The emergence of commodity supercomputing has driven clusters based on the Linux operating system into engineering and scientific research organizations that couldn't afford their own supercomputing resources before. But Linux clusters have the potential to become a hot topic in traditional information technology circles as well. The coming year may well be the technology tipping point when Linux cluster technology escapes its current home in research organizations and inhabits the traditional data center.

Building Clustered Linux Systems by Robert Lucke, recently published by Prentice Hall Professional Technical Reference and HP Books, attempts to provide a starting point for organizations interested in building or evaluating their first Linux cluster.

We took this opportunity to have a chat with Robert and ask him a few questions on the subject of Linux and clustering.

What made you want to write a book about building clusters?

Before starting work on clusters, I spent a considerable amount of time tackling workgroup architecture and large-scale system administration problems for my engineering and scientific customers. When I got the opportunity to work on a prototype Itanium 2 cluster at Pacific Northwest National Laboratory, I was fascinated by the new techologies like the high-speed interconnect from Quadrics, and recognized many familiar management and architectural issues. The more I learned, the more I saw applications for clusters in other more "traditional" areas. The book itself was a learning experience for me and an attempt to collect and organize cluster-building information for organizations that are investigating clustered solutions.

Why do you think that clusters are an important architecture?

If you have the proper application software, a cluster can scale high-performance, high-availability or high-throughput resources far beyond anything that is available in a single SMP system. Being able to do this with commodity hardware brings tremendous compute resources in reach of organizations that previously couldn't afford them. I see cluster architectures and techniques as the gateway to some of the resource virtualization that seems to be the Holy Grail of traditional IT departments today. I think this is exciting! I guess it's the love of finding an elegant solution to a problem that drives my excitement. Instead of being a marketechture, clusters represent a real solution to a group of scalable problems.

When someone says cluster, what does it mean to you?

I have learned to be very, very careful with the word cluster. It's overloaded and the meaning depends on the audience. Using cluster in a scientific context evokes a different mental picture than if it were used by traditional IT folks. In general, I think of a cluster as a group of separate resources like systems, CPUs and RAM that gets poured into a mold. The shape of the mold determines the final shape and behavior of the clustered solution. Sizing the problem and determining the shape of the mold is the fun part for me.

Besides scientific and engineering environments, do you see any other applications for clusters?

I sure do! There are database clusters, web server clusters, file server clusters, visualization clusters, and on and on. Instead of building clusters that push the upper limits of RAM and CPU resources, thousands to tens of thousands of CPUs, for example, the company I work for concentrates on application-specific clusters. These are smaller, single-function clusters that are meant to run application configurations that would have required a large and expensive SMP system. The intent is to lower the complexity of building and managing the cluster, but still provide a more cost-effective solution for the application. I think this type of approach is generally applicable in any type of computing environment.

 What are some of the common mistakes you've encountered in cluster building?

The single biggest problem I run into is what I call pile o'hardware syndrome. That's the notion that you just buy a whole bunch of cheap hardware, rack it up, and a cluster will magically appear out of a pile of pieces. It's still very common to underestimate the amount of work required to make physically separate resources work together as if they were one very large, manageable SMP system. A cluster is still a systems engineering problem that can turn nasty if you aren't careful. But, with advances in pre-racked, pre-cabled hardware from some of the hardware vendors and the cluster software toolkits like OSCAR and Rocks, I see cluster building getting easier all the time.

Why do you think that Linux is the best cluster operating system?

One simple answer is choice. There are commercial distributions, free distributions, white-box distributions and so forth. If you have a commercial software package like an Oracle database that's qualified against a particular Linux commercial distribution like SuSE or RedHat, you can build a fully supported cluster configuration. If you want to do research or custom work, there are free distributions like Debian or Fedora. Because the source code is available, you can choose your starting point and degree of customization. This is the best of all possible worlds.

The Linux operating system is stable, manageable and flexible. You are free to configure Linux as you see fit instead of trying to chip away at a black-box operating system that fights you every step of the way. There's a wealth of free management and development tools available. Oh, did I mention that Linux runs on a wide range of commodity hardware, both 32- and 64-bit? What's not to like? Nothing else comes close in my estimation.

What do you see ahead for clustered architectures?

I definitely see Linux clusters moving into mainstream information technology environments. If you look back, the scientific community tends to drive computing technologies that are later adopted by the more conservative IT organizations as business solutions. One modest example I can think of might be the World Wide Web and the Mosaic browser. I firmly believe that clusters, specifically Linux clusters, are poised to repeat this type of adoption pattern. I think we are very close, if not past, the tipping point.

What would you say to someone who is thinking about building his first cluster?

Do it, but do it with your eyes open. Do your homework before starting. Give yourself time to learn. Try not to fall into the pile o'hardware trap. Start small and scale up. Investigate starting points like openMosix, Rocks and OSCAR first. If you don't have time for the learning curve, then have a replicable solution designed and implemented for you.

Conclusion

About Rob Lucke