A Holistic Optimisation - Success Mantra for HPC Performance
Ashish Bisht, Deepika H. V, P Haribabu, Sathish Kumar, S D Sudarsan · 2023
High Performance Computing (HPC) aids in solving numerous complex scientific problems such as weather forecasting, drug discovery, physical simulations, molecular modeling, nuclear research, cryptanalysis, oil and gas exploration. To scale these kinds of applications across the nodes of a cluster or supercomputer Message Passing Interface (MPI) is used. The performance of MPI is one of the key aspects to achieve the expected speedup in the application performance on a HPC cluster, which in turn depends on the architecture of the servers and hence it is important to understand the same for HPC practitioners. Application performance is dependent not only on an applications' capability to scale but also on the efficiency of application to execute on an architecture. In this paper, we compare the performance of applications using MPI on two architectures of based on Intel and ARM. This is important as the former is a predominant architecture and the latter is one of the upcoming architectures among the supercomputers. We have used benchmarks which cover a wide variety of applications belonging to the Berkley dwarfs. This will help to analyze the comprehensive behavior of HPC applications on both the architectures. Finally, we present our observations in terms of computation, communication, data size and functional behavior of the applications.