Clusters Make a Stand

More and more companies are turning to high-power grids and clusters of servers to provide the horsepower needed in their data centers. But the flurry of new clustering announcements also raises the question about how these large projects will link this infrastructure together.

For example, take GlobeXplorer LLC., a specialist in providing satellite images and aerial photography, which has become the latest firm to build a major cluster of servers in its data center. The Walnut Creek, Calif.-based firm announced this week that it has implemented 40 PowerEdge 1750 servers from Dell Computer Corp. (Nasdaq: DELL) in an attempt to tame 400 terabytes of raw image data.

In total, the cluster offers a compute engine that can work at speeds of more than 448 gigaflops, which, in case you were wondering, stands for a million floating point operations per second.

Such an ambitious project raises some good quesions: How would the networking infrastructure handle the massive amounts of data being transferred?

GlobeXplorer is using Fast Ethernet and Gigabit Ethernet to link its cluster, which is performing well. Rob Shanks, the company's president and CEO says, We have very little downtime – we don’t have any bandwidth issues.”There's more to building a cluster than raw processing power. Bill St. Arnaud, a grid computing expert at Canadian internet development organization Canarie Inc., warns that data center managers should do some serious thinking about bandwidth.

He says, "The big challenge in these clusters is the bandwidth between the machines – the issue of using technologies such as InfiniBand or even 10-Gbit/s Ethernet needs to be carefully thought through."

"In getting to these gigaflop speeds – bandwidth is usually the biggest choke point," he adds.

Not surprisingly, the same applies to grid computing, which is when significant computing resources are applied to achieve a specific task or solve a problem. This, in particular, demands high-speed connectivity and ultra-low latency.

But at least the technologies are becoming available to help achieve this, according to a Light Reading report on grid networking that came out last year. The report said that the advent of high-performance Ethernet and the availability of line-rate 10-Gbit/s Ethernet are significant steps toward enhancing the network layer to support grid computing initiatives (see Grid Networking).With grid projects already under way at such firms as Charles Schwab & Co. Inc., more and more data center managers could be confronted with similar schemes. No more, it seems, are grids the sole preserve of government research agencies and academic institutions (see Blades for Buffalo ).

Linux could be a key enabler for companies. GlobeXplorer, for example, is using Red Hat Linux for its Dell deployment. There is more to this than the cost savings that are often attributed to open source software.

St. Arnaud says, "Almost all of the big clusters out there are Linux – the Linux kernel is easily modified, so if you need to move a bus, it's relatively easy to do."

"And you're not dependent on the vendors' choice of operating system," he adds.

— James Rogers, Site Editor, Next-gen Data Center Forum0