Large-scale data processing reached new heights when Silicon Graphics Inc. (SGI) on Thursday announced that its Altix 3000 server for clustering will scale up to 256 processors in one node.
This means users need only one instance of the Linux operating system and one Linux kernel to address all 256 processors, said Jason Pettit, product line manager at SGI in Mountain View, Calif.
The Altix 3000 server is based on Intel Corp.’s 1.5GHz Itanium 2 64-bit microprocessor. It also contains the NUMAflex shared memory architecture, allowing processors in one node to share memory, negating the need to pass data through I/O or networking bottlenecks, the company said. NUMAflex is contained in the SGI ProPack 2.4, consisting of a message passing toolkit (MPT) shared library and a science and math libraries (SCSL) shared library. It also contains asynchronous I/O, the Vitesse SATA driver and diskless booting of partitioned systems. ProPack 2.4 enables the Altix 3000 to store up to 8TB of data.
To enable an Altix 3000 cluster, Pettit said users need only download the 2.6 Linux kernel and install the ProPack 2.4.
“The things in ProPack don’t alter the kernel but they do enhance the performance of the system,” Pettit noted.
Users can also create superclusters by connecting two Altix 3000 servers, scaling to up to 512 processors. By May 2004, SGI expects to be able to connect four Altix 3000 servers, creating a supercluster with up to 1,024 processors, Pettit said.
“These larger [type] systems are utilized by academia, government national labs and supercomputing centres around the world,” Pettit said. “Even within commercial spaces we’re seeing adoption of these large-scale systems, so that it can speed their time to discovery or their time to producing new products and improving the accuracy of the models they’re working on today by being able to incorporate more data.”
When SGI released its Altix 3000 server in September 2003, it could only scale to 64 processors in one node, Pettit said.
That is when the company embarked on a beta program to increase the clustering ability of the Altix 3000 to 128 processors, he added. The project was so successful it culminated with the Naval Research Laboratory in Washington D.C. deploying a 128-processor Altix 3000 cluster on a single distribution of Linux. This achievement prompted SGI to continue its efforts, scaling now to 256 processors in one node.
SGI has now set a new goal of scaling one Altix 3000 server up to 512 processors by the end of 2004, Pettit said. This means a 1,024-processor supercluster could be created by joining two 512-processor nodes.
SGI’s NUMAlink architecture would enable data to travel between these supercluster nodes in as little as 50 nanoseconds, the company said. NUMAlink is also part of ProPack 2.4.
Right now, SGI is embroiled in an experiment with NASA’s Ames Research Center in Moffett Field, Calif., whereby NASA is running a single 512-processor cluster node. This allows NASA’s Ames to conduct research such as ocean simulations.
Before running an Altix 3000 server, NASA Ames ran 1,000 processors on a SGI Origin server, which runs on Irix, SGI’s flavour of Unix, said Bob Ciotti, TeraScale applications group lead at the NASA Ames Research Center.
In summer 2003, NASA Ames switched to the Linux-based Altix 3000 server. The research group started out with a 128-processor node, then moved to the 256-processor node and finally ramped up to the 512-processor node, Ciotti said.
“Our interest in doing this is [that] the Itanium 2 is a very fast processor — it does very well on scientific problems, things that have high demand for memory bandwidth,” he said. “[Additionally] since we’ve already built a 1,000-processor machine and it has these characteristics of shared memory, high-bandwidth interconnect and low latency interconnect, we know that we can get [the Altix 3000] to scale fairly well.”
Users can purchase the Altix 3000 system with a minimum of four processors at US$75,000. The server scales in four processor increments to 256 processors, reaching a price of US$4.1 million, Pettit said.
