Speculative Parallelization – Eliminating the Overhead of Failure

Existing runtime parallelization techniques impose severe performance penalties when a speculative parallelization is attempted and fails. Some techniques require a sequential restart of the speculative execution while others only disregard the work after the first point of failure. This paper intro...

Full description

Saved in:
Bibliographic Details
Published inHigh Performance Computing and Communications pp. 460 - 471
Main Authors Luján, Mikel, Gustafson, Phyllis, Paleczny, Michael, Vick, Christopher A.
Format Book Chapter
LanguageEnglish
Published Berlin, Heidelberg Springer Berlin Heidelberg
SeriesLecture Notes in Computer Science
Subjects
Iteration Space
Loop Body
Memory Access
Memory Address
Speculative Execution
Online AccessGet full text

Cover

More Information
Summary:Existing runtime parallelization techniques impose severe performance penalties when a speculative parallelization is attempted and fails. Some techniques require a sequential restart of the speculative execution while others only disregard the work after the first point of failure. This paper introduces a new technique that reduces the performance overhead of failure to less than 1% on standard processors through a combination of hoisting the failure path and partitioning work to a Coinspector Thread.
ISBN:9783540754435
3540754431
ISSN:0302-9743
1611-3349
DOI:10.1007/978-3-540-75444-2_45