Silicon Graphics (SGI)
Tuesday revealed servers and software that make it possible for the Linux
operating system to tackle massive computing chores.
The Mountain View, Calif. firm, known for its development of high-octane
supercomputers for life sciences and government markets, unveiled its Altix
3000 family of servers and superclusters. The systems combine SGI’s
supercomputing architecture with Intel Itanium 2 processors and the Linux
operating system to dash a number of scalability and performance records.
In fact, each node in the Altix 3000 series runs a single Linux operating
system image with up to 64 Itanium 2 processors and 512GB of memory. Because
several nodes use the SGI NUMAlink system interconnect fabric, data is
transmitted up to 200 times faster than with conventional clustering
methods, making it possible for the server to scale to hundreds and,
ultimately, thousands of processors. How fast is the NUMAlink interconnect
fabric? Data crosses over an SGI NUMAlink switch, round-trip, in as little
as 50 nanoseconds, which is about as much time as it takes a beam of light to travel 50 feet.
What makes the Altix 3000 line so special? Most supercomputers require
massive amounts of global shared memory to tackle complicated models, such
as global climate prediction. Such tasks cannot be easily solved in smaller
pieces. Altix 3000 superclusters work like supercomputers because they
provide global shared memory across nodes and operating systems. The
marriage of global shared memory and Linux creates great breakthrough
opportunities for technical users on a standards-based platform.
The Altix 3000 family drew rave reviews from a number of computing experts.
“A few months ago the idea of Linux nodes beyond eight processors was only a
dream,” said Jon “maddog” Hall, president and executive director, Linux
International. “By moving the performance of clustered memory to approach
that of main memory, SGI has unlocked the doors to address whole new
categories of problems that were either difficult or incredibly inefficient
with the previous generation of standards-based clusters. Yet SGI did it
with a standard set of programming interfaces.”
Illuminata analyst Gordon Haff said some of the benchmarks were impressive, but that Altix 3000 is geared for very specific functions and that the news shouldn’t be posed as a means of saying linux can scale to 64 processors 00 that just isn’t true.
“Fundamentally, Linux has been looked at as a distributed OS,” Haff said. “And that’s fair. But the way SGI has used it with Altix shows that it can be used as a replacement for Unix in certain capacities. SGI has shown Linux can scale quite effectively in certain environments.”
The Altix 3000 family will be available later in the first quarter of 2003 in
both entry-level and scalable supercluster models. The entry-level server
starts at $70,176 at four processors with up to 32GB of memory and scales to
12 processors and 96GB of memory. The supercluster model scales to hundreds
of processors and over 1TB of memory today, with future scalability to 2,048
processors and 16TB of global shared memory. The king of the hill
64-processor system begins at $1,129,262.
SGI, who competes with the likes of Cray, IBM and HP in the supercomputing
space, is taking a page out of Big Blue’s book in terms of attractive price
points: the firm said its Altix 3000 family offers roughly twice the
performance of comparable IBM machines for half the price.