ASPhALT: an Automatic System for Parallel AppLication Transformation

Colloq: Speaker: 
Anthony Danalis
Colloq: Speaker Institution: 
University of Delaware
Colloq: Date and Time: 
Tue, 2008-07-08 13:00
Colloq: Location: 
5700, D307
Colloq: Host: 
Philip Roth
Colloq: Host Email:
Colloq: Abstract: 
In order to provide ever increasing processing power, modern clusters have grown to machines with several thousand processing units. As a result, communication between large numbers of processing units can induce performance overheads that need to be addressed for a parallel application to scale well and harness the full potential of a modern cluster.<br><br>In this talk I will describe ASPhALT, an Automatic System for Parallel AppLication Transformation. ASPhALT aims to reduce the communication cost that MPI applications experience on computing clusters by increasing the communication computation overlap achieved by the applications. Doing so requires vertically integrating knowledge and optimizations at the system, library and compiler layers. In my talk I will discuss the different components of the system and demonstrate how an MPI aware compiler optimization phase combined with a specialized MPI-companion communication library and knowledge of the cluster environment can increase communication computation overlap and achieve better performance for parallel MPI applications.
Colloq: Speaker Bio: