BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20151117T200000Z
DTEND:20151117T203000Z
LOCATION:19AB
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: Graphics processing units (GPUs) are increasingly utilized=0Aas throughput engines in the modern computer systems.=0AGPUs rely on fast context switching between thousands=0Aof threads to hide long latency operations, however, they still=0Astall due to the memory operations. To minimize the stalls,=0Amemory operations should be overlapped with other operations=0A as much as possible to maximize memory-level parallelism (MLP).=0AIn this paper, we propose Earliest Load First (ELF) warp scheduling,=0Awhich maximizes the MLP by giving higher priority to the warps=0Athat have the fewest instructions to the next memory load.=0AELF utilizes the same warp priority for the fetch scheduling=0Aso that both are coordinated. We also show that ELF reveals=0Aits full benefits when there are fewer memory conflicts and fetch stalls.=0AEvaluations show that ELF can improve the performance by 4.1%=0Aand achieve total improvement of 11.9% when used with other=0Atechniques over commonly-used greedy-then-oldest scheduling.
SUMMARY:ELF: Maximizing Memory-Level Parallelism for GPUs with Coordinated Warp and Fetch Scheduling
PRIORITY:3
END:VEVENT
END:VCALENDAR