besides communication improvement, this modification allows another necessary tuning, described as MPI process distribution planning. With a novel approach to CPU load balancing and adequate interconnect hardware integration, we achieved the execution time reduction from 57 percent to several orders of magnitude, depending on the parallelization level. Finally, we managed to produce a near-linear increase of simulation speedup with each increment of the parallel nodes, which is a remarkable result considering Inter-Server latency limitations.