This presentation is intended for the new comer users of ALCF Sunspot and Aurora systems. It is expected that the users have prior experience on large ALCF and non-ALCF machines and familiar with general workflow of setting login and compute node environment, compilation, interaction with scheduler, and project allocation details. I will focus on specifics of messaging software and present how to choose the right messaging software stack, discuss the affinity and process placements to make sure that the cores, GPUs, and NICs were interacting over the shortest path with maximal efficiency. Intel has made significant contribution to improved interaction of GPUs with the host system within an MPI context, and we will discuss how to use those improvements. Some time will be spent on optimized collectives. Familiarity with Aurora hardware specification is desirable, but not required for this event.
Vitali Morozov is the Software Engineering 5 at the ALCF. He received his B.S. in Mathematics and MS in Computer Science from Novosibirsk State University, a Ph.D. in Engineering from Ershov’s Institute for Informatics Systems, Novosibirsk, Russia. At Argonne since 2001, he has been working on computer simulation of plasma generation, plasma material interactions, plasma thermal and optical properties, and applications to laser and discharge-produced plasmas. At the ALCF, he has been working on performance projections, performance analysis and simulation, studying the hardware trends and evaluates experimental and non-conventional hardware. He is also porting and tuning applications to large-scale supercomputers.
