SGI Scales Linux Servers To 256 Processors

The Altix 3000 systems, running Itanium 2 processors, are initially aimed at high-performance computing in government and academic research.
Silicon Graphics, Inc., revealed Wednesday that it's solved bottlenecks in Linux, and is now able to scale its Altix systems up to 256 processors running off a single Linux kernel.

Key to the breakthrough, said SGI officials, are changes within its Advanced Linux Environment and the latest version of SGI ProPack 2.4 software, which was released in February.

"Before Altix, SGI had done a lot of work within and with the Linux community to make it possible to scale to 64 processors in a single instance of Linux, and we've continued to work on the performance bottlenecks," said Jason Pettit, the product line manager for the Altix 3000 systems.

A large system, such as a 256-processor Altix, said Pettit, "puts a magnifying glass over the Linux kernel and allows for what would be small performance hits on a single desktop or workstation to really come to light. That let us identify these smaller bottlenecks and to fix them.

"Once we identified how to scale beyond 32 processors, there was really no top end to Linux," he added.

In fact, SGI's original intent was to push from 64-processor scaling to 128 later this year, but as work continued, it became apparent that there were no problems speeding up the timetable, and scaling Altix to run 256 CPUs off one Linux kernel, said Pettit.

The Altix 3000 systems, which rely on Intel's Itanium 2 processors -- customers have a choice between Itanium 2's running at 1.3GHz or 1.5GHz -- are initially aimed at high-performance computing in government and academic research. NASA's Ames Research Center, for instance, already uses a demonstration version of the Altix that has been scaled to 512 processors for ocean and climate change research.

But SGI plans to also push the high-processor-count Altix 3000 into corporate environments, such as R&D departments of companies in the oil, manufacturing, and aerospace industries. SGI had previously planted beta versions of the Altix with, for instance, a French oil company, which is using a 256-processor system to model oil exploration and drilling scenarios.

By boosting the number of processors running off of a single Linux kernel -- as opposed to, for instance, tying together several cluster nodes, each running its own iteration of the operating system -- the Altix offers simpler administration, easier programming, and direct access to more memory, said SGI executives. Pettit pointed out that the Altix, when loaded with SGI's Linux and ProPack 2.4, can address as much as eight terabytes of memory.

"A benefit over clustering is the simplicity of administration," said Beverly Bernard, the company's program manager for Linux. "If you have 20 modes, each node needs to be identical in its configuration and its applications. Change one and you have to change them all. It's much simpler with Altix."

Petitt touted the scaling as a way for research facilities, and researchers with less-than-cutting-edge programming skills to still harness lots of computing horsepower. "The shared memory system of the Altix allows you to complete complex modeling without worrying how to divide it up, decompose it, to the separate nodes of a cluster." This is perfect, he said, for researchers who need high-performance computers, but aren't skilled enough programmers to split the computing tasks among separate systems.

Altix 3000s scaled up 256 processors are available immediately, added Pettit, who ball-parked the price of such a system at around $4.1 million. The company's shooting for delivering systems that run up to 512 processors from one Linux kernel by the end of the year.

Editor's Choice
Brian T. Horowitz, Contributing Reporter
Samuel Greengard, Contributing Reporter
Nathan Eddy, Freelance Writer
Brandon Taylor, Digital Editorial Program Manager
Jessica Davis, Senior Editor
Cynthia Harvey, Freelance Journalist, InformationWeek
Sara Peters, Editor-in-Chief, InformationWeek / Network Computing