SGI Previews Next-Generation, Blade-Style Altix Supers
by Timothy Prickett Morgan
With the SuperComputing 2005 tradeshow raging in Seattle this week, a lot of the major players in the high performance computing market are making announcements, and Silicon Graphics, which has come under pressure to shore up its revenues in recent quarters, is hoping that a new line of blade-style supercomputers code-named "Tollhouse," due early next year, will help give it the sales bump it needs to turn the company around.
Like the previous Altix 300 and 3000 series machines, the future Altix 4000 machines are based on Intel's Itanium processors and use SGI's own NUMAlink NUMA memory clustering technology to create a single, global shared memory space for multiple Itanium nodes to share. There have been some rumors swirling around in the HPC industry that SGI, as part of its financial restructuring, might be thinking about making the jump from Itanium processors to Opterons from AMD. Jill Matzke, high-end server marketing manager at SGI, just laughed at this suggestion, then said "No" very curtly with another laugh, and then other people on the call during my briefing with SGI laughed as well at the suggestion. I suspect that this will put the Opteron rumors to rest. But, the interesting new blade architecture of the Altix 4000s certainly looks as if it would allow different processor architectures to be plugged into the NUMAlink structure, if SGI did decide, someday and for whatever reason, to support alternative processors. Once you see the modular architecture, you can see that one CPU module in a blade setup is replaceable with another, and you can then understand the confusion this might have caused.
The current top-end Altix machines, the Altix 3700s, can lash together up to 512 processors and 128 GB of memory into a shared memory supercomputer running either modified Linux or off-the-shelf Linux. The modifications are tweaks to the math libraries and other kernel features that can boost the performance of Linux running on the NUMAlink clustering. While these Altix 3700 machines are made out of individual motherboards that can be thought of as akin to blades, they are very much architected like systems. The Altix 4000s, says Matzke, have been re-architected from the ground up to be a much more modular and scalable system, allowing customers to add features they need--processor capacity, memory, graphics processing, I/O, and other features--independently of each other within each chassis of the configuration. This allows customers to better match the configuration of the Altix 4000 to their workloads, not try to make a generic machine fit their particular workload.
The first machine to come out of the Altix 4000 line is the Altix 4700 (presumably there will be variants of the Altix 4700 announced next year, much as the smaller Altix 330 and 350 machines were offshoots of the Altix 3700s). The Altix 4700 chassis has 10 blade slots, and up to four chasses can be fit into an industry standard rack. Matzke says that this chassis provides three times the density of computing resources as the current Altix 3700s. The Tollhouse blades have two processor sockets, and they support the "Madison" single-core Itaniums today and will support the future dual-core "Montecito" Itaniums, expected around the middle of 2006 now after several delays, as well as the kicker to Montecito, the dual-core "Montvale" Itaniums. The exact configurations on the Tollhouse chassis remains something of a mystery, but it looks like customers will be able to plug a lot up to 10 two-socket processor blades in a chassis. That's 80 processor cores per rack today and 160 processor cores when Montecito starts shipping, representing about 450 gigaflops per rack today (with the 1.66 GHz Itanium) and maybe just under 1 teraflops per rack with Montecito if the larger caches help performance.
Right now, SGI is expecting to deliver 1.6 GHz/6 MB cache and 1.66 GHz/9 MB cache Itanium options in the initial Tollhouse servers, which will begin shipping at the end of the first quarter of 2006. Up to 512 processors can be used to support a single global memory, which has been expanded from 24 TB with the Altix 3700s to 128 TB in the Altix 4700s. The Tollhouse machines will also have graphics blades and I/O blades (for linking the cluster to peripherals through hot swap PCI-X and PCI Express links), and will have special field programmable gate array (FPGA) blades, which are used to significantly speed the number-crunching performance of specific algorithms. The Cray XD1 has FPGAs as well, and these are becoming all the rage in HPC now that people are beginning to understand how to program them. The reason FPGAs are important, says Matzke, is that they can boost the performance of HPC workloads by as much as three to four orders of magnitude. These FPGA blades are called RASC blades by SGI, short for Reconfigurable Application Specific Computing.
Both the original and the future Tollhouse versions of the Altix servers are based on the NUMAlink 4 interconnect, which SGI says is almost twice as fast as the RapidArray interconnect used in the Cray XD1 Linux-Opteron parallel supercomputer, and offers 60 percent more bandwidth per link. NUMAlink is also faster than InfiniBand, Quadrics, Myrinet, and IBM's High Performance Switch interconnect as well. It is this speed as well as the NUMA technology that allows the Altix machines to create a single memory space for HPC applications to play in. Other federated architectures--massively parallel processors, commodity clusters, and constellations--typically have memory allocated to nodes in the supercomputer, not a single memory space. What the NUMA architecture means is that a single copy of the operating system can run across the nodes in the cluster and applications run in a single memory space; this simplifies system administration and programming.
Using single-core Madison processors in the Altix 4700, Matzke says that SGI can cram just under 60 gigaflops per square foot of floor space, which is about three times the density of the current Alrix 3700s as well as the Itanium-based Integrity Superdomes from Hewlett-Packard. And while IBM's p5 575, which has eight processor slots and can have either eight or 16 Power5 or Power5+ cores activated, offers about 90 gigaflops per square foot of floor space (presumably using the single-core Power5 chips), it does so by burning twice the electricity and dissipating twice the heat. Moving to the Power5+ chips will allow IBM to significantly increase the power density of the p5 machines, but similarly, SGI is counting on Intel's dual-core Montecito chips to help it give IBM a run for the money. Just moving to the Montecitos will allow SGI to deliver about 120 gigaflops per square foot of floor space. It would not be surprising to see SGI boost the processor count to 1,024 (or higher) for that shared memory, too. It is unclear if that can be done with NUMAlink 4, or if customers will have to wait until NUMAlink 5 for this.
As far as pricing goes on the Tollhouse machines, mum's the word. SGI is not saying until the machines start shipping in March 2006 or so. But Matzke did offer a few hints. "Historically, when we announce a new generation, it tends to be cheaper, and this product will be less costly than our current product." How much so remains to be seen. That depends more on market pressures than anything else. With SGI under the gun to boost revenues and get back into the black with some profits, it will have to figure some way to balance these two opposing forces. A week ago, SGI was delisted from the NYSE exchange and is now trading on the Over The Counter market. This is not a lot of fun, obviously, since having your stock drop to below a buck a share and your company's market capitalization fall to around $110 million (which is a about one-sixth of SGI's annual revenues) is not a good thing.
But SGI is pretty excited about the Altix 4700, and Matzke says that the company has secured $70 million in pre-orders for the Tollhouse servers prior to their launch today. These pre-orders include a 1,500-core, 6 TB machine at the Dresden University of Technology, rated at more than 12 teraflops, and a cluster with 6,656 cores (using dual-core Montecitos) that is rated at 69 teraflops that will be installed by the Leibniz Computing Center.
Helping SGI's cause is the fact that the Altix 4000s will run either Red Hat Enterprise Linux 4 or Novell SUSE Linux Enterprise Server 9 right out of the box. Customers who need to tune their machines for the best performance can deploy the ProPack extensions that SGI created at first for its own Red Hat-derived Linux (which debuted in early 2004) and then grafted into SUSE Linux; these ProPack extensions are not yet available on RHEL 4, but I am guessing that they will probably make it into RHEL 5 when in comes out at the end of 2006.