Sun Revamps HPC Offerings
Even as analysts and customers wonder whether Sun Microsystems will continue to survive on its own — or whether it wants to — the company continues to push new products out the door. On Tuesday, at Sun’s Parter Summit in Las Vegas, the company introduced a number of new offerings, mostly centered on the recently launched Xeon 5500 (Nehalem EP) chips. The products aimed at the HPC space include a rack server, two new Constellation-class blades, a Lustre-based storage system, a number of rather interesting InfiniBand products, and a Sun cooling door. According to Michael Brown, Sun’s marketing manager for HPC, the upgraded product set represents “almost an end-to-end revamp of the HPC offerings.”
Perhaps the simplest new offering is the X2270 rack server, a 1U dual-socket box that incorporates the new Nehalem EP chips. The X2270 is basically an upgrade of the X2250, which used the previous generation Harpertown processors. The rack servers are targeted at mid-sized commercial HPC environments, such as you might find in financial services or electronic design.
But Sun has directed most of its engineering smarts at the Constellation blade systems, where the new Nehalem processors and Quad Data Rate (QDR) InfiniBand technology have been used to build a more advanced platform for high-end clusters. The key new product is the Nehalem-equipped X6275 blade, a dual-node blade, where each node can house two quad-core chips. (In essence, Sun built a four-socket blade with dual-socket hardware.) Doing the math, that means each blade provides 16 cores, and thanks to Nehalem’s simultaneous multithreading, up to 32 threads.
The blade fits into Sun’s 6048 chassis, and because it’s a dual-node setup, 96 nodes (768 cores) can be squeezed into a single 42U enclosure. At this maximum configuration, a single chassis can deliver 9 teraflops. Although that’s rather impressive, according to Sun’s own Web site, that would work out to only 0.8 teraflops better than an enclosure fully populated with the X6440, the company’s four-socket AMD quad-core blade. Note also that the Intel blade memory maxes out at 192 GB (that is, as soon as the 8 GB DDR3 server DIMMs hit the streets), while the AMD blade can house up to 256 GB, although the latter uses the somewhat slower DDR2 memory.
What really sets the X6275 apart are the new networking and I/O capabilities, which will allow the blade to inhabit petaflop-sized systems with thousands of CPUs. Each node includes an onboard QDR InfiniBand host channel adapter (HCA), Gigabit Ethernet, and a PCIe ExpressModule slot. A SATA interface is also available to connect to an optional Sun flash module, which offers 24 GB of high performance storage per node. It’s designed for users interested in saving state, having a scratch data area, or booting an OS. Since the flash module is hooked up to a SATA controller, to the apps it looks like a hard drive.
The other new blade is the X6270, which is less computationally dense and is geared for more general-purpose HPC and commercial duty. This one is a full-height blade that is a single-node version of the x6275, and can hold up to 144 GB of memory. Since it only has half as many cores as its dual-node sibling, the x6270 actually offers a better byte per flop ratio. It also provides four interfaces for on-board disks, with optional RAID, plus two GigE ports. The better memory ratio, additional I/O and extra networking make this blade more versatile, and it would tend to be a better fit where compute density is not the overriding factor, as, for example, in the head node of an HPC cluster.
Along with the blades, Sun announced a number of new Sun-branded InfiniBand products. The first one is a QDR InfiniBand Network Express Module (NEM) for the 6048 chassis. It’s essentially an InfiniBand leaf switch that can link up to 24 nodes. Since four of these modules fit in a single 6048 enclosure, all 96 nodes can be accommodated without any external switch hardware. The NEMs can be directly connected to datacenter InfiniBand switches in a fat-tree topology or to other NEMs in a 3D torus mesh. The goal here is to reduce cables and extra switch hardware in these ultra-dense blade setups.
Sticking with the InfiniBand theme, Sun also introduced a PCI Express QDR InfiniBand expansion module, which can provide a second QDR link via the PCIe interface. The additional link means you can have 80 Gbps of InfiniBand per node, which could be split between compute and storage, or simply aggregated for additional bandwidth.
In addition, the company previewed its “Project M9,” a 648-port InfiniBand datacenter switch. The hardware will be based on the same technology as Sun’s current 3,456-port switch used in TACC’s Ranger supercomputer cluster. According to Brown, the M9 will use 75 percent less space than traditional InfiniBand switches and will make use of 12X InfiniBand cabling, which will allow it to route three connections per cable. Once again, the idea is to minimize the hardware footprint. Brown notes the upcoming switch could be used to hook together non-Sun servers and storage.
Also announced was Sun’s new cooling door, which was previewed in November at SC08. The door fits in the rear of a 6048 chassis and relies on passive cooling, so no additional fans or power is required. There are two flavors: one that uses chilled water and one that uses a refrigerant gas. They are designed to handle a thermal load of up to 35KW per rack. Since studies show this type of system can reduce cooling costs by up to 84 percent, more and more datacenters are turning to liquid cooling to cut down on power consumption.
On the storage side, Sun has unveiled an integrated Lustre storage system, which is designed to scale from 48 TB up to multiple petabytes. The storage component options include the Sun Fire X4540 and X4250, and the Sun Storage J4400 and J4200 storage arrays. Expansion is accomplished by adding more storage modules. The idea here is to offer a pre-packaged Lustre solution for HPC apps. Since the system is not tied to Sun server gear, Brown thinks there’s an opportunity to sell these systems to HPC users whose systems are under-configured from a storage perspective. He says Sun has some early customers for the systems, but they haven’t gone public yet.
A number of customers have already signed up for Constellation supers based on the new hardware, including the Australian National University (ANU), Australia’s Bureau of Meteorology, South Africa’s Centre for High Performance Computing (CHPC) and the University of Zurich. These are in addition to installations of Nehalem-based blades at Sandia National Laboratories, Forschungszentrum Jülich, and RWTH Aachen University, which were announced back in November.
As far as how these new offerings will play in a depressed economy, Brown thinks that despite the current downturn, there’s still a lot of demand for high performance computing gear. “We’re seeing very strong uptake in the HPC area,” he says. “We’ve sold over two petaflops of HPC solutions based on Sun blade design.” Brown says he’s spoken with a number of people in the higher education sector that are applying for supplemental NSF and NIH funding that will be drawn from the US government’s stimulus package. Brown realizes that not all of that money will be heading to Sun, but he’s optimistic that the company will see its fair share.
Overshadowing all these announcements is the question of whether Sun plans to sell the business or go it alone. Since the IBM deal devolved into an April fool’s joke, Sun’s uncertain status has left customers wondering about the future of the company. Sun is not speaking publicly about its next move, so for the time being, it looks like the company will let its products do the talking.