SGI Proves You Can Go Home Again.

New SGI 3000 series reaffirms SGI's strengths

Since its founding in 1982 as Silicon Graphics, SGI has been best known for its technical workstations, servers, and supercomputers and its emphasis--epitomized by its name--on visualization, high-end graphics, and other compute-intensive applications. This is not to say that the company doesn't offer prod-ucts in other areas; indeed, through much of the '90s, some observers might say that the company almost lost sight of its strengths, allowing itself to be drawn into the general-purpose Unix server wars--where it has been thoroughly trounced by Sun and HP and then into the world of Windows servers--the very definition of a me-too, commodity market and, now, SGI is an enthusiastic player in the open-source Unix world, as well, and has made valuable contributions to Linux, includeing the company's XFS journaling file system, which addressed a serious weakness in Linux.

Yet while the company may have wandered in the wilderness for a time, it appears never to have lost sight of its forte, as evidenced by the introduction last month of the SGI 3000 series of high-performance servers and visualization systems. These distributed shared memory systems are based on the company's NUMA 3 architecture, a third-generation non-uniform memory access technology. Rather than giving all processors in a system direct access to all of main memory via a bus with its slow speed and high latency, a NUMA system gives each processor its own local memory for high-speed direct access and, then, uses low-latency crossbar switches to deliver access to the rest of the memory. The result--given the right OS hacks--is a cache-coherent multiprocessor system with a single memory image accessed by a single operating system kernel.

NUMA, of course, is not new, although SGI claims their version to be the most advanced (thus, "third-generation"). What is new in the SGI Origin 3000 servers and SGI Onyx 3000 visualization systems is the NUMAflex architecture, which makes the 3000 series completely modular. A NUMAflex system is assembled as a series of "bricks"--computational modules that encapsulate each of the various functions of a server. The various bricks, all of which are hot-pluggable, include:

* C-brick: a CPU module (two or four MIPS processors) and local memory (up to 8GB). C-bricks are based on the Intel Itanium (64-bit) and, eventually, the forthcoming McKinley (128-bit) will also be available. Currently, the 3000 series scales to 512 processors; SGI promises scalability to 2,048 processors with the McKinley bricks.

* R-brick: the router interconnect module, which provides eight NUMAlink channels. Four of these are for connecting C-bricks to the interconnect fabric: thus, each R-brick can connect up to 16 processors. The other four ports are for connecting to other R-bricks.

* I-brick: a base I/O module containing an 18GB system disk with a second drive bay for OS mirroring, a CD-ROM, five PCI expansion slots, a 10/100 Ethernet port, two USB ports, one 1394 port, and two SGI Xtown2 ports (1.2GB/sec) to connect to C-bricks.

* P-brick: 12 additional hot-plug PCI slots with six 64-bit, 66MHz PCI buses.

* X-brick: supports SGI XIO high-speed expansion slots for proprietary SGI expansion cards.

* G-brick: houses SGI advanced hardware graphics subsystems.

* D-brick: up to 12 dual-ported hot-pluggable Fibre Channel disk drives.

The result of this modularity is a system that can be configured and reconfigured to match almost any imaginable computational demand. Of course, all this would be overkill for a file server. What SGI is aiming the 3000 series at is "big" data--things like econometric modeling, weather analysis, and the like. Yet the company points out that the 3000 series can actually start out in cluster mode (running Linux, if you like) and transition to distributed, shured-memory computing, partitioned any way you please. This should be especially attractive to academia; it's unlikely that anyone else would want to transition from, say, frame rendering (low-to-moderate grained parallelism and moderate I/O, for which a cluster is ideal) to weather modeling (finegrained parallelism, high computational and communication loads, for which distributed computing is ideal).

However, although the technical computing community is a small one compared to general purpose computing, the SGI 3000 series may be worth a look by VARs and systems integrators outside it. The insatiable demand for compute power in almost every field is redefining the meaning of technical computing or at least widening its applicability and, in the future, SGI may again find itself in the mainstream of business computing without even trying. Integrators who understand and can exploit the flexibility of a system such as the SGI 3000 will, then, he well positioned for profit.