For those of us who have worked in the high-performance computing industry for some time, the annual Supercomputing conference is something of a reunion, since you either know everyone or are at most one degree of separation from all of the 10,000 or so people who attend.
Supercomputing is where people involved in scientific and technical computing have gathered for almost 20 years to meet and talk about what is new. This year, I thought I'd record some of my impressions from Supercomputing 2006, which was held this week in Tampa.
Why should you care about a geekfest for high-performance computing types talking about technology that does not affect you? As I have often said, high-performance computing technologies tend to lead the way into the mainstream.
These are the trends in technologies I saw that you might see from a vendor near you sometime soon as products make their way to the mainstream. Of course, I am assuming that most readers of this column are not from the HPC community. The following is not a complete list of the hundreds of demos from hundreds of vendors, but highlights based on trends. I am in no way endorsing or recommending any of the products or architectures discussed, just giving you my impressions.
Big Blue Thinks Big
IBM was demonstrating a combined cluster file system with hierarchical storage management (HSM) using GPFS and HPSS.
Big Blue has added a DMAPI interface (Data Migration API) to GPFS and showed this interface. IBM combined this with YottaYotta's technology to provide remote data mirroring using block-level mirrors across large geographies.
This seemed interesting for two reasons: the combination of a clustered file system and HSM is very attractive to sites running large clusters, and remote replication of data is also promising. This has some exciting potential as file systems grow across geographic boundaries while still being managed by an HSM.
Prepare for iWARP
The Internet Wide Area RDMA Protocol (iWARP) is an Internet Engineering Task Force (IETF) update of RDMA TCP/IP that allows what is termed a zero copy TCP/IP (and is different from the late 1980s DARPA project called Iwarp). Currently, most TCP/IP stacks move from the NIC using multiple memory copies.
iWARP has the potential to allow remote direct memory addressing over WAN connections along with the ability to have low latency connections between local machines. Since iWARP runs on top of standard Ethernet, it could be used for the interconnect fabric between clustered computers for parallel applications. Today this is usually done with InfiniBand connections at a much higher cost than Ethernet connections; of course, this is 10GbE, not standard 1GbE.
With large clustered systems — say, over 4096 nodes — you have many levels of switch topology, and 10GbE switches with large port (512+) counts are being developed and should be available next year. If iWARP takes off, this would create a couple of advantages. For one, large core switches could allow high-speed connection to large numbers of nodes instead of using smaller switches and having to manage, maintain and architect a complex switch fabric with multiple levels of switching fabric. This could also create a single management fabric, since you could both manage the systems and communicate over standard Ethernet.
If iWARP becomes accepted, then it will not be too long before we see iWARP chipsets on higher-end PC boards. We all know that commodity parts always win in the market over specialized parts that have to be modified and maintained. For example, it was only a few years ago that people were talking about "Fibre on," meaning Fibre Channel chipsets on PC boards. If I remember correctly, IBM sold PC boards with Fibre on for a while. Fibre on did not take off, since the majority of PC owners did not need or want Fibre Channel. I certainly don't know anyone who has a Fibre Channel RAID for their home PC. So for the mass market, the cost was too great and IBM soon dropped this. Fibre Channel is relegated to HBAs placed into PCI (e and/or X) slots.
I just looked at the Dell Web site, and at least today, Dell's high-end home gaming machine supports Gbit Ethernet. What this means, in my opinion, is that 1Gbit Ethernet is now truly a commodity part. Can it be long before we see 1Gbit Ethernet ports on Wireless routers? I think not. If 1Gbit technology is now commodity even for home systems, it is not a big leap of faith that we will see the same for 10Gbit. If this happens, the cost of 10GbE is a commodity, just like we saw for 10 BaseT, 100 BaseT and 1GbE. Think about some of the advantage for iWARP for home use — downloads and gaming — in terms of CPU usage. I believe this is a technology that should be watched.
Two companies — DataDirect Networks and Linux NetworX — announced some interesting product combinations.
DataDirect unveiled what it claims are the highest-performing and most dense cluster storage solutions, the S2A Cluster Storage Solutions. The solutions are specifically designed for high-performance, high-capacity network cluster applications for the "best price-per-performance, best price/capacity-per-square foot solutions."
Linux Networx announced performance-tuned supersystems, the LS-P Series of turnkey, production-ready systems offering high application throughput and significant reductions in total cost of ownership (TCO) for leading product design applications. LS-P systems are performance-tuned for computational fluid dynamics (CFD), crash/impact analysis and structural analysis applications.
What is interesting about these two announcements is that both vendors are combining offerings from a number of vendors and packaging software with hardware. You get one-stop shopping for a solution and one finger to point. HP has been doing this for a while, and it seem like a number of other vendors are following that trend. It sure takes the burden off customers, since the vendors are doing the same thing over and over again. You are seeing this trend in the commercial world also, where Oracle is now providing Red Hat support.
And In Other News...
Here are some of the other interesting announcements from SC06.
SGI is moving into the cluster computing market using large core counts and InfiniBand on the board, with the new Altix XE1300 Ultra-Dense Cluster built with the new quad-core Intel Xeon Processor 5300 Series. Based on the "Atoka" board co-designed with Intel and Supermicro, the system integrates 16 cores, InfiniBand and Gigabit Ethernet in a single 1U high chassis.
The newly launched Cray XT4 supercomputer (formerly known as "Hood") and the Cray XMT massively multithreaded system ("Eldorado") are both part of Cray's Rainier program, the initial phase of the company's Adaptive Supercomputing vision that will bring all Cray's processing technologies together onto a common infrastructure. The Cray XT infrastructure provides a common user environment for login, compilation and running scripts, a common resource manager and work scheduler, a unique high-performance parallel file system and various intersystem interfaces.
These are two very different types of systems. The XT4 uses AMD processors connected via a proprietary interconnect. The other system uses custom made chips to allow multiple threads to be simultaneously executing.
HP announced improvements to its Unified Cluster Portfolio for high-performance computing. Product updates include support for new HP servers and workstations powered by quad-core Intel Xeon 5300 series processors. The HP Cluster Platform 3000 system now runs new quad-core processor-based HP ProLiant servers that offer customers greater energy efficiency and can improve system performance by nearly 50 percent.
The More Things Change...
The Supercomputing conference was not much different this year than last year, the year before or even 15 years ago. New products, new software, new offerings happen on a yearly basis, but Supercomputing is often not about all of that, but about seeing and being seen. If you or a company does not come to Supercomputing, people wonder what's wrong. The result is a pretty consistent picture of trends and developments in the industry.
Over the years, a number of trends have been shown at Supercomputing first, such as: clusters and rack-mounted computers for customers; Fibre Channel; InfiniBand; and 10Gbit Ethernet.
I am sure vendors show these technologies at other shows, but you'll be hard-pressed to find them all at the same show.
Henry Newman, a regular Enterprise Storage Forum contributor, is an industry consultant with 26 years experience in high-performance computing and storage.
See more articles by Henry Newman.