High Performance Computing and Parallel Systems Research

Data Mining is the process of extracting useful and previously unknown information out of large complex data collections.

Large amounts of data are collected routinely in business, government departments and research organisations. They are typically stored in large data warehouses or databases. For data mining tasks suitable data has to be extracted, cleaned and integrated with other sources. Further data analysis is required to find accurate, useful and understandable information.

Parallel and high-performance computing techniques are essential in data mining in order to be able to handle and process these large data collections within reasonable time. The ANU Data Mining Group is working on various aspects of high-performance data mining:

• Development of open source software for record linkage
  Data integration (linking records from one or more data sets that represent the same entity) is a crucial first pre-processing step in many data mining project, as well as in biomedical, statistical and business oriented data analysis applications. The aim of this project is to develop free, open source software for biomedical record linkage that allows the linkage of larger data sets by using techniques from parallel and high-performance computing with increased linkage quality by applying machine learning and data mining algorithms.

• Development of improved algorithms for predictive modelling that are scalable both with the size and complexity of real world data collections, as well as the parallel platform they running on (number of processors).

For more information please visit the ANU Data Mining Group web site.

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, and was designed over 1999-2000. Its entry won the prestigious Gordon Bell Prize for Price/Performance at Supercomputing 2000. Current work is based on the following themes:

• Optimizing Communication Patterns (with Wi Bing Tan)
  This involves both user-level and kernel level enhancements to communication patterns on a Beowulf cluster using commodity fast ethernet switches.
• Scheduling Strategies for Cluster Computers (with John Uhlmann)
  Gang scheduling is co-ordinated scheduling of parallel tasks, under the SCore cluster middleware facility. However, implicit co-ordinated and even local scheduling policies have the advantage of being able to overall inter-job computation and communication, and are of increasing interest.
• A Cost-Effective Node Architecture for Cluster Computers (with Bob Edwards, Tom Rowlands and John Uhlmann)
  This originally involved developing a `supernode' of 4 dual PIIIs using IDE buses. The new design involves driver development and porting MPI to PIII nodes connected by the Dolphin interconnect.

Sample Publications

• 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.
• Wi Bing Tan and Peter Strazdins. The Analysis and Optimization of Collective Communications on a Beowulf Cluster, In Proceedings of ICPADS'02: 2002 International Conference on Parallel and Distributed Systems, 8 pages, IEEE Press, Taipei, Dec. 2002
• Peter Strazdins and John Uhlmann. Local Scheduling out-performs Gang Scheduling on a Beowulf Cluster 11/03.

Collaborators

• Dr Bing Bing Zhou
  School of Computing and Mathematics, Deakin University

Related Links

• The Beowulf Project

Our interest is in the efficient exploitation of parallel computers for computational science type applications.

• OpenMP for Non-Uniform Memory Access (NUMA) Architectures
  OpenMP comprises a series of compiler directives and library calls for programming shared memory parallel computers. It is based on the assumption of uniform memory access. In reality, however, almost all large shared memory parallel architectures have a memory hierarchy with different access times and bandwidth associated with different memory addresses. Currently there is no consensus as to whether the OpenMP programming model should be extended with data placement directives to account for NUMA architectures or whether these issues are best handled through the O/S via sensible page placement and migration algorithms. Our research seeks to address this question. To date our work has focused on the SGI Origin 3000 system exploring the benefits of data placement for the iterative solution of the Laplace equation. Future work will target computational chemistry algorithms as part of the CC-NUMA project.
• Software Distributed Shared Memory (DSM)
  Although basic message passing still provides the most portable means of programming distributed memory parallel computers, there are a number of applications for which use of this paradigm is either unnatural or simply too complicated. For these applications software DSM models, that provide each process with one-sided access to a region of memory that may be physically distributed over many CPUs is particularly attractive. Currently our research is focused on developing a restricted version of the Adsmith object based DSM system that uses MPI-1 rather than PVM.

The aim of this project is to produce a complete machine simulator for an UltraSPARC SMP system. The project's main web page is at http://ccnuma.anu.edu.au/sulima/.

This project began as part of the ANU-Fujitsu CAP Program, Phase III, in 2000. It will continue from 2003, as part of the CC-NUMA Project, a collaborative project with Sun Microsystems and Gaussian Inc.

Team Members

The current Sparc-Sulima team members are Bill Clarke, Andrew over and Peter Strazdins. The team collectively can be contacted at sim@cap (.anu.edu.au).

Sample Publications

• Bill Clark, Adam Czezowski and Peter Strazdins, Implementation Aspects of a Sparc V9 Complete Machine Simulator , (11/01).

Collaborators

• Professor Gernot Heiser,
  School of Computer Science and Engineering, The University of New South Wales
  

Related Links

• The Sulima ISA Simulator

The objective of this project is broadly to investigate parallel linear algebra algorithms chiefly for dense matrices and also provide usable, high performance dense linear algebra codes. This project began as part of the ANU-Fujitsu CAP Program, in 1990.

Past and current work is based on the following themes:

• Distributed BLAS and LAPACK libraries
• BLAS libraries on SPARC processors: this work was incorporated in the UltrSPARC kernels for the Developer's Release of the ATLAS project (with Viet Nguyen)
• dense parallel algorithms: Symmetric Indefinite Solvers (codes incorporated in Data Mining and Electromagentic Compatibility Analysis applications), Parallel Matrix Factorization, Out-of-core algorithms.

Sample Publications

See publications of Peter Strazdins.

Collaborators

• Professor Jack Dongarra
  Innovative Computing Laboratory, University of Tennessee Knoxville
• Dr John Lewis, Boeing Corporation