Fujitsu goes dense with Nehalem blades
It's a dynamic cube
Server maker Fujitsu this morning took the wraps off a new generation of half-height blade servers and a new blade chassis. Collectively, they're known as the Dynamic Cube, but they'll be sold by the more boring (yet more easily trademarked) name: the Primergy BX900 S1.
Like the latest blade server chassis designs from Hewlett-Packard and Dell, the BX900 S1 is a double-decker chassis that stacks two lines of half-height blade servers into the box, along with switches, power supplies, fans, and other infrastructure. Density is always an issue among blade server makers. That was one of the key selling points when blades came out at the turn of the century. But compute density cuts both ways because density also means you need to supply 20 kilowatts or more of juice to a single rack and then remove most of that juice as heat.
Among the tier one server makers, Fujitsu's new Dynamic Cube blade chassis has bragging rights in terms of density, with up to 18 half-height blades in a 10U chassis. The chassis will eventually support full-height blades, but they were not announced today. Hewlett-Packard's c7000 blade enclosure is also 10U high and supports 16 half-height (two-socket designs) or 8 full height (usually for quad-socket designs) blades. Dell's latest blade enclosure, the M1000e, also packs 16 half-height blades into a 10U space, and IBM's BladeCenter H chassis packs 14 full-height blades into a 9U chassis.
Using the latest quad-core "Nehalem EP" Xeon 5500 processors, launched in late March, Fujitsu can get 14.4 cores per 1U of rack space with the Primergy BX900, compared to 12.98 cores per 1U for HP and Dell and 11.2 per 1U for IBM. (Big Blue has a 7U BladeCenter E low-energy chassis that sports 14 blades, but they have to be relatively low-powered chips to get up to the 16 cores per 1U compute density that IBM is offering with this variant).
According to Joe Duran, product manager for the Primergy server business at the United Kingdom unit Fujitsu Technology Solutions, the design goal with the Dynamic Cube blade servers was not just density but increasing airflow to cool the blades, while at the same time reducing the amount of energy that the cooling systems use and the blades consume as they do their work.
FTS comprises the Fujitsu and Siemens server units that Fujitsu took whole ownership of in Europe on the same day Intel launched the Nehalem EPs. While Fujitsu added Nehalem EPs to its six-year-old Primergy BX600 blade chassis - which packs 10 half-height or five full-height blades in a 7U chassis - the new BX900 S1 chassis is really a whole new box, tuned to take advantage of the energy efficiency and virtualization features inside the Nehalem EP chips and their associated chipsets.
The BX900 S1 chassis has a new high-speed midplane that has four redundant fabrics for linking the blades to the redundant management processors in the chassis. This chassis has 6.4 Gb/sec of aggregate I/O bandwidth and six power supply/fan modules in the back as well as eight I/O connection blades. (Each power supply has two fans). One key difference that Fujitsu is highlighting with the BX900 S1 is that it is using power supplies with 90 per cent or higher efficiency and that the machine is actually designed to work under full load.
"Customers are hitting 75 to 80 per cent utilization these days," says Duran, thanks in large part to the virtualization and consolidation of server workloads. "So you have to make sure these machines can actually handle such loads in practice, not just in theory."
Duran says that HP's 16-blade c-Class chassis will draw about 5.5 kilowatts when loaded up with Nehalem EP chips, but that an 18-blade configuration from Fujitsu using the same processor and memory on the blades will draw just under 5kw of juice. That's about a 10 per cent reduction in power consumed for 12.5 per cent more computing delivered. The more efficient power supply is key, but so is opening up the fronts of the blades to let air in.
Duran says that 39 per cent of the surface area of the front of the BX900 S1 chassis is open for sucking in air, compared to as low as 18 per cent for some other blade server designs. More air intake means that Fujitsu can use larger fans and spin them more slowly, which radically drops the amount of energy it takes to cool the box and cuts back on noise.
The HP c7000 chassis moves about 300 cubic feet of air per minute through the chassis, according to Duran, and it is whipping through the box at 87 miles per hour. The BX900 S1 chassis can move 500 cubic feet of air per minute through the chassis, and to keep those 18 blades and their switches cool only requires the air to be moving at 53 miles per hour.
Fujitsu's homegrown ServerView system management tools can lash up to four BX900 S1 chassis together into a single management domain, putting a total of 72 blades - or 576 Nehalem processor cores - under the control of a single toolset. And Fujitsu's virtualized I/O for blades - called Virtual I/O Manager, or VIOM for short and akin to HP's virtual connect for server and storage networks - also can span across four chassis.
Eventually, according to Duran, the high availability clustering features of ServerView, called Resource Coordinator, will be tweaked to allow orchestration of workloads and processor, memory, and I/O capacity across the four chassis in a single rack. This capability is expected in the first quarter of 2010.
Along with the new BX900 S1 chassis that Fujitsu debuted today, there are two new blade servers: the BX920 S1 compute blade and the SX940 S1 storage blade.
Sponsored: Benefits from the lessons learned in HPC