Fujitsu Extends Blade Server to Eight-Way SMP

Published: April 5, 2006

by Timothy Prickett Morgan

The Fujitsu-Siemens partnership dabbed its toe into the market for Opteron-based servers last fall with a two-socket rack server and a blade server that made use of the HyperTransport interconnect to lash blades together into symmetric multiprocessing (SMP) clusters. And now, Fujitsu-Siemens is pushing the extent of that SMP support to the current limit of HyperTransport technology.

The Primergy BX630 Opteron-based blade server is based on AMD's "Thor" chipset, which allows two, three, or four two-socket motherboards to be coupled together to create a single SMP machine, which looks like one single motherboard to an operating system because SMP servers have tightly coupled, shared main memory. With the Primergy BX630 server that shipped last October, Fujitsu-Siemens supported a two-socket blade based on single- or dual-core Opteron 200 series processors and a four-socket blade based on Opteron 800 series chips. The interesting bit in the Primergy BX630 machine is that customers could buy a special cable from Fujitsu-Siemens that allows the HyperTransport links to be connected, turning two physically separated two-socket SMP servers (possibly each with four active Opteron cores) into a single, logical four-socket SMP server (possibly with as many as eight active cores). What this means is that if you need a bigger blade, you don't have to swap out your two-socket blades for four-socket blades; you can make one on the fly. As we pointed out last fall, the Thor chipset could lash up to four blades together to create a logical eight-socket SMP, but Fujitsu-Siemens started out with only two boards. Now, Fujitsu-Siemens is enabling HyperTransport to go all the way to four boards and eight sockets.

The Primergy BX630 chassis takes up 7Us of space in a rack and has 10 blade slots, which means customers using dual-core Opteron 200 series processors can cram two 16-core SMP servers into a single chassis--and still have two blade slots left over.

According to Richard McCormack, senior vice president of product and solutions marketing at Fujitsu Computer Systems, the North American unit of the Fujitsu-Siemens partnership, this is the first time a tier one server vendor has used the NUMA-like SMP clustering in the Opteron architecture to scale out a box. He also says that while Fujitsu-Siemens has PrimePower Sparc64, PrimeQuest Itanium, and Primergy Xeon servers that all offer at least 16 cores in a single system image, the growing popularity of the blade form factor among some customers is driving Fujitsu-Siemens to add 16-core support into the Opteron blades. And because Intel does not have a direct memory connection architecture for its Xeon machines, providing such functionality using Xeon-based blades would require the development of an entirely new chipset. (IBM's "Summit" chipsets for high-end servers could provide such functionality, and it is peculiar and significant that Big Blue has never put the Summit chipset into its own blades to offer such functionality.)

While Fujitsu-Siemens can claim to be the first tier one server vendor to use the Thor chipset as AMD intended, it cannot claim to be the actual first vendor. Appro, a specialist in the high performance computing (HPC) portion of the server market based in Milpitas, California, has been shipping a similar kind of Opteron blade server since early 2005. Its XtremeBlade design also allows two-way, four-way, and eight-way blades (which take up one, two, or four slots in the chassis) to be mixed and matched inside a single chassis. Many early blade server designs required different chasses for blades of differing SMP scalability, but now vendors have settled down to the concept that they need to have one chassis and support it for a long time. The XtremeBlades will use the Opteron 200 series chips for two-ways and the Opteron 800 series chips for four-ways and eight-ways. Customers can use regular Opterons, or, if they are worried about heat, they can use the Opteron HE variants of the chips, which drop the voltage on the chips by about 25 percent and cut back on power use and heat dissipation by as much as 65 percent.

Because Fujitsu-Siemens has aspirations in both enterprise computing and in the HPC space, the extended Primergy BX630 machines could be an important differentiator for the company. First, customers who want to deploy enterprise applications--particularly messaging servers like Microsoft's Exchange Server--and who like the blade form factor have been running out of gas on four-socket blades. Moreover, since Fujitsu-Siemens will be supporting up to 128 GB of main memory on the four-blade, eight-socket SMP using 4 GB DIMMS, this could be a perfect HPC node for applications that require large main memories on relatively powerful nodes. As supercomputing centers play around with X64 clusters, they are figuring out that bigger nodes with bigger memory are often a better deal when it comes to performance than a lot of smaller, cheaper nodes lashed together with Gigabit Ethernet, InfiniBand, or other networking schemes.

The extended SMP capability for the Primergy BX630 will be available early in the second quarter. A two-socket blade is expected to cost $2,250, while an eight-socket setup will cost under $36,000.

RELATED STORY

Fujitsu-Siemens Finally Opts for Opteron in Servers