SGI Targets Midrange HPC with Linux-Based Altix 350s
by Timothy Prickett Morgan
Supercomputer maker SGI may be working on getting a commercial 128-way Itanium-based server out the door for high performance computing (HPC) customers sometime this spring, but it has its sights on the broader market for so-called midrange HPC gear. The high-end of the HPC market gets all the glory, but the money is really made in the midrange. To that end, SGI has rolled out the Altix 350, a dense, midrange supercomputer that is a chip off the current 64-way Altix 3000 block.
SGI was showing off the new Altix 350s at the LinuxWorld show in New York this week. Almost exactly a year ago, SGI rolled out the 64-way Altix 3700 HPC server that employed Intel's 900 MHz/1.5 MB L3 cache or 1 GHz/3 MB L3 cache "McKinley" Itanium 2 processors. SGI also debuted a deskside variant called the Altix 3300 that scaled from four to twelve of the 900 MHz McKinleys. Last July, when the 1.5 MHz/6 MB L3 cache "Madison" Itanium 2 processors were launched by Intel, they could be immediately plugged into the Altix 3000 line, yielding about 50 percent more performance.
Both the Altix HPC servers, which are based on the 64-bit Itanium processors and run Linux, and the Origin HPC servers, which are based on SGI's own 64-bit MIPS RISC processors and its own Irix Unix variant, have SGI's NUMAFlex clustering technology at the very heart of their systems. With the NUMAFlex cache-coherent, non-uniform memory architecture, each processor runs its own instance of an operating system (in this case, Linux or Irix), but like processors in a symmetric multiprocessing (SMP) cluster (a clustering method that is common in generic servers and workstations today), all of the processors in a NUMAFlex cluster can access the same shared pool of memory. Any time processors and applications have a single main memory space to play with, it simplifies programming and allows bigger programs and more programs to run more efficiently on a machine than is possible in a loosely coupled machine.
The new Altix 350s bear some resemblance to the Origin 350 midrange boxes that SGI launched in April 2003. Both use the third generation NUMAFlex architecture, and are based on rack-mounted servers that are denser than their bigger brethren. Chip for chip, the performance of the Altix 350s is probably going to be better than the Origin 350s for certain workloads, given that the Origin 350s use 600 MHz or 700 MHz R16000 processors. However, the Origin 350s can scale to 32 processors in a single memory space, compared to 16 processors with the Altix 350. That's because the Origin 350s are based on four-way cell boards instead of the two-way cell boards used in the Altix 350s.
The Altix 350 two-way cell boards are not the same as the four-way boards used in the Altix 3300 and Altix 3700 machines, obviously, but they share plenty of technology. The Altix 350 supports any Itanium 2 processor or the new Low Voltage Itanium 2 chip, which was code-named "Deerfield." Deerfield support is significant, because the Deerfield processor throws off only 63 watts of heat running at 1 GHz with 1.5 MB of L3 cache and it only costs $744, compared to $4,227 for the fastest 1.5 GHz Madison with the 6 MB L3 cache. The machine will also support the low-cache, HPC variant of Madison that Intel announced last fall, which runs at 1.4 GHz, has 1.5 MB of L3 cache, and costs only $1,172. These low-voltage Itanium 2 processors can, says Andy Fenselau, Altix product line director at SGI, make an Altix 350 configuration anywhere from one-half to two-thirds the cost of an equivalently powerful Altix 3000. However, if customers want to use faster, hotter, and more expensive Madison chips in the Altix 350s, they can do this.
The Altix 350, which was code-named "Opus," comes in a 2U chassis. A base machine with a single 1.4 GHz/1.5 MB L3 cache LV Itanium 2 processor, 1 GB of main memory, and a NUMAFlex link costs $12,199. A four-processor configuration in two chassis linked by NUMAFlex into a single 4 GB memory space costs $21,600. The Altix 350 can be expanded to include up to eight of these 2U, two-way nodes and up to 192 GB of shared main memory and 32 PCI-X peripheral slots. Customers who invest in Altix 350 machines can cluster them with Gigabit Ethernet or InfiniBand interconnects if they need to add power. But the Altix 350s boards cannot be directly upgraded to the Altix 3700s, although processors, memory, NUMAFlex, and other components can be moved to the bigger frames if need be.
Fenselau says that over $2 billion of the $2.6 billion annual midrange HPC server market is comprised of clustered Unix SMP servers that can be replaced by the Altix 350 configurations. This is what SGI is gunning for--replacing Unix machines with these Linux boxes, even if it means replacing its own Origin boxes. He says that the Altix 350s have about 50 percent lower costs and about two to three times better price/performance than the Unix servers that IBM, Sun Microsystems, and Hewlett-Packard ship into the HPC midrange. "We hope these numbers will not only raise eyebrows," says Fenselau, "but will compel customer to consider Altix and Linux instead of Unix for HPC workloads."
Interestingly, Fenselau says that the typical HPC site administrator has about $25,000 a quarter in discretionary spending, and it is no coincidence that the cost of a four processor Altix 350 node is under that limit. In the course of a year's time, an HPC site could have a fairly powerful 16-processor Altix 350 installed and not have to go through budgetary nightmares. He also says that for certain workloads, having 16-processor Altix 350 nodes clustered better suits HPC applications than buying the same number of two-way Xeon DP servers running Linux, as IBM, Dell, and HP might have customers do. Moreover, for a given number of Itanium or Xeon CPUs, it is clearly easier to manage as few nodes as possible.
The Altix 350 runs a variant of Red Hat's commercial Linux distribution that has been optimized for 128-way processing and NUMA clustering by SGI. It also includes a number of optimizations for math libraries that boost HPC performance that are proprietary to SGI. Novell's SuSE 8.0 and 8.1 Linux distributions are also supported off the shelf on the entire Altix line, but this SuSE support does not include the full set of optimizations that the Red Hat version has. It is reasonable to assume that now that SuSE has been acquired by Novell and is going after the HPC market with gusto, these optimizations will be done for SuSE's Linux within a few months time.
SGI will roll out the next generation of its NUMAFlex clustering in the second half of 2004, presumably for both Origin-MIPS-Irix and Altix-Itanium-Linux machines. Exactly what SGI has in store for the fourth iteration of NUMAFlex is unclear, but lower latency, higher bandwidth, and more memory scalability seem likely.