Intel offering more power to Melbourne researchers

For decades, Intel has struggled to build processors that could compete with specialised chips from supercomputing rivals. Today's chips, however, are cheaper, more scalable and fast enough that they're winning over life sciences researchers -- and redefining the rules of supercomputing.

Researchers at the Melbourne Advanced Research Computing Centre (MARCC) are waking to the analytical possibilities of a new class of high-performance supercomputer that was delivered to the centre's University of Melbourne data centre this month.

The new system is an IBM 1350 Linux cluster - a 100-processor supercomputer that will help researchers at the University of Melbourne, APAC (Australian Partnership for Advanced Computing), VPAC (Victorian Partnership for Advanced Computing), and other affiliated bodies churn their way through increasingly complex problems.

MARCC's new system quadruples the computing power available to the centre, which also runs a NEC SX4 supercomputer that was commissioned with much fanfare over three years ago but is now showing signs of age. Other HPC resources within MARCC include a cluster of HP AlphaServer systems and 16 standard Intel Pentium III-based servers; some existing systems will be decommissioned once the IBM 1350 is up to full speed.

The group's latest upgrade is significant for more than the fact of modernisation, however: bucking the conventional wisdom of high-performance computing (HPC) circles, MARCC has done away with the proprietary Unix servers that have long been a mainstay in computing-based research. Instead, the new system is based entirely on 2.4GHz Intel Pentium III Xeon processors, high-performance chips designed for intensive server-based applications.

The system's 100 Xeon processors are spread in pairs across 50 dual-processor computing nodes linked into a single computing cluster (two of these pairs manage the other 48, which are dedicated to number-crunching). This approach has kept the costs of the IBM system low -- around $400,000, compared with several million dollars for traditional proprietary or Unix-based supercomputers.

Dirk van der Knijff, University of Melbourne-based manager of the MARCC, says the cost benefits of an industry-standard platform make the move to commodity processors a no-brainer. "The fastest CPU that you can buy at the moment will do around 8 gigaflops [8 billion floating point operations per second]," says van der Knijff. "A 2.4GHz Xeon chip sustains well in excess of 1 gigaflops but costs one-hundredth of the price. That gives us around 10 times the computing power [of a conventional supercomputer] for the same price."

The possibilities that emerge from so much power are still becoming apparent.

It will certainly allow for added complexity in projects such as a computer model of the enteritic nervous system that's been established by one team of MARCC researchers; another project is crunching data on kidney cells, nephrons and other organ structures to build a working model that simulates the chemical structure of the kidney.

There will also be a significant focus on for the panoply of projects that are defining compute-based life sciences research: modelling the complex patterns of protein folding, analysing nuclear magnetic resonance scans, and so on.

"In the past, we haven't had the power to offer the biology people here," says van der Knijff. "There are a number of areas that the university is interested in, and we hope to be able to take on [other organisations] as customers as well."

The new system's potential will become even better realised as researchers build applications that take advantage of GrangeNet, an extremely high-speed network that is connecting Australian universities and research institutions. GrangeNet's ubiquity within research circles will allow MARCC to share the IBM system's processing power with other institutions. This would pave the way for practical telemedicine applications that combine remote connectivity with substantial computing power -- for example, providing real-time visualisation of surgical procedures for a specialist involved in remote surgery.

MARCC isn't the only Australian university making the Intel switch. In January, the Queensland Parallel Supercomputer Foundation (QPSF) -- which involves six Queensland universities -- announced it would take delivery of a pair of SGI Altix 3000 supercomputers, one with 64 CPUs and one with 16 CPUs. Both will run Intel's Itanium 2 processor, the company's next-generation 64-bit processor that provides increased data crunching capabilities compared with the 32-bit Pentium and Xeon families.

Bioinformatics, systems modelling, computational physics, computational chemistry and engineering are among the applications already tagged by the QPSF. The systems will also be linked to GrangeNet, allowing sharing of resources between Australian research institutions and -- via GrangeNet interconnections with similar research networks in the US, Canada and elsewhere -- around the world.

According to Intel, a total of 15 Asia-Pacific research institutions have now switched to Intel-based systems for compute-intensive research, a trend that is sending traditional high-end suppliers scrambling as it changes conventional notions of the supercomputer. Much of the platform's appeal lies in the flexibility provided by its clustered architecture: universities no longer need to buy one massive computer and the attendant complexity of the technology needed to synchronise dozens of processors.

Rather, Intel-based computing clusters aggregate the power of many lower-powered systems, distributing tasks between connected nodes. Hong Kong Baptist University, for example, is using Intel Xeon chips in a 64-node computing cluster that has been set to research on traditional Chinese medicine. The Indian Institute of Science has installed a 32-processor Itanium 2 system for use within its Supercomputing Education Research Centre, while China's Tsinghua University is using a 144-processor cluster for bioinformatics research.

Hong Kong University has built a cluster of 300 standard Intel Pentium 4-based desktop PCs that are used for life sciences projects involving molecular, dynamic N-body simulations. Singapore's MIT Alliance is using a system with 60 Itanium 2 processors, while 32-processor Xeon clusters are being used for life sciences-related work at three other Singapore research institutions.

Acceptance of systems running Intel processors represents a major coup for the chipmaker, which dominates the conventional computer market but has struggled in the past to produce chips that could compete with specialised HPC processors from niche makers such as NEC, SGI and Cray. Recognising that their proprietary days are numbered, even those companies have moved towards Intel-based solutions.

Three years ago, just three of the world's top 500 supercomputers (listed at www.top500.org) ran Intel processors; today, the number is 56. One of the main reasons for that, says MARCC's van der Knijff, is the fact that Intel-based clusters can be scaled to the nth degree: "Demand for compute power has been going up so everybody has to use large numbers of processors," he explains. "The subdivision mechanism you use to spread jobs across that many [Intel] processors is generally able to be extended."

Clustered computing is particularly appropriate for funding-hungry research organisations since they can extend modular clusters in stages each time additional grant monies roll in. Life sciences will lead the way in further increasing Intel's stature among supercomputing circles - and Corey Loehr, the company's recently appointed Australian business development manager for life sciences, is excited by the prospects of a market that IDC has valued at $US6 billion with annual growth of 24 per cent.

The local market will be particularly ripe for the benefits of Intel technology, which Loehr sees as a critical equaliser between big and small research organisations. "In Australia, a lot of bio-IT companies are not big," he explains. "The fact we can give them great performance at a realistic price is great, because they can get this power in their hands and compete on a research basis with some of the larger companies."