#Number
TR-PDS-1997-009

#Title
The Challenges of Low-Overhead Message-Passing Communication
Using Commodity Superscalar Processors

#Author
Brian Grayson
Craig Chase

#Abstract
Commodity microprocessors are generally used as the compute
base for high performance parallel machines, including networks
of workstations (NOWs) and parallel computers.  In spite of
recent advances, communication between processors remains
relatively expensive.  We present a message passing model and
network interface design that reduces communication overhead
nearly to the absolute limit.  Unlike previous work, we assume
a highly-parallel superscalar processing node with a
weakly-consistent memory system.  We show how traditional
techniques (e.g. polling for message arrival) are poorly suited
for superscalar processors, and introduce new synchronization
methods that allow computation to proceed in a purely dataflow
fashion, without modifications to the processor architecture. 
Messages are sent as soon as their data has been written,
computation commences as soon as incoming data arrives, and
local computation can be arbitrarily intermixed with
communication subject only to the true data dependences that
exist in the program.  

	
#Bib
@TechReport{,
  author = 	 "Brian Grayson and Craig Chase",
  title = 	 "The Challenges of Low-Overhead Message-Passing Communication
                  Using Commodity Superscalar Processors",
  institution =  " Parallel and Distributed Systems Laboratory, ECE
                   Dept. University of Texas at Austin",
  month =	 "July",
  note =	 1997,
  note =	 "available via ftp or WWW at maple.ece.utexas.edu
		  as technical report TR-PDS-1997-009"
}