Congresso de M´ etodos Num´ ericos em Engenharia 2015 Lisboa, 29 de Junho a 2 de Julho 2015 c APMTAC, Portugal 2015 PERFORMANCE ANALYSIS OF A PARTICLE-IN-CELL PLASMA PHYSICS CODE ON HOMOGENEOUS AND HETEROGENEOUS HPC SYSTEMS Xavier S´ aez 1∗ , Alejandro Soba 2 , Edilberto S´ anchez 3 , Mervi Mantsinen 1,4 and Jos´ e M. Cela 1 1: Barcelona Supercomputing Center (BSC-CNS) C. Gran Capit` a 2-4, 08034 Barcelona, Spain e-mail: xavier.saez, mervi.mantsinen, josem.cela@bsc.es, web: http://www.bsc.es 2: Centro de Simulaci´ on Computacional para Aplicaciones Tecnol´ ogicas. (CSC-CONICET) Godoy Cruz 2390, CABA, Argentina e-mail: soba@cnea.gov.ar 3: Laboratorio Nacional de Fusi´ on (CIEMAT) Avenida Complutense 40, 28040 Madrid, Spain e-mail: edi.sanchez@ciemat.es, web: http://fusionsites.ciemat.es 4: Instituci´ o Catalana de Recerca i Estudis Avan¸ cats (ICREA) Pg Llu´ ıs Companys 23, 08010 Barcelona, Spain web: http://www.icrea.cat Keywords: Particle-in-Cell Plasma, Heterogeneous HPC systems, GPU, ARM Abstract. PIC methods are one of the most used methods in plasma simulations. We present a comprehensible evaluation of the PIC code performance on four current parallel platforms: IBM PowerPC, Intel Nehalem (SMP), Intel Sandy Bridge (SMP) and ARM GPU. The behavior of computational algorithms and data structures are analyzed to deduce which code optimizations will make the best use of each platform. 1 INTRODUCTION During the last few decades, high-performance computing (HPC) has been dominated by the rapid scaling of the CPU clock frequency (c.f. Moore’s Law) (figure 1). Currently, the performance of next-generation supercomputers is limited by the power efficiency and, as a result, several novel hardware designs have emerged. Originally, many scientific codes were not developed for present designs of supercom- puters based on multi-core and heterogeneous architectures. Basically, they were only leveraging task level parallelism through message passing models such as MPI. Unfortu- nately, a hand-tuning of these codes is often required to exploit the modern platforms capabilities. 1