This paper describes the parallel software design used to increase the speed of data collection in experiments for a polymorphic computing architecture study which uses a genetic algorithm to search for desirable instruction arrangements within the architecture. The software was designed to run on the NSF-supported Chameleon Cloud Computing Project Cluster. The software design, implementation principles, and container contents are described in detail. This software is divided into two separate suites: the Multi-Experiment Management Software and the Single Experiment Management Software to control parallelism across compute nodes and within experiments, respectively. Within the Single Experiment Management software, the control scripts, the worker threads, the parallel execution control application, and the genetic algorithm are discussed. Additionally, an execution time study of a single experiment using the Single Experiment Management Software on a single node is presented. This study shows the speed increase for this particular case versus the number of threads compares favorably to the ideal execution times. At the end, a list of lessons learned and a list of future research areas are presented.

In this study, seven types of Chameleon Cloud compute nodes (ARM64, Haswell, Skylake, Cascade Lake, Ice Lake R650, Ice Lake R750, and Zen 3) are benchmarked using an HPC software suite that is meant to study polymorphic computing architectures. This benchmark is sensitive to the compute nodes' integer, memory, and threading performance. Each of the compute nodes is evaluated for execution time, speedup, and efficiency across a sweep of worker thread counts. The ARM64 compute node is comparatively the least performing compute node by a wide margin. The best performing compute nodes were the Ice Lake and Zen 3 systems. These nodes performed roughly equally.