SAN JOSE, Calif. -- PolyServe Inc., a Beaverton, Ore.-based clustering software vendor for Linux and Windows data centers, took network-attached storage (NAS) out of a box and put it in a Linux cluster this week at the ClusterWorld Conference & Expo 2004.
The PolyServe NAS Cluster integrates network file system (NFS) protocol functionality with a symmetric cluster file system, high-availability services and cluster and storage management capabilities, said Steve Norall, PolyServe's general manager of Linux Solutions. As many as 16 low-cost Linux-on-Intel servers can be clustered for high-performance, fault-tolerant file serving across a storage area network (SAN), Norall said.
Why NAS in a cluster and not in a box? NAS is a self-contained, single-headed appliance with its own storage disk built in, Norall explained. "The trouble is that the buyer will pay more for the NAS disk and additional disks than he would pay for the disks for a server," he said. PolyServe offers a software-only solution, so a company can leverage existing servers, eliminating in many cases the need to buy new hardware, according to Norall. "Just hook them up to your Fibre Channel SAN, and build a NAS cluster with this software," he said.
It's a good idea to put NAS file serving on commodity servers, rather than linking a couple of independent NAS filers for failover, according to storage analyst Arun Taneja, founder of the Hopkinton, Mass.-based Taneja Group. Clustering on industry-standard servers delivers better performance than standalone NAS at a much lower price point, he said.
Taneja estimates that it would cost more than $450,000 to buy the products and NFS licenses costs for clustering two NAS filers for 12.6 terabytes (TB) of enterprise storage. The price tag for a 12-plus TB PolyServe NAS cluster on two Intel servers with two gigabytes of RAM each would be about $80,000, according to Norall. Generally, the PolyServe NAS Cluster costs $3,500 per processor.
PolyServe promises that its software approach to NAS file serving will centralize and simplify management.
"It's easy to manage one NAS appliance but when it comes to a cluster of two or three, it's more complex," said Norall. "The system administrator has to partition the data and has multiple entities to back up and provision with new storage." As demands grows, the administrator ends up managing multiple entities.
With PolyServe's solution, you have one cluster, said Norall. "You manage the cluster as a single entity. So, you have a single pool of storage that you provision and that backs up and is backed up at SAN speed."
PolyServe's customer benchmarks show that a representative enterprise-class NAS failover cluster can deliver approximately 250 megabytes per second (MB/second) of highly available throughput performance based on sustained sequential writes to a single file system translating to a price/performance value of $1,904 per MB/second. The two-node PolyServe NAS Cluster delivers the same 250 MB/second of highly available throughput performance -- at a price/performance value of $317 per MB/second, Norall said.
"With these metrics, a three- or four-node PolyServe cluster would surpass a top-of-the-line NAS filer at a fraction of the price," Taneja said.
There will be training costs associated with moving NAS file serving out of appliances and on to commodity servers. "Initially, training on that kind of infrastructure is needed," said Norall. "This really is a new approach." He adds that performance gains and lower costs overall can offset training costs.
Moving NAS file serving to commodity servers will help speed adoption of clustering in mainstream businesses, according to Norall.
"People are waking up to the fact that clusters aren't just for the scientific community," he said.
Also, he noted, clustering is tightly linked to two dynamic and popular concepts: grid computing and the utility data center. Grid computing and utility data centers are being adopted in commercial data centers, not just in the high-performance computing sector. "At the heart, these concepts deal with multiple servers together and having them act as one; essentially, that's clustering," he said.