Server clusters move toward mainstream

Server clusters have long been the domain of scientific computing, but as the technology matures and becomes available on more platforms, enterprise users will start reaping the rewards it has always offered, such as flexibility, availability and high performance.

Some recent examples of how this technology is moving to more mainstream servers include Dell partnering with Cray to deliver Intel-based servers with Cray's powerful clustering software; Compaq's plans to deliver cluster-technology on Intel's IA-64-based machines; and Stratus Technologies coming out with turnkey appliance cluster servers suitable for users who need simple failover.

Observers say what makes clustering attractive is its ability to let users string together multiple servers and have them act as one to provide the power needed to run compute-intensive applications, such as those used for financial transactions, database queries or back-end applications. Users also can configure clusters so that they don't have a single point of failure - in fact, multiple machines can fail without bringing down the cluster. And users can configure clusters in physical or virtual subsets, assigning them to work on specific jobs when they are needed.

User clusters

A good example of mainstream clustering technology can be found at Google. The search giant has more than 10,000 home-built low-end PCs - many clustered to handle large workloads. The distributed nature of the environment ensures that no single server going down can compromise the whole.

Google also employs automated management and provisioning tools that let network managers update code and monitor machines efficiently. Jim Reese, director of operations at Google, says the company has achieved flexibility and availability through clustering. He can dynamically provision systems based on site traffic on the fly, and by having many systems shoulder the load, redundancy is a given.

"One of the biggest benefits we get from clustering is the ability to dynamically allocate resources where we need them, because at any given time, we have to respond to page hits that can fluctuate quickly," Reese says.

At the other end of the scale, the world's largest nonmilitary computer - a cluster of 750 AlphaServers with a total of 3,000 processors, which runs at the Pittsburgh Supercomputing Center (PSC) - helps users run multiple projects simultaneously.

"The point of having this system is that we can use the entire cluster, or parts of it, to work on certain projects," says Beverly Clayton, executive director of PSC. An automated resource manager lets IT staff partition the system as needed, and because the single virtual system actually consists of many physical layers, an outage won't threaten downtime.

Vendors are building on examples such as Google and PSC looking to add features that promise to make clusters more efficient for everyday use. Features such as partitioning let users quickly add resources such as memory, processors and storage. Sun's recently announced Sun Cluster 3.0 software, which is designed for use with the company's Solaris operating system and clustered servers, lets customers partition servers to designate resources to help meet peaks and valleys in demand for applications and data, observers say.

Compaq's TruCluster is another good example of the mainstreaming trend. TruCluster is clustering software based on OpenVMS, Compaq's high-end operating system acquired through its acquisition of Digital. TruCluster now can run on Alpha-based systems, and the company plans to offer it on Intel's 64-bit processor-based systems in the future, says Mark Silverberg, technical marketing manager for Compaq's high-performance systems group. TruCluster lets users tie together clusters of Alpha-based servers and manipulate them as a single virtual supercomputer or partition them to work simultaneously on different jobs. The ability to manage many systems through one view cuts down on management and maintenance headaches and simplifies operations.

Silverberg offered other tidbits about product direction - if Compaq merges with Hewlett-Packard, the two companies will offer a hybrid server-cluster product designed to run on IA-64 - Intel's next generation of 64-bit processors. The move to offer cluster technology on Intel-based systems should give users easier access to the mainstream applications they will want to run on them, such as Oracle's database software, which requires lots of horsepower.

Dell also recently jumped into the clustering trend, partnering with Cray, a longtime provider of high-end server cluster gear. Cray will use Dell's PowerEdge servers along with its own HPC cluster technology. The Dell eight-, 16-, 32- and 64-way systems will be Intel-based boxes, offered at a significantly lower price than traditional supercomputer clusters, with an eight-server system starting at $75,000.

Stratus is also lowering the bar. The company already offers fault-tolerant systems for the Windows operating system environment, but now is taking that approach a little further. Patrick Farley, director of business development for Stratus, says the company plans to offer a turnkey appliance server. The new four-way system actually will function as a fully redundant two-way server, with the two extra processors as backups. The system ideally is suited for places such as bank branches, and retail and remote offices, where IT staff may be nonexistent but uptime is a priority.

High end, low end

Tom Manter, an analyst at Aberdeen Group, says that new cluster products, designed to offer features traditionally reserved for high-end systems, will go much further than the first Intel-Microsoft-based options available since the mid-1990s via Wolfpack. Wolfpack allowed for two cluster nodes - something that severely limited the scalability users could gain.

Manter also notes that InfiniBand, the next-generation switch fabric I/O technology, will bring clusterlike functionality to users via features such as shared direct memory access, which will let customers pool memory and access it from multiple computers.