First in a series of four articles about the TeraGrid

Deep, wide, open. This three-pronged conceptualization underlies the TeraGrid, the National Science Foundation’s cyberinfrastructure initiative. “Deep” means digital muscle—more than a petaflop of aggregated computing power, highlighted by the addition of NSF “track 2” systems, Ranger (579 tflops) and Kraken (607 tflops). “Open” means extensibility, the ability to include new resource providers and university partnerships to broaden the resource base.

“Wide” means that the TeraGrid wants its resources to be useful to as many researchers as possible. To that end, TeraGrid has created “Science Gateways” -- diverse entry points for the uninitiated to pass into the realm of computational science and get things done with the array of resources available through TeraGrid. Implemented in 2005, the Science Gateways program, led by Nancy Wilkins-Diehr of the San Diego Supercomputer Center, has grown rapidly and now comprises 35 Gateways—each of them tailored to the needs of and designed by a specific research community.

The garage sale (or should I say tag sale given the New England location) that is SiCortex is disheartening. I worked with SiCortex enough to know that it was "SiCortex" not "SciCortex" as so many were apt to write. I also knew that these were smart people. They were doing the good work. There was no smoke and mirrors, no shining up yet another mass market turd, just bright people delivering on a daring idea. We need more of this not less.

The idea was simple. Instead of doing HPC with clusters employing faster and hotter processors, use many power efficient processors and a great interconnect. Oh yea, and use Linux from the ground up. Keep it open, keep it right. And, it worked quite well.

I have in the past talked with marketing people who seemed clueless about what they were really selling. Not SiCortex. They lived and breathed their technological value proposition. I could tell when I wrote a white paper for them. Theirs was not a "me to" product, nor was it another 1U server or blade with the latest x86 platform in it.

I suspect that the demise of SiCortex is more about the inability of the venture firms to fund the company than their ability to sell supercomputers or push the envelope. They had not yet turned a profit, but seemed to be on their way. I wish the employees of SiCortex a good transition and thanks for being brave.

Dammit.

A few pieces of news crossed my way recently. First, AMD has released the x86 Open64 Compiler Suite (binary and source). This is a free as in beer and in speech compiler suite that is the basis for the PathScale Compiler. AMD also provides a collection of libraries and HPC applications that can be built with the compiler (instructions on how to build the packages are provided.)

While we are talking about compilers. I also found a nice bullet point overview of OpenCL (pdf). If you recall OpenCL is a new language that is designed to be portable across GPU and CPU architectures. It even has a simple FFT example. As I have said in the past, things like CUDA, OpenCL, and BrookGPU are nice, but they don't cover the cluster computing model. And, it is step in the right direction.

Finally, here are two papers (pdf) that discuss using the cloud for HPC. They even include benchmarks! Take a look at Benchmarking Amazon EC2 for High-performance Scientific Computing and Can Cloud Computing Reach The TOP500?. Don't sell your cluster just yet.

Let's give a warm welcome to HPC Community. You may have noticed the web feeds from HPC Community on the right side of the main page. The simians here at ClusterMonkey are working with HPC Community to help build a bigger/stronger community and nothing builds community like free software! HPC Community is the home to a pile of cool software projects. The two most notable are Kusu and Lava. Of course, they have to be good because in addition to open code they have cool names and logos.

Kusu is the foundation for Platform Cluster Manager (previously known as Platform Open Cluster Stack OCS 5), is a standardized approach to easily build, manage and use Linux clusters and is a freely available cluster distribution!

Platform Lava is an open source entry-level workload scheduler designed to meet a wide range of workload scheduling needs for clusters up to 512-nodes.

Check out both projects and more at the HPC community site. And if you are wondering what the name Kusu and the little turtle are about, just ask Why the Turtle?