SGI ends Itanium era with UV supers

After years of development and financial struggle, Silicon Graphics has at last begun shipping its "UltraViolet" Altix UV supercomputers.

With that, the Itanium era of SGI is essentially over at the company, even though it will never say that publicly, and the NUMAlink interconnect is now available on standard x64 processors, giving SGI a real chance to start making some money.

Bill Mannel, vice president of product marketing at SGI, says that the first Altix UV 1000 boxes actually went out the door last Friday, ahead of the Memorial Day holiday and concurrent with the International Super Computing conference in Hamburg, Germany. Another batch of machines left the SGI factory on Tuesday, and full volume shipping is expected by the middle of June.

The initial configurations of the Altix UV 1000 machines are limited to the four-rack, 2,048-core versions of the machine, which link 128 two-socket blade servers using Intel's high-end "Nehalem-EX" eight-core Xeon 7500 processors together in a 2D torus topology using the NUMAlink 5 interconnect created by SGI.

The full Altix UV 1000 machine, which first lashes together 256 blades into a fat tree configuration and then links these clusters together in an 8x8 2D torus, for a total of 16,384 cores, will be available by the end of the calendar year, delivering 74.3 teraflops in a global shared memory system running Novell's SUSE Linux Enterprise Server 11 and SGI's ProPack extensions. Red Hat's Enterprise Linux is now supported on the Altix boxes, too.

SGI can double up the performance again using the NUMAlink 5 interconnect with 32,768 cores, is working to push the UltraViolet design to break the petaflops barrier by supporting 262,144 cores and then out to exaflops many years from now.

In their initial configurations, the base Altix UV 1000 machines have twice as many cores as the top-end Itanium-based Altix 4700s, which used the NUMAlink 4 interconnect and which was limited to 512 dual-core Itanium 9000/9100 processors. The Altix 4700s topped out at 2TB of shared memory, but the 2,048-core Altix UV 1000 machine supports 16TB of memory.

Initial customers for the UltraViolet supers include the Pittsburgh Supercomputing Center, Oak Ridge National Laboratory, the Leibniz Supercomputing Centre, and the University of Cambridge. An unnamed group of systems integrators working for unspecified federal governments are also lining up for the machines, according to SGI.

Concurrent with the shipping of the Altix UV 1000 supers, SGI is rolling out its ProPack 7 extensions to SLES and RHEL Linux, which includes optimizations for Itanium and Xeon 7500 processors and the NUMAlink interconnects as well as math libraries and MPI stacks created by SGI specifically for its supers. ProPack 7 includes SGI's Unified Parallel C compiler, its first parallel C, and also has a tweaked version of SGI's Message Passing Toolkit that can scale across tens of thousands of processors.

The MPT implementation adheres to the MPI-2 standard and includes a new feature called Performance Boost, which allows the MPT stack to intercept calls from programs compiled to other MPI stacks (HP MPI and Intel MPI being the two important ones) and run them as MPT, which runs better on the box. But you don't have to recompile the HPC applications to use MPT to get the benefit.

The updated MPT stack inside of ProPack 7 also has hooks for the MPI Offload Engine (MOE) that is a key component of the Altix UV 1000 machines. The MOE electronics in the NUMAlink 5 hub chipset can process MPI reductions two to three times faster than competitive clusters and massively parallel systems, according to SGI, and can handle MPI barriers up to 80 times faster. By having the hub do some of the MPI work, more CPU cycles are left to do the work they are supposed to be doing, such as doing calculations.

SGI has also updated its SGI Management Center cluster management tool, which does bare metal provisioning of cluster nodes as well as performance monitoring and management, to work on the new Altix UV 1000 supers. The management tool is currently certified to work on Altix UV, Altix XE, Octane III, Rackable, and CloudRack systems, with support for Altix ICE blade clusters coming later this year. T

he ProPack 7 extensions run on all SGI machines except the Origin 400. ®