Start-up Evergrid enters virtualization fray

Evergrid, a start-up coming out of stealth mode this week, is promising mainframe-like reliability for applications running on clusters of commodity Linux servers.

The company’s virtualization software is designed to separate applications from the operating system so that software processes can be saved at any point in time, thus reducing downtime and allowing applications to be shifted as necessary. Evergrid is launching at the SuperComputing 2006 show going on this week in Tampa, Fla.

“Today, complex parallel applications . . . are constrained by hardware,” says Dave Anderson, Evergrid’s CEO. Anderson was formerly CEO of e-mail security firm Sendmail and also served as CTO and general manager at Amdahl before coming to Evergrid.

When there are hardware failures, applications running in parallel on commodity clusters typically have to be completely restarted. In addition, there is no easy way to move applications among different hardware platforms, he says.

“We end up with silos and servers dedicated to a particular application because we don’t have the ability to move servers back and forth from one kind of application to another, so that leads to increased management costs,” Anderson says.

Evergrid tackles both the availability and resource-management issues with its Availability Management Suite of software, which includes Evergrid Availability Services for application recovery and restart and Evergrid Resource Manager for workload scheduling, according to Anderson.

The software inserts an application virtualization layer between the operating system and the application, without any modification to the application itself, Anderson says. It runs on physical servers as well as atop virtual machines created with software from VMware and Xen, adding “less than 5% overhead,” Anderson says.

“What we’re able to do with this application virtualization layer is we can do checkpoints of the distributed application without stopping the application and without stopping I/O,” he says. “We end up with a globally consistent snapshot of an entire set of nodes so if there is a failure we can reload all the nodes and restart the application from the point that we last did one of those snapshots.”