BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20151117T220000Z
DTEND:20151117T223000Z
LOCATION:18CD
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: We present a new approach for multithreaded communication=0Aand asynchronous progress in MPI applications, wherein we offload=0Acommunication processing to a dedicated thread. The central=0Apremise is that given the rapidly increasing core counts on modern=0Asystems, the improvements in MPI performance arising from=0Adedicating a thread to drive communication outweigh the small=0Aloss of resources for application computation, particularly when=0Aoverlap of communication and computation can be exploited. Our=0Aapproach allows application threads to make MPI calls concurrently,=0Aenqueuing these as communication tasks to be processed=0Aby a dedicated communication thread. This not only guarantees=0Aprogress for such communication operations, but also reduces load=0Aimbalance. Our implementation=0Aadditionally significantly reduces the overhead of mutual=0Aexclusion seen in existing implementations for applications using=0AMPI THREAD MULTIPLE. Our technique requires no modification=0Ato the application, and we demonstrate significant performance =0Aimprovement (up to 2X) for QCD, FFT and deep learning CNN.
SUMMARY:Improving Concurrency and Asynchrony in Multithreaded MPI Applications using Software Offloading
PRIORITY:3
END:VEVENT
END:VCALENDAR