BEGIN:VCALENDAR PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN VERSION:1.0 BEGIN:VEVENT DTSTART:20151117T203000Z DTEND:20151117T210000Z LOCATION:19AB DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: With the increased popularity of multi-GPU nodes in modern HPC clusters, it is imperative to develop matching programming paradigms for their efficient utilization.=0AIn order to take advantage of the local GPUs and the low-latency high-throughput=0Ainterconnects that link them, programmers need to meticulously adapt parallel applications with respect to load balancing, boundary conditions and device synchronization.=0AThis paper presents MAPS-Multi, an automatic multi-GPU partitioning framework =0Athat distributes the workload based on the underlying memory access patterns. The framework consists of host- and device-level APIs that allow programs to efficiently run on a variety of GPU and multi-GPU architectures.=0AThe framework implements several layers of code optimization, device abstraction, and automatic inference of inter-GPU memory exchanges.=0AThe paper demonstrates that the performance of MAPS-Multi achieves near-linear scaling =0Aon fundamental computational operations, as well as real-world applications in deep learning =0Aand multivariate analysis. SUMMARY:Memory Access Patterns: the Missing Piece of the Multi-GPU Puzzle PRIORITY:3 END:VEVENT END:VCALENDAR