Project Title:
Piranha Parallelism: Distributed Self-Management of Computing Resources in a Network
06.01-7442
912118
Piranha Parallelism: Distributed Self-Management of Computing Resources in a Network
Environment
Scientific Computing Associates, Inc.
One Century Tower, 265 Church St
New Haven
CT
06510-7010
Robert D.
Bjornson
203-777-7442
ARC
NAS2-13556
100
06.01-7442
912118
Abstract:
Piranha Parallelism: Distributed Self-Management of Computing Resources in a Network
Environment
At the present time, typical large-scale computations may comprise at least three
major steps: input preparation, possibly on modest, text-oriented workstations; numerical
calculation on a large vector or parallel supercomputer; and output examination,
most likely on high-performance graphics workstations. Unfortunately, this entails
very unbalanced utilization of computer hardware. All the workstations are largely
unused (at least in terms of compute-power), and the parallel supercomputer can be
so oversubscribed that users have long waits for available time slots. The goal of
this project is to alleviate this situation. It will exploit networks of workstations
to provide additional supercomputer power through the development of software to
dynamically allocate unused workstation cpu cycles to fill the computational demands
of large jobs. This will be achieved in the framework of a simple and powerful scheduling
model called piranha parallelism. Users will generate parallel tasks following a
fixed format, and the tasks will be released into a network-wide task pool. These
tasks are guaranteed to be attacked by as many computational piranhas (aka workstations)
as have available idle cycles. Phase I will focus on the design and implementation
of a prototype piranha system.
Commercial and government users will see greatly enhanced output on important applications
and will incur far lower hardware acquisition costs, since it will make more effective
use of in-place computers in situations where local area networks of workstations
are rapidly becoming the standard computing environment.
parallel programming, network operating systems, Linda, distributed computing