Cluster Systems Research

This is an Associated Project of the ANU-Fujitsu CAP Program, Phase III.

This Project is based on the ANU Beowulf cluster parallel computer, known as the Bunyip. This contemporary cluster has 96 dual Pentium III nodes connected using a Fast Ethernet switch. Its entry won the prestigious Gordon Bell Prize for Price/Performance at Supercomputing 2000.

Objectives

This Project aims to develop and evaluate system-level software and hardware for the Bunyip Beowulf cluster, and its possible successors.

Sub-Projects

  1. Gang Scheduling
    Co-ordinated scheduling of parallel tasks (gang scheduling) is essential in order to use the resources of a parallel computer efficiently. The implementation of gang scheduling within an operating system kernel is the most powerful approach, as the kernel has the greatest potential of knowledge and control of the system's resources.

    This project will investigate, analyse and improve practical implementations of gang scheduling within an operating system such as Linux, and within middleware cluster facilities, such as SCore. A comparison of the two approaches, kernel-based and middleware-based, is also an interesting question.

  2. Optimizing Communication Performance
    This project seeks to improved the performance of communication primitives and patterns on cluster computers using commodity fast ethernet switches, such as the Bunyip. This may be both at the MPI (applications) level and and the kernel level. Communication-intensive applications such as dense linear system solution can benefit from these improvements.

    Approaches include evaluating algorithms fir standard communication patterns such as `all-gather', the use of Kernel MPI for faster basic messaging performance, and utilizing the IP layer multicasting capabilities.

  3. Developing a Cost-Effective Node for Bunyip2
    This project seeks to develop a highly cost-effective `supernode' for a cluster computer, based on dual Pentium-based processor systems. Communication can occur within the supernode using IDE buses. The use of supernodes reduces the size of the main cluster switching network, thus enabling it to use a faster, more expensive technology (eg. Myrinet) for the same overall price.

    This will involve implementing a PCI-based card to drive the target end of the IDE bus, building the supernode prototype, and evaluating and optimizing its communication performance.

Researchers

These include Bob Edwards (Technical Support Group, DCS ANU - Bunyip Design), Wi Bing Tan (Honours Student - Communication Patterns), Peter Strazdins, and John Uhlmann (Honours Student - Gang Scheduling).

For enquiries on this Project, Peter is probably the best first point of contact.

Publications

1
D. Aberdeen, J. Baxter and R. Edwards, A 98c/MFLOP Ultra-Large Scale Neural Network Training on a PIII Cluster, Proceedings of Supercomputing 2000, November, 2000.
2
P.E. Strazdins, High Performance Dense Linear Systems Solution on a Beowulf Cluster.
In Proceedings of the 5th International Conference and Exhibition on High-Performance Computing in the Asia-Pacific Region, 15 pages, Gold Coast, September 2001.
See also the relevant sections (1, 2) in the CAP Program Report for 2001.

Related Links

Peter Strazdins
2001-11-20