With Sun Microsystems' recent introduction of its new Opteron-based system, including the eight-socket, 16-core Sun Fire X4600, the industry witnessed a Tier 1 OEM significantly expand its commitment to the x64 (64-bit x86) server market. Sun wasn't the first vendor to venture into the eight-socket Opteron space. Silicon Valley start-up, Fabric7 Systems Inc., introduced its x64 eight-socket, 128 GB machine back in November of 2005. Fabric7 actually has two offerings — the Q80 server, their basic eight-socket model, and the Q160 server, the company's high-end machine, which scales to 14 Opteron processors and incorporates a built-in low-latency, high-bandwidth I/O fabric.
When the Q80 and Q160 were unveiled last year, Fabric7 had only enterprise server customers on their minds. The company was — and still is — primarily focused on businesses providing financial equity trading and financial services, media entertainment, content distribution, telecommunications and web services. Using the commodity components — the AMD Opteron processor and the Linux and Windows operating systems, Fabric7 is aiming for the sweet spot in the mid-range enterprise server market.
But the company's offerings are not just vanilla x64 servers. With its Q-Par hardware partitioning capability, Fabric7 systems can dynamically carve the eight-socket server into two four-socket servers or four two-socket servers. Until the company made hardware partitioning available on their x64 systems, this capability was restricted to high-end mainframes, typically RISC machines running some flavor of Unix. Hardware partitioning is an efficient mechanism for sharing computational resources within a single box and avoids the performance penalty associated with software virtualization. However, software virtualization can run on top of the hardware partitions, adding another layer of resource sharing.
Resource virtualization allows datacenters to consolidate hardware, while increasing flexibility. The growing popularity of this model is encouraging the IT industry to focus on the type of architecture that can best accommodate it. Customers are starting to wonder if they can get a different dimension of computational performance and I/O throughput with a smaller collection of fat nodes rather than with a larger collection of skinny nodes.
“Traditionally, or for the past several years, it's been a build-out of lots of little 1U pizza boxes,” explains Bryan Sweeley, VP of marketing at Fabric7. “Customers have been very successful taking their applications and breaking them down to run across hundreds or thousands of servers. Now we are starting to see the pendulum swinging back again — people are looking at larger SMP and larger memory footprints.”
Larger SMP systems can make sense for HPC workloads as well. The recent deployment of Tokyo Tech's 38-teraflops TSUBAME supercomputer, which consists of 655 Sun Fire X4600 machines, is an example of building a large high performance system from a relatively small number of fat nodes. The fewer the nodes, the less you have to rely on expensive I/O switches and comparatively slow system interconnects to shuttle data around. A smaller number of boxes also means fewer individual pieces of hardware and software to manage and maintain.
The Tokyo Tech machine has not escaped the notice of other supercomputer users who are exploring different ways of scaling out systems. Sweeley revealed that Fabric7 has been contacted by a number of unnamed HPC users wanting to know more about what the company can offer.
“We've had some calls from the HPC crowd and have a couple of prospects cooking,” says Sweeley. My sense is that there is a trend developing in HPC to move towards the larger SMP footprints. And so the conversations we're having with some of the supercomputing installations tend to be focused around that.”
A single eight-socket machine may be powerful enough to run some types of industrial HPC applications such as electronic design automation (EDA). One of Fabric7's customers, Magma Design Automation, needs the large memory space — in their case 64 GB — to support their EDA applications. The larger memory footprint is used to improve performance for their full chip design runs. Less demanding test runs may require a hardware partition of only two processors and this can be reconfigured on the fly in 10 minutes.
Though both the Sun and Fabric7 systems share the same eight-socket Opteron architecture, the resemblance basically ends there. Besides the hardware partitioning capability, the other big feature of Fabric7 machines is their high-performance network I/O, that is, the fabric. It uses four HyperTransport links for I/O, as compared to two for the Sun Fire X4600.
On the Q160, Fabric7's high-end machine, the native interconnect supports both Ethernet and Fibre Channel. According to the company, the fabric supports 128 Gbps of non-blocking, switched I/O that provides up to 40 Gbps bandwidth to the processor and memory complex. The switched I/O can be extended across the fabric to provide 30 Gbps of Ethernet (or 16 Gbps Fibre Channel) between chassis or across sites. A software switch can dynamically reconfigure the Fibre Channel and Ethernet bandwidth, creating a flexible interconnect. For many enterprise applications, this is probably more networking capability than they can use. But to the extent that I/O is important for HPC applications, Fabric7's implementation of a high performance, flexible network makes for an interesting differentiator.
“Our fabric is a blend of 10 Gigabit Ethernet with a low-latency InfiniBand-like technology, says Sweeley. “On top of that, we can run Fibre Channel over the same fabric. For the HPC folks, we have more networking capability than they've seen from anyone else.”
Both Sun and Fabric7 are following the AMD Opteron/HyperTransport technology roadmap pretty aggressively. These companies are pushing the envelope with their 16-core offering, although this represents only a fraction of what will be possible in the next year or two . When the Opterons go quad-core in 2007, OEMs will be able to build 32-core SMP servers. When another HyperTransport link is added, the number of supported sockets will go from 8 to 32, allowing for the possibility of a 128-core machine. With Sun scaling up its Opteron systems, Fabric7 will face some tough competition from a Tier 1 OEM. But the team at Fabric7 seems comfortable with the prospect.
“Right now Sun and Fabric7 are the only two vendors with an eight-socket Opteron [system],” says Sweeley. “That's a healthy competitive environment.”
As far as Fabric7 pursuing the high performance computing market more aggressively, that probably depends on how the larger SMP machines are perceived by HPC users over the next six to eighteen months. The company certainly seems open to the possibilities. When asked about what Fabric7 offers high performance computing, Sharad Mehrotra, the company's president, CEO and founder, had this to say:
“Fabric7 does think that the pendulum will swing back towards larger SMP servers. The Tokyo Tech example is the lighthouse that is making heads turn and prompting people to question the conventional wisdom of today. Fabric7's implementation of hardware partitioning across its entire product line, provides HPC users with the flexibility to move quickly from small SMP to large SMP configurations in a matter of minutes. Additionally, the switched, virtualized I/O capability available in our larger system, the Q160, provides customers the flexibility in network infrastructure that is required to keep up with the variable compute needs of the processing farm. We believe that the HPC world will shift from 'grid' to 'fabric' computing in the future as the benefits of our approach become more apparent with real-world deployments.”
We will be providing an extended interview with Fabric7 CEO Sharad Mehrotra during our special coverage of the LinuxWorld Conference and Expo, which takes place August 14 – 17, in San Francisco, California.