Sunnyvale (CA) – A new Linux box is scheduled to arrive this August at the NASA Advanced Supercomputing (NAS) facility at the Ames Research Center in Moffett Field, CA. But, this is no ordinary Linux box. What makes this new machine distinct is that it will be the largest supercomputer ever created which runs a single instance of an operating system. In this case : Linux.
Supercomputers can be broken down into two general categories : big iron and cluster. Big iron refers to the high-end machines where all (or most) processors are controlled by a single instance of the operating system. These machines use a single system image (SSI) and have certain advantages which, even today, make the more desirable over the alternative for specialized tasks. Cluster machines are the other kind, and they typically run hundreds or thousands of OSes simultaneously. There is often one OS instance for each mainboard. NASA’s new Linux machine will be of the big iron variety.
It’s an SGI Altix 4700 comprised of 1,024 dual-core Itanium 2 processors (2,048 cores total). The entire machine, including all of its memory, networking and storage, will be controlled by a single operating system. As such, every running program will have full access to the machine’s total memory which will be isntalled at 4TB (terabytes, 2GB per core, 4GB per processor). The machine will be capable of ustained 13.1 Teraflops. It should debut in the Top 100 list of Supercomputers when the next list comes out in November, somewhere around #70. NASA will also install a 250 TB SGI InfiniteStorage 10000 system to handle this machine’s massive data storage requirements.
The largest supercomputers of today contain a physical core count in excess of 50x this machine’s. The difference here is that while they run thousands of copies of their native OS (typically UNIX or Linux), this machine need only run one. Supercomputers typically require a couple hours to boot up. At the rate of 1 GB per second it would take this machine 67 minutes just to test it’s system memory (were it to be done in serial). At a rate of 100 MB per second it would take this machine 11 hours to populate all of its memory from disk. With a power consumption rate of around 500 wats per mainboard (processor, memory, bus), this machine would consume a constant input of about 1.5 Megawatts (with cooling requirements factored in).