The servers run the hospital's mission-critical radiation therapy system used to treat cancer patients. When the power went out, one server -- in a room lacking emergency power -- immediately went down. But the application kept running with no loss of data or interruption as the other server kept humming along.
At another time, smoke from a fire in one of the server rooms brought a server down, even as the overall system continued running without failure. Overall, the radiation therapy system has maintained "better than 99.999 percent uptime," Michaelson said. "We're very happy with its performance."
Littleton, Mass.-based Marathon hopes the latest iteration of its nearly-crash-proof technology will reach an even wider audience. Yesterday, the company launched EverRun HA, which it claims can provide nearly 100 percent uptime for Windows-based clusters. EverRun HA follows the two other applications Marathon has released since its comeback from bankruptcy in January 2004: EverRun Splitsite, which helps servers -- even if they're many miles apart-- replicate data and back up operations, and EverRun FT.
Installed on an identical pair of Windows single- or dual-processor servers, EverRun FT essentially runs an application on both machines simultaneously while using a virtualized layer to create the appearance of a single server, according to Gary Phillips, CEO of Marathon. That eases deployment of applications that are not specially written to divvy up workloads among multiple processors or machines and guarantees virtually no downtime.
EverRun HA is more flexible. Users can deploy an application to run over small groups of Windows servers, including racks of blade servers -- even those with multiple processors. The tradeoff is that EverRun HA guarantees only the industry standard of 99.999 percent uptime, or high availability, instead of true fault-tolerance.
"EverRun HA nicely leverages their FT product to give a lower price and more flexibility to customers who are willing to trade-off ultimate availability for a lower price and greater configuration flexibility," said Gordon Haff, an analyst with Nashua, N.H.-based Illuminata Inc. "Marathon seems to have finally found an overall strategy, product set and go-to-market model that is working for them."
Marathon was founded in 1993 by former Digital Equipment Corp. workers. Companies have traditionally looked to mainframes or Unix systems when they demand fault-tolerant or highly-available systems. But that option is not always possible -- especially when the software only runs on Windows, as is the case for Princess Margaret Hospital's radiation therapy management system. So Marathon sought to bring fault-tolerance to Windows via customized Windows servers that competed with clustering software from EMC Corp. subsidiary Legato Systems Inc., Symantec's Veritas unit or even Microsoft Corp., with its Windows Server Cluster software.
While nearby rival Stratus Technologies Inc. continue to sell hardware, Marathon became a software-only vendor after a tough post-9/11 period that led to bankruptcy, a management shakeup and a relaunch in January 2004.
While few current Marathon users had much experience with EverRun HA, they praised other existing Marathon software.
Among them is Mike Huntley, the network manager for the Washington state Secretary of State's office who oversees Web sites that post election results in real time. Huntley's problem was that on election nights, traffic spiked so high that the sites always crashed.
Two years ago, Huntley installed Marathon software instead of simply adding more servers and load balancers. The Marathon software created a set of three "logical" servers, each of which was actually comprised of three Web servers and a database server running Microsoft Corp.'s SQL Server. That created both fault tolerance and redundancy, so that no single failure -- be it server, storage or software -- could cause the system to crash, he said.
"It was really taking a beating on election night. We were getting thousands of hits a second. But it stayed up, and I was really pleased by that," he said.
Though EverRun FT and HA use a virtualization approach, they don't offer the consolidation and partitioning features of more traditional virtualization software such as VMWare or Xen, according to Jerry Melnick, Marathon's vice-president of engineering.
EverRun HA's list price starts at $7,500 per license, which includes one year of support. Marathon claims 1,000 customers, and grew its license sales 80 percent between 2004 and 2005.
Michaelson, from the Princess Margaret Hospital, is already planning an upgrade to EverRun HA. It will allow him to switch to four cheaper and more powerful 4-way Intel servers, an upgrade Michaelson hopes to finish this fall.