The company which built the world’s fastest computer has drawn on its experience in supercomputing to build a series of Linux-based high end computers with a host of innovative features.
NEC — maker of the Earth Simulator supercomputer, which currently holds the #1 spot on the list of the world’s 500 fastest computers — today is unveiling in the United States its high-performance Express5800/1000 line of servers, aimed at the high performance computing market.
The Japanese computer maker, whose U.S. headquarters are in Sacramento, Calif., has been shipping the systems in Asia and Europe for several months.
The high-end Express5800/1000 can hold up to 32 64-bit Itanium 2 processors. Those processors are divided into 4-CPU subsystems, each of which can operate as an independent computer system.
That allows users to mix and match the operating systems on one box, and even to change operating system on the fly, says Dave McAllister, NEC’s director of 64-bit systems. “You could bring down four of the CPUs, and reboot them under a different operating system, without taking down the whole system,” McAllister says.
The computer currently is available running Linux, but will also run Windows .Net Server 2003 when it becomes available from Microsoft.
The Express5800/1000’s mainframe-type architecture provides a great deal of resiliency, according to McAllister. The system can be configured so there’s no single point of failure, so if the pathway between a controller and an I/O subsystem fails, there’s another pathway to get to that same I/O subsystem.
In addition, if a component does fail, it can be replaced without taking down the whole system: CPU cards, PCI cards, fans and power supplies are all hot-swappable.
Pushing Linux Past Eight Processors
Perhaps the most intriguing aspect of NEC’s new system, however, is that it scales past eight processors without the usual drop-off in performance that has bedeviled SMP Linux systems.
In spite of the SMP improvements added to the 2.4 version of the Linux kernel, Linux does not scale well past eight processors. Adding CPUs beyond that generally provides diminishing returns: each additional CPU provides proportionally less processing power.
The Express5800/1000, however, scales linearly all the way to 32 CPUs, according to NEC. Each additional CPU beyond eight provides the same amount of processing power, with no drop off beyond eight.
Part of the credit for this goes to the system’s 4-CPU architecture, says McAllister. “The Express5800/1000 is broken into cells of four CPUs,” he says, “and Linux does very well on four CPU systems.”
The linear scaling is also a result of improvements NEC has made to Linux itself. NEC is both a member of the Open Source Development Lab (OSDL), and an active participant in the Linux on Large Systems foundry, formerly known as the Atlas Project.
Working with these groups, NEC has made improvements to the Linux kernel which allow the system to dynamically allocate additional memory — the Express5800/1000 can hold up to 256 GB of main memory — to applications which need it. The system can also perform load balancing, shifting applications onto less-busy CPUs to maximize performance. These changes will be available in the 2.5 version of the Linux kernel.
Pricing for an 8-way Express5800/1000 server begins at about $180,000; a high-end 32-way system will cost about $400,000, according to NEC.