Less than two months after the University of Queensland announced plans to build a 128-node Linux supercomputer, the Australian Centre for Advanced Computing and Communications (ac3) will eclipse that by installing a 147-node cluster running at some 1.5 teraflops in its Sydney data centre.
The new ac3 system, costing about A$750,000 (US$486,000), will use dual 3.06GHz Xeon processors for each node. The supercomputer will also have 300G bytes of memory, 1.2T bytes of storage, and run Red Hat Inc. Linux with a 2.4 series kernel.
The Australian Research Council (ARC) and a consortium of NSW universities including the UTS, Sydney University, and UNSW jointly fund ac3. Ross McPhedran from the Sydney University Physics department is one of the principal instigators behind the project.
“The first boxes will go in early this month for evaluation and we will then have an official launch towards the end of August,” Professor McPhedran said. “The cluster is the largest of its type in Australia and be a very powerful tool for academic research. Applications such as designing harder wearing materials – including better disk brakes for cars – and better drugs, will be among its uses.”
McPhedran said the overwhelming direction in supercomputing is clustered computing.
“Putting clusters together is much easier now than the approach of 15 years ago,” he said. “Beowulf clusters today are a cost-effective sweet spot for academic institutions.”
The system, to be supplied by Dell, is being funded by an initial A$500,000 grant from ARC with further funding to be supplied by the university consortium.
“A tender was put out for the system and while we were very happy with all the tenders, Dell offered the best price performance ratio,” he said. “The cluster will consist of blade servers with a fibre optic back plane switch. We are yet to do teraflop science but hope to achieve it with this system.”
Ac3 CEO Phil McCrea said although clusters are cheap, not all applications easily lend themselves to parallel processing.
“We have done financial modelling work on clusters for certain applications and have found that adding extra hardware actually slows down processing,” McCrea said. “This is due to the inter-processor communication between nodes becoming saturated. For example, one application showed an increase in processing speed up until 20 nodes, but at 21 nodes it tapered off.”
The new Linux cluster will join ac3’s impressive family of high-performance computing hardware including an NEC SX-5 vector computer and a 64-processor SGI Origin 2400.
Ac3 recently decommissioned a 68-processor IBM RS/6000 cluster, which McCrea described as a “normal commercial decision” for the leased machine.
“Clustered machines add another degree of difficulty in that MPI (message passing interface) code is needed which is the main disadvantage of clusters compared with shared memory architectures.” “We have a variety of systems at ac3 because of the range of specialist applications we provide,” McCrea said.