In the HPC market, Dell has established itself as the number three system vendor, trailing only its larger competitors, HP and IBM. Known for offering no-frills performance servers at reasonable prices, Dell has garnered a particularly strong following in higher education and government labs, especially for small and mid-sized clusters. But a recent spate of purpose-built HPC products from the company points to a subtle shift in Dell’s high performance computing strategy.
During a recent conversation with Donnie Bell, senior manager of HPC Solutions in the Dell Product Group, and Tim Carroll, Dell’s HPC Global Lead, the two reps outlined how the company is treating HPC more as a distinct opportunity, and less like an extension of their enterprise business. The result is that Dell has developed more HPC-specific products and is backing that up with more system testing and validation prior to deployment. “It’s not just about throwing gear out there,” explained Bell. “It’s got to be the gear that they want, put together in the solution they want.”
The shift in strategy has come about over the last three years. Attracted by the bullish HPC market (or at least bullish forecasts thereof) and a seemingly untapped demand for high performance computing, Dell is focusing particularly on the so-called “missing middle,” a term the Council on Competitiveness came up with to identify the potentially large group of unserved users between entry-level and high-end HPC practitioners. “That’s the market that Michael [Dell] said we’re going to invest in,” said Bell.
Of course, what this class of users ultimately wants are turnkey systems that are as easy to use as their desktop systems and don’t require an advanced degree in high performance computing in order to maintain. So far this is beyond the reach of Dell, as well as any of its competitors. Making HPC clusters act like appliances is still the stuff of fantasy.
Where Dell is staking out new ground is in its product mix, which now includes a range of HPC-centric offerings. It wasn’t too long ago that the PowerEdge 1950 was the workhorse server for Dell’s HPC customers. For all intents and purposes, though, the 1950 was an enterprise server pressed into HPC service by necessity. Today Dell offers servers and blades aimed specifically at the performance sector, including the latest HPC-friendly gear: the PowerEdge 6100, M610x, C410x.
The C6100 is the company’s new HPC workhorse, an ultra-dense rackmount server that encapsulates four dual-socket nodes in a 2U form factor. It offers twice the density of an average dual-socket server and is even 20 percent denser than blades. Dell accomplished this feat by sharing the internal infrastructure: power supply, fans and backplane. You can service the nodes individually, and the hard disk drives (either 2.5″ or 3.5″) are hot-pluggable.
The C6100 is available with either Intel “Nehalem” 5500 or “Westmere” 5600 processors. Outfitted with 6-core Westmere CPUs, a single 2U box will deliver 48 cores. Because of its density and power, it’s specifically targeted as a building block for HPC clusters, but can also be used for general Web and cloud installations, where maximum performance is a priority. The C6100 has been shipping since the spring.
Dell recently announced C6100 deployments at the University of Colorado and University of Kentucky. Both systems will be supporting a range of scientific research at those institutions, including climatology, genomics, energy studies, pharmaceutical design, and physics, among others. The Colorado system is big enough to warrant the number 31 spot on the TOP500 list.
The brand new PowerEdge C6105 is the AMD counterpart to the C6100, offering Opteron “Lisbon” 4000 series processors in the same dense 2U enclosure. The 4000 Opterons are the less performant, lower power siblings to the Opteron 6000 processors, so the C6105 is geared more toward the large-scale cloud and Web 2.0 deployments than strict HPC. Availability is still a couple of months away.
On the blade side, the dual-socket PowerEdge M610x is an M610 variant for HPC that includes two x16 PCIe Gen2 slots and two I/O mezzanine cards. (The M610, by the way, is the building block for the newly announced 300 teraflop Lonestar super at TACC.) The PCIe slots on the M610x lets you install a single NVIDIA Tesla (Fermi-class) GPU card, if you want to accelerate data-parallel workloads; or perhaps a Fusion-io ioDrive Duo, if you’re looking for ultra-fast storage. The two mezzanine slots makes dual-rail InfiniBand a possibility, but you also can slot in Ethernet, Fibre Channel, or whatever networking combo you might desire. Like the C6100, the M610x is available with quad-core Xeon 5500s or six-core Xeon 5600s.
Because of the extra connectivity options, the M610x is a full-height blade (unlike its half-height M610 sibling), but still fits neatly in Dell’s M1000e blade chassis. The new blade was announced in June and has been shipping for a couple of months.
If a single GPU per server isn’t enough, Dell is now offering the PowerEdge C410x, a CPU-less 3U box that can house up to 16 Tesla M2050 GPU modules. As of today, that represents the biggest commercial GPGPU box on the market. At the maximum 16-GPU configuration, the C410x can deliver 16.5 teraflops of raw performance.
Of course, tapping into that requires a CPU host, so the C410x conveniently allows connectivity for up to 8 servers. The idea here is to decouple the CPU and GPU so that a customer can mix and match the processor ratios as needed by the application. This could be especially useful in those cases that can take advantage of a high GPU:CPU ratio, like some seismic and physics codes, or where the work is such that the optimal processor ratio varies from one application to another.
If you’re getting the idea that Dell is a little GPU-happy these days, you’re right. According to Bell, the company believes a lot of their HPC customers will be opting for GPU acceleration now, as they chase ever denser performance. Even the new Dell Precision T7500 has a slot for a Tesla C2050 GPU, for those CUDA desktop apps that need a few hundred extra gigaflops to really shine.
“Quantitatively, there are so many more thousands of researchers doing their work on desktops,” said Dell’s Tim Carroll. “But it’s only a matter of time before those people are performing their research on a server somewhere, whether it’s their own, the institution’s, or in the cloud.”
Whether Dell’s new HPC investment yields big dividends is difficult to gauge. Because of the sharp downturn in the global economy in the last couple of years, IT spending has dipped considerably, although less so for HPC. According to Carroll, though, Dell’s HPC business is “seeing growth across the board,” adding that the market seems to be really breaking loose over the last three to four months.
The latest IDC numbers for 2009, which splits out HPC system revenue by vendor, has Dell with a 12.7 percent market share. That’s about half that of IBM’s share at 29.6 percent and HP’s at 28.2. But for mid-sized (departmental) systems, Dell is at 29.8 percent, edged out only by HP at 35.6 percent. That’s a good starting place, especially considering that the size of the HPC pie is forecast to start growing again now that the recession seems to be easing.
Despite the evolution in strategy, Dell still relies on partnerships with vendors like Platform Computing and Terascala to fill in the cluster management and HPC storage pieces of their solution, respectively. And even though the cluster maker is now designing purpose-built HPC systems, it is doing so to fulfill established market demand, rather than for the sake of invention. Contrast that with former HPC maker Sun Microsystems, and its enthusiasm for building exotic hardware, like 3,456-port InfiniBand switches and proximity communication chips.
Dell’s much more conservative innovation strategy is designed to serve the large sweet spot in the middle of the performance market, relying on the acceleration of HPC demand to drive revenue. According to Carroll, the company is still fundamentally about delivering open standards-based commodity clusters, adding, “we want HPC to be widespread and we want to be the ones who deliver that.”