Grids over the enterprise WAN

Grid computing is touted by many as the killer application for wide area networks. The whole idea of grid computing -- that of reusing spare processing power on one computer for a job running on another -- is predicated on the idea that the computers are connected over a network.

This network can be a local area network or a WAN, but in practice is more likely to be the latter, as grid computing is right now largely the preserve of universities, research establishments and large corporations in the pharmaceutical, financial and oil industries. At its most basic, grid computing -- commonly referred to as the next-generation Internet -- can run over the flimsiest of networks. SETI@Home and other distributed computing applications, which are akin to a poor man's grid, run quite happily over the public Internet, with nodes (that is, PCs or servers that take on a share of the processing load) on everything from corporate LANs to dial-up connections.

But for serious grid computing the message is emerging that a robust and fast network is essential. The Globus Alliance -- a team of academic, government and commercial researchers that created an open-source Globus Toolkit of interfaces and source code -- cites the problems of making something work fast and consistently over a network that may be neither of those things.

Some organisations are taking a good look at their networks now, in anticipation of the demands that grid computing may place on them in future. This does not mean every company out there should be doing so, but some universities in particular are already upgrading their networks to 10Gbit backbones just to support grid computing -- and finding that in a buyer's market they are able to get good deals on the upgrades.

Grid evangelist Ian Foster, who is head of the distributed systems lab at Argonne National Laboratory in Illinois, agrees that grids are currently more academic curiosity than commercial venture: "Grid computing today is where the Internet was in 1991... today's mini-grids will grow into a huge global grid." However, the commercial applications of grid computing are beginning to catch on. IBM, HP and Sun are all pushing their grid initiatives as they feel the adverse affects of diminishing budgets for new server and network infrastructures in enterprises. The Globus project now has version 3 of its grid toolkit -- an open-source implementation of the Open Grid Service Architecture (OGSA) that companies can use to create grid applications. Even Oracle is getting in on the act with its 10g database and 10g application server. But while Larry Ellison will laud the potential of grid computing at OracleWorld, in private Oracle executives admit they have few corporate customers on the horizon.

Pharmaceutical giant GlaxoSmithKline is one company that is using grid computing to achieve jobs that would previously have taken months. "There are real examples out there," says Una du Noyer, head of infrastructure and security at Cap Gemini Ernst & Young, who is working with companies to implement grid-computing solutions. But, she admits, most applications are still confined to life sciences and financial applications. "It's finding uses in anything that you would traditionally use supercomputers for," she says.

In life sciences, grid computing is used for cancer research and protein sequencing and folding (the Folding@home project is one example of a distributed computing application for this). In financial services, applications include derivatives analysis, statistical analysis and portfolio risk analysis, where the statistical margin of error can be significantly reduced.

There are other applications: in the field of energy, grid computing can be used for seismic analysis and reservoir analysis; in manufacturing it can find a home in mechanical design and process simulation; and in the entertainment field there is digital rendering.

But all of this takes the right infrastructure. The computing element is easy: the whole point of grid computing is to use fallow computing power on underused computers in -- and sometimes even outside -- an organisation. It is the network part that needs careful consideration.

Cardiff University is nearing the end of a major overhaul of its already substantial network, prompted by several factors, one of which was its pioneering work on grid computing. Network team leader Tom Wearsma is nearing the end of a major project to upgrade the university's substantial WAN that will span two campuses (Cardiff University will merge with the University of Wales College of Medicine in August 2004), and has 25,000 outlets.

There are other applications: in the field of energy, grid computing can be used for seismic analysis and reservoir analysis; in manufacturing it can find a home in mechanical design and process simulation; and in the entertainment field there is digital rendering.

But all of this takes the right infrastructure. The computing element is easy: the whole point of grid computing is to use fallow computing power on underused computers in -- and sometimes even outside -- an organisation. It is the network part that needs careful consideration.

Cardiff University is nearing the end of a major overhaul of its already substantial network, prompted by several factors, one of which was its pioneering work on grid computing. Network team leader Tom Wearsma is nearing the end of a major project to upgrade the university's substantial WAN that now spans two campuses (Cardiff University recently merged with the University of Wales College of Medicine), and has 25,000 outlets.

"We could not do grid computing on the old network," says Wearsma. "It just was not up to the job." There were other reasons for upgrading the network to a 10Gb infrastructure supplied by Foundry Networks, says Wearsma, but grid computing was a key issue. "The whole point of grid computing is to make the most of the resources we have and to create a virtual supercomputer across our networks." Some 10,000 computers distributed over the university's WAN go quiet every night at 7.00pm. "Grid computing is all about how you use that infrastructure," says Wearsma.

To build the infrastructure, the university obtained a Science and Research Infrastructure grant to find a solution that would see it through at least the next five years. This meant a multi-gigabit backbone -- driven, says Wearsma, by grid computing.

Network integrator Pervasive Networks won the contract with its proposal to install a 25,000-outlet, 10Gb backbone based on Foundry Networks equipment that the university will in future be able to upgrade to 40Gb bandwidth using Foundry's MG8 technology.

"All manufacturers will have new hardware in the next five years so we were keen to see roadmaps," says Wearsma. "Foundry were willing to show us their roadmaps." For a multi-gigabit backbone, Wearsma was expecting to be given trunk gigabit -- essentially a number of 1Gbps backbones bundled together -- but, he says, Foundry and Pervasive came along with 10Gbps from day one. There are some 100 switches on the network, with 10/100Mbps Ethernet to the desktop and the potential for Gigabit to the desktop -- or to the 300 servers connected to the network. Between the two campuses lies 2km of dark fibre.

Wearsma reckons he got a good deal on the upgrade. "We happened to be in a buyer's market so there was some very aggressive bidding. Only academic institutions like ourselves have been on the market -- big corporations have been holding back a bit." Wearsma got the network for a cool £3m, and thinks it was worth it.

"We have a number of parallel processing machines, and a large 200-node Linux-based cluster in earth sciences," he says. "Processing jobs from this are then offloaded at Gigabit speeds on to the WAN -- this is a great example of grid computing. One development project in here is an access control program that goes round looking for spare computing cycles and can distribute work loads. Earth sciences has vast amounts of data that needs to be mined, and this all makes it far easier."