BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20121114T204500Z
DTEND:20121114T210000Z
LOCATION:155-F
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: The design and development of current generation supercomputing=0Asystems is fueled by the increasing use of multi-core=0Aprocessors, accelerators and high-speed interconnects,=0AHowever, scientific applications are unable to fully=0Aharness the computing power offered by current generation=0Asystems. Two of the most significant challenges are communication/synchronization latency=0Aand power consumption. Emerging parallel programming=0Amodels offer asynchronous communication primitives that=0Acan, in theory, allow applications to achieve latency hiding. =0AHowever, delivering near perfect communication/computation overlap=0Ais non-trivial. Modern hardware components are designed=0Ato aggressively conserve power during periods of inactivity. =0AHowever, supercomputing systems are rarely idle and software libraries=0Aneed to be designed in a power-aware manner. In our work, we address both of critical problems. We propose hardware-based non-blocking MPI collective operations and re-design parallel applications to achieve better performance through latency hiding. We also propose power-aware MPI collective operations to deliver fine-grained power-savings to applications with minimal performance overheads.
SUMMARY:High Performance Non-Blocking and Power-Aware Collective Communication for Next Generation InfiniBand Clusters
PRIORITY:3
END:VEVENT
END:VCALENDAR