BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20141118T223000Z
DTEND:20141118T230000Z
LOCATION:393-94-95
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: This paper presents a new methodology for implementing SPMD distributed memory kernels.=0AIn this methodology, the programmer describes a generic implementation strategy for a particular class of kernels by writing a partial, generic implementation, and wrap it as a library with a concise and clean interface.=0AThe programmer then relies on our system to derive a concrete MPI implementation that follows the strategy and matches a given reference implementation.=0AThis methodology is made possible by a new set of language constructs that allow programmers to relate the behavior of a sequential reference implementation to a distributed implementation, and a new synthesis algorithm for distributed memory implementations.=0AWe demonstrate the methodology by implementing non-trivial kernels from NAS Parallel Benchmarks. Our approach can automatically infer challenging details and produce efficient implementations which performs within 95% of hand-written Fortran code, while promoting reusability and reducing programmer effort by leveraging generative programming and synthesis.
SUMMARY:MSL: A Synthesis Enabled Language for Distributed Implementations
PRIORITY:3
END:VEVENT
END:VCALENDAR