Computerworld - Russ Miller runs a monster of a server cluster that eats storage at an incredible rate. The bandwidth requirements alone on his 22TFLOPS system force Miller to look outside the storage box, so to speak, for better throughput and scalability.
As the director of the Center for Computational Research at the University at Buffalo, Miller oversees a supercomputer comprising 6,600 processors that is used by the university and many businesses in western New York.
To support all that computational power, Miller turned to a clustered storage system that could alleviate bottlenecks and automatically load-balance and grow on the fly to accommodate user demand.
Like many IT managers who have seen the benefits of server clusters, Miller chose to try the relatively new technology of storage clusters as a means of attaining a fully redundant infrastructure that's highly scalable and easy to manage. Clustering provides massive throughput because of an increased port count that comes from cobbling many storage servers together into a single pool of disks and processors, all working on a similar task and all able to share the same data.
Management functions are distributed across the storage server farm. To an application server, the farm looks like a single, block-level storage system. Storage capacity can be added without disrupting applications running on the cluster.
There's lots of talk about storage clustering among vendors these days, but few market leaders have fully embraced the concept, according to analysts. Most of the development is still being led by start-up companies such as Ibrix Inc., Isilon Systems Inc. and Intransa Inc.
In April, Miller selected a system from Dell Inc. and Billerica, Mass.-based Ibrix that gave him storage read rates of 2.3GB/sec. and about half that rate for data writes - far above what any monolithic storage array could produce, he says.
"We don't have any single points of failure. So if and when we need to make additional investments in storage, we can do that without any major downtime or major reconfiguration. We didn't see any downside to going this route," Miller says.
Tony Asaro, an analyst at Enterprise Strategy Group in Milford, Mass., agrees. "The beauty of some clustered architectures is you can start small and grow as much as you want," he says.
The University at Buffalo's storage cluster consists of three EMC CX700 storage arrays, each with 70 146GB drives that are managed with Ibrix's software.
"If one of these I/O nodes goes down, we won't lose anything except a little performance," says Miller.
Ibrix is a clustered file system than runs on hosts, but it can also run on storage arrays. For example, the internal disk drives on low-end Dell servers can be combined to create a storage pool. The result is a compute farm that also clusters its storage. "It adds no greater complexity by adding more servers to the cluster," Miller says.
Free Insider GuideCopyright © 1994 - 2012 Computerworld Inc. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of Computerworld Inc. is prohibited. Computerworld and Computerworld.com and the respective logos are trademarks of International Data Group Inc.
