Linux cluster for performance computing

A wave of linux clustered machines has been anticipated since 1994, when the Goddard Space Flight Center produced the first Beowulf cluster. But only recently have the tools become available for the Linux cluster to become a more manageable engine on the Internet.

Clusters are multiple computers connected by a high-speed, local area network. If one computer or node of a cluster fails, other nodes of the cluster will pick up its work. The technology is seen as an inexpensive way to ensure 24 by 7 operations, although cluster software is required and one unit of the node must act as referee.

Today, Linux clusters are functioning as high-availability Web servers, domain name servers and mail servers running standard, nonparallel applications. One of the largest Linux cluster sites is the 6,000-node engine behind the Google search site, says Richard Schuh, chief operating officer at Linux NetworX, a Sandy, Utah, Linux cluster start-up.

In addition, Linux clusters functioning as scientific problem-solvers have grown far beyond the original 16-processor Beowulf scale to become a low-cost alternative to expensive parallel processors. Linux clusters are tackling large problems in unraveling the human genome, meteorology and pharmacology.

One cluster, Computational Plant, or CPlant, at Sandia National Laboratories, is a 580-processor cluster that ranks 62nd on the Top 500 Supercomputing Sites list presented by the University of Mannheim and the University of Tennessee. And Houston energy giant Conoco has built a Linux cluster to analyze seismic data from oil and gas exploration.

"We believe in the superior price/performance of the Linux cluster," Schuh says, "but they won't get adopted into the mainstream unless there are tools to administer them like single systems."

Linux NetworX and other suppliers - such as Rainfinity, Red Hat, Silicon Graphics Inc., SteelEye Technology, TurboLinux and VA Linux Systems - have brought clustering management software to market over the past few months to make Linux clusters more manageable.

API Networks, a strategic business unit of Alpha Processor Inc., packages and sells skeleton Linux servers running Compaq Computer's high-speed Alpha processors. Since Linux was ported to the Alpha chip second, after it was first ported to Intel's central processing units (CPUs), Alpha has proven a frequent base for high-performance Linux clusters.

Since Beowulf, "a body of expertise has been built up in Linux clusters," says Tom Morris, senior manager of product strategies at API. His firm is shipping 833-megahertz Alpha CPUs in its just-released rack-mount CS20 servers, in a design "optimized for cluster implementations," he says.

SGI recently unveiled the Advanced Cluster Environment for SGI 1000 Series Linux rack-mount servers. All the units of an ACE cluster can be viewed through one management console. Any node of the cluster can be restarted remotely from the console, and each node's performance can be presented visually on the screen.

Amy Chung, director of SGI's software product marketing, says SGI demonstrated a 128-node Linux cluster with an SGI Origin 3000 server front end at LinuxWorld in August in San Jose. The Origin server can handle a much larger input/output datastream when running Irix, SGI's Unix, than when running Linux, but the Linux cluster provided plentiful, cheap processing power. "We took the best of each - mature Unix and Linux - and put them together into one system," Chung says.

Linux distributor TurboLinux issued LinuxCluster Server 6 last month, an update to the cluster management software first issued in November 1999. It provides load balancing across nodes and can pass off HyperText Transfer Protocol traffic if one node fails. The software is behind the Birkenstock footwear online store, which formerly ran on a single 350-MHz Pentium server. It now runs on five machines - two cluster managers, two Web servers and a database server - according to Howard Lee, marketing manager at Birkenstock.

In June, Red Hat brought out the High Availability Server, which provides load balancing and some failover protection for a two-node cluster.

But for transaction processing failover, where a server fails while retrieving or entering data in a database, third-party failover software is needed to ensure the transaction is not lost. SteelEye's LifeKeeper for Linux 3.0 is available at $2,995, Linux NetworX's Schuh says. His firm has a partnership with SteelEye.

In addition, Rainfinity provides its Rainfront and Rainwall load balancing and cluster management software for firewalls, Web servers and other servers in the Internet infrastructure. VA Linux offers VACM 2.0 for managing clusters.

Until now, managing the cluster has been a matter of managing all the individual nodes. Now, with TurboLinux's Linux Cluster Server 6, "you can take out one node and put in another on the fly," while the rest of the cluster continues running, says Lonn Johnson, vice president at TurboLinux. TurboLinux and other suppliers aim to make the cluster run as if it were a single machine, he adds.