BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20111115T190000Z
DTEND:20111115T193000Z
LOCATION:TCC 303
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: Graphics processors (GPUs) have emerged as an important computing platform for general-purpose and high-performance computing, because they offer a large number of parallel cores and high memory bandwidth. However, data layouts in GPU memory often lead to sub-optimal performance for programs designed with a CPU memory interface in mind. This implies that application performance is highly sensitive to architecture and memory access patterns. This paper proposes a simple API, Dymaxion, that allows programmers to declare access patterns.  The runtime then optimizes memory layouts accordingly, hiding the latency by overlapping it with inter-device transfers; and then transforms subsequent memory access addresses as necessary.  We show that applications with a variety of access patterns achieve significant speedups with Dymaxion, and also explore the importance of maintaining per-device data layouts and cross-device data mappings with a case study of concurrent CPU-GPU execution.
SUMMARY:Dymaxion: Optimizing Memory Access Patterns for Heterogeneous Systems
PRIORITY:3
END:VEVENT
END:VCALENDAR