DHPC Technical Report DHPC-064

DHPC Technical Report DHPC-064

Remote Application Scheduling on Metacomputing Systems

H.A.James and K.A.Hawick

Archived: 15 February 1999

Published in Proc. of High Performance Distributed Computing (HPDC'99), Redondo Beach, California, August 1999

Abstract

Efficient and robust metacomputing requires the decomposition of complex jobs into tasks that must be scheduled, or placed onto processors within the distributed environment. In addition to the placement of tasks, there is need to manage data exchange and control of a job's computational flow. Depending on the level of knowledge that the node which creates the execution schedule has about the global system state, there are various ways of creating a schedule and implementing it efficiently. Most computations may be structured as a process network, where data is either pushed from the creating (or storing) node to the node where it will be used, or at the instigation of a node, retrieved from where it was created. The DISCWorld metacomputing infrastructure employs the concept of a rich data pointer, the DRAM, which can point to data and services, and can be traded between clients and servers, and servers. We present an extension of the DRAM to represent and describe data that has not yet been created, the DRAMF. We show how the use of the DRAMF facilitates efficient scheduling in the DISCWorld, and how computations are able to be optimized. We also present a simple recursive algorithm for determining the optimal placement of a job's components in the presence of partial system state information by only considering a selected subset of all available processing nodes. The model that we describe is not restricted to our metacomputing framework, but is applicable to any system that supports a global data naming scheme.

PDF version

PostScript version (gzip compressed)

PDF version of abstract published in HPDC 99

PostScript version of abstract published in HPDC 99 (gzip compressed)

webmaster@dhpc.adelaide.edu.au