Honours project proposal:
Evaluation and Optimization of Communication Patterns on a Beowulf

co-supervisor: Bob Edwards,

This project has been undertaken by the Honours student Wi Bing Tan, 07/01

A cluster computer such as the ANU Beowulf is a parallel computer using a Commercial-off-the-Shelf switch-based network to communicate between the processors. Usually, the TCP/IP protocol, accessing the operating system is used to send these messages.

Compared with custom built networks of parallel computers such as the Fujitsu AP3000, these networks are relatively slow, often limiting application performance. On the other hand, they exhibit less message contention,which is a desirable property.

Thus for these networks, the evaluation and optimization of typical communications, including normal point-to-point messages, and widely used communication patterns such as broadcast, multicast and reductions, is an interesting and important issue.

This project can look at and examine one or both of the following aspects:

• Evaluation and optimizations of communication patterns within MPI. This would include performance comparisons of various algorithms (eg. binary-exchange methods vs tree-based broadcasts), on the cluster, as well as modifying algorithms to match the Beowulf's special topology.

• Optimizing the Beowulf TCP/IP and network drivers, in order to improve communication performance. It includes improving bandwidth by using all three of the Beowulf's network interface cards, short-cutting the TCP/IP protocol stack for traffic between cluster nodes,a nd optimizing broadcast and multicast operations.

  This requires some experience in writing device drivers for Linux,

References

• The cluster computer home page.
• The ANU Beowulf home page.
• for communication patterns, see Peter for one of his papers and other papers.