Linear algebra (LINPACK) benchmarks of the UP Data Commons hyperconverged infrastructure

Abstract

Essential for evaluation of readiness for high-performance computing (HPC) utility is the testing for a facility's speed in common computing tasks. We evaluate and analyze the HPC performance of the UP Data Commons (UPDC), a research facility of the University of the Philippines (UP) for data hosting and high performance computing (HPC). The floating point operations per second (flop/s) is measured using the widely used Highly-Parallel LINPACK (HPL) benchmark. Since publicly available benchmark scores are often from top-ranking facilities having significantly higher specifications (i.e. TOP 500), a direct comparison of maximum scores has limited utility to smaller server facilities. Hence, for more fair comparisons, we extrapolate from a series of measurements at different CPU core counts. Our tests with the hyperconverged infrastructure (HCI) cluster show favorable results at a lower (n ≤ 30) number of cores, n. However, as n is increased, a relative drop in performance is observed after a highest score of 1638 Gflop/s at n = 34 cores using the Intel MKL HPL benchmark. Extrapolations indicate that the HCI performs at around 2% to 10% compared to top-ranking supercomputers. Recommendations are explored for possible maximization of actual computational performance to avoid the effect of data transfer between cores during computations.