Print
Hits: 7793
Grids, Clusters and Making Money

Earlier this week I attended High Performance on Wall Street in New York City. The conference was well attended and interest levels were quite high. Of course, clusters have become the workhorse of Wall Street due to the shear magnitude computation they need to complete right away. After all, time is money.

Although I had limited time, I managed to find one piece of significant technology news and three presentations worth mentioning.

The technology news comes from PathScale who was pleased to announce that the InfiniPath adapter has been shipping since August and has been picked up by Voltaire in an OEM agreement. A press release is available here. This news represents a significant change in the InfiniBand market as now there are two families of "InfiniBand" adapters available for clustering. And, as everyone knows with clusters, choice is a good thing. Previously InfiniBand HCAs were only available from Mellanox derived technology. Some may argue that InfiniPath is not true InfiniBand, but in a sense, who cares, a full complement of APIs are available (e.g. TCP, MPI, etc) and all the user cares about is that their MPI application runs fast.

For those of you not familiar with the InfiniPath HTX Adapter, it has the lowest available MPI latency (1.32 μseconds) and lowest half-bandwidth message size in the industry (385 bytes), as well as providing full InfiniBand 4X-level bandwidth (952 MB/s peak uni-directional, 1875 MB/s peak bi-directional). As you might infer, the InfiniPath adapter also uses standard InfiniBand Switches, further leveraging the commodity market.

As for the presentations, you can find the presentation materials on line at the conference site. The program included a mix of panels and presentations. There were three presentations that I found interesting. The first was by Adrian Kunzle, of JP Morgan Chase Investment Bank. His message was loud and clear that Grid computing will be a big win for Wall Street. And based on his arguments and examples, I would agree. The only issue I have is that although "grid" was the focus of the show, I must have talked to at least five people who asked "What would you consider a grid?" or "What is the difference between a grid and cluster?" This question was a recent discussion topic on the Beowulf Mailing List. The jury is still out on exactly constitutes a grid, but some background can be found in this ClusterMonkey article.

The second interesting talk was by Alan McCarter & David V. Gelardi of IBM. For me the real telling slide came when they compared and Intel Xeon cluster to a Blue Gene cluster that had the same theoretical peak performance of 11.4 TFLOPS. The Xeon system requires 20 racks, 315 square feet, and 210 KW of peak power. The Blue Green System requires 76.5 square feet, and 56 KW of peak power. When the electric bill arrives, I'll let you do the math. Although the Blue Gene systems are not really commodity clusters, the need for FLOPS/Watt rating has never been more important.

The other interesting talk was by Kurt Ziegler of ASPEED Software. I have read about ASPEED in the past and know that they have been doing much of the grunt work in the HPC financial sector. Kurt outlined a collection of case studies where ASPEED was able to migrate and run applications in parallel environments within a reasonable time frame and cost using ASPEED technology. Of all the talks, I attended, this was the one that actually showed real work getting done on real problems. No offense to the other speakers, but we all know grids and clusters are great things. It is nice to know somebody is putting them to work. (And, I am sure there are others)

Finally, I did walk past the table for the new Itanium Solutions Alliance. I recall how Itanium was heralded by Intel and HP as the next generation of CPU. It was going to allow HP to use "commodity technology" to deliver HPC performance. I also realize there are many companies/competitors with a vested interest in Itanium that are now working together, but I often wonder, what's the point.

The next show of this type is the Linux On Wall Street conference in April of 2006.