Cluster of workstations and PCs make an attractive platform for
parallel computing. Cluster computers can be formed out of existing
desktop workstations, which makes it easy for groups to embark on
parallel computing projects without a capital outlay; Also, even for
dedicated usage, clusters make economical parallel computers, because
of the use of inexpensive commodity hardware.
We have targeted this architecture since the early days of Charm++
development. Our first versions of Charm (then called "Chare Kernel",
around 1990) ran on networks of SUN workstations.
Our cluster computing research includes:
Portable implementation of parallel programming paradigms on
clusters (Converse, Charm++, AMPI, and Java).
Adapting to extraneous loads : In a desktop environment, if one
of the workstations being used by a parallel program starts running a
desktop job, the performance of the entire parallel application
suffers. Our research has led to a system which can detect this
situation and readjust the workload accordingly.
Handling machines with different speeds : This is done by
using initial speed measurements, and measurement-based load
balancing.
In addition to the cluster of SUN workstations in the CS department,
and in our lab, we have installed and used:
-
The "Cool: cluster: Intel gave us 64-processor cluster (16
quad xeons). This has been broken into to cluster. The one named
"Cool" is our workhorse cluster. The other cluster is being used for
Operations Research by Prof. U. S. Palekar.
-
The "Thrift: cluster: A cluster of 8 Linux PCs, put together
for under $4,500 in 1998. (This is now being dismantled).
-
In the Theoretical Biophysics collaboration, we had three
clusters of 16 HP machines, expanded with another 4 4-processor HP
K-boxes, in the Theoretical Biophysics collaboration, since 1993. This
was one of the early use of clusters for real applications.
-
Three cluster of Linux machines, in the Theoretical Biophysics
Collaboration (Beckman Institute). It consists of 32 processors (16
dual processor boxes) connected by 100 MB ethernet!
- In addition, we have access to several other clusters, including
the "Turing" cluster housed in the Computer Science department (with
about 400 processors). This is part of Computational Science and
Engineering effort, of which our group is a participant.
|
- 99-02
Robert K. Brunner and Laxmikant V. Kale, Adapting to Load on Workstation Clusters, The Seventh Symposium on the Frontiers of Massively Parallel Computation
- 99-01
Laxmikant Kale, Robert Brunner, James Phillips and Krishnan Varadarajan, Application Performance of a Linux Cluster using Converse, 3rd Workshop on Runtime Systems for Parallel Programming
|
