Parameter optimization for high performance computing systems using genetic algorithm

Abstract

Though Von Neumann Architecture (vNa) has had unprecedented success for general- purpose computing for the past seven decades. While being heavily used in High Performance Computing (HPC) systems, vNa is becoming unpopular among industry professionals due to vNa bottleneck and unable to address the ever-increasing demand for performance gain by latest HPC applications. Though there are few methodologies to overcome vNa bottleneck; namely Processor caching, Prefetching, multithreading, Processing in Memory (PIM) and RAMBUS, which we will not discuss in this paper, the latency between ALU and memory still hold back the performance of vNa systems, even in HPC environments. Alternatively, there are few different technologies are emerging as opposed to vNa such as Quantum computing, Parallel computing, Harvard Computing architecture (HCA) which are addressing the vNa bottleneck and taking computing performance in to a whole new level in terms of performance in Floating Points (Flops). On the other hand, another alternative is to use accelerators like Graphic Processing Units (GPU, Field Programmable Gate Arrays (FPGA) and Dell Process Acceleration Tool (DPAT) to overcome the latency between ALU and Memory subsystem. Even though there are several alternatives available to overcome vNa bottleneck, the performance requirement from almost every HPC application seems never to be achieved by the underlying HPC hardware systems. Hardware system configurations are always performed at a sub-optimal level, and no one has the time or patience to go through thousands of different permutations of hardware parameter finetuning. Moreover, we simply cannot blame any system administrator who is responsible for HPC system configurations, especially processor, memory, and network parameter finetuning to provide best performance benchmarks, as it’s a very time consuming and tedious task of many man-days of manually configuring all those parameters to find an optimal convergence point. On the other hand, one cannot be satisfied of a particular combination of parameters is the peak performance point as there always be a doubt whether there can be another set of combination exist for the optimal performance. Thus, it will be a never- ending manual task to find the optimal performance of HPC hardware systems. To overcome HPC hardware parameter finetuning to find the best performance benchmarks, we proposed use of Genetic Algorithm based approach to find the best combination of Processor, memory, and network parameters to deliver highest hardware performance for any HPC application