Memory Performance of KDD Applications

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

Objectives

The objective of this project is to analyse the cache, memory and I/O behaviour of Knowledge Discovery in Databases (KDD) applications, to detect system architecture or operating system related bottlenecks for such applications, and to come up with ideas and techniques of how modern shared memory multiprocessors and operating systems can be modified so that applications with irregular memory access become more efficient.

Approach

Our approach to analyse, using performance analysis tools on UltraSPARC SMPs and UltraSPARC SMP simulators, of parallel KDD applications including the decision tree induction (C4.5), Predictive Modelling, Probabilistic Record Linkage, and SQL-aware Data Mining.

Currently, we are using hardware performance counters to collect and analyse various hardware events like cache misses, TLB misses, MIPS, MFLOPS, CPI, etc.

The Research Team and Contact

The current people working on this project are Adam Czezowski and Peter Christen. Peter is probably the most suitable person for first contact on the Project.

Publications and Presentations

How Fast is -fast? Performance Analysis of KDD Applications using Hardware Performance Counters on UltraSPARC-III
Peter Christen and Adam Czezowski
Proceedings of the Australasian Data Mining Workshop, Canberra, December 2002.
Paper (ps.gz, 82 KB) Paper (pdf, 174 KB) Slides 4up (ps.gz, 335 KB) Slides 4up (pdf, 745k KB)
(also available as Technical Report TR-CS-02-03, ANU Department of Computer Science, September 2002)
Performance Analysis of KDD Applications using Hardware Event Counters
Peter Christen and Adam Czezowski. ANU Computer Science Technical Report TR-CS-02-01, February 2002.
Available online from http://cs.anu.edu.au/techreports/2002/TR-CS-02-01.html.
Performance Analysis of KDD Applications using Hardware Event Counters
Peter Christen and Adam Czezowski. Presentation at the CAP Workshop 2002, Fujitsu, Kawasaki, Japan, 6 February 2002.
Slides (pdf.gz, 66 KB) Slides (ps.gz, 58 KB)

Related Publications

1: Peter Christen, Markus Hegland, Ole M. Nielsen, Stephen Roberts, Peter E. Strazdins and Irfan Altas, Scalable Parallel Algorithms for Surface Fitting and Data Mining. Elsevier Journal of Parallel Computing, 27(2001), September 2001.
2: Peter Christen, Ole M. Nielsen, Markus Hegland and Peter E. Strazdins, Parallel Data Mining on a Beowulf Cluster, Proceedings of the HPC-Asia 2001 Conference, Gold Coast, September 2001.
3: Ole M. Nielsen, Peter Christen, Markus Hegland, Tatiana Semenova and Timothy Hancock, A Toolbox Approach to Flexible and Efficient Data Mining, Proceedings of the PAKDD-2001 Conference, Hong Kong, April 2001, Springer Lecture Notes in Computer Science, Artificial Intelligence series, LNAI-2035.

See also the relevant chapters in the CAP Program Report for 2001