HP Sets Up Cloud-HPC Computing Unit, Launches Two Server Blade
Published: May 29, 2008
by Timothy Prickett Morgan
The companies coping with Web 2.0 workloads and trying to create massively scalable computing infrastructure on which to run those applications--think Google, Yahoo, Amazon, and eBay and you have the poster children--have a lot in common with the supercomputing labs of the world, which are trying to build big clusters to run models and simulations for good and sometimes nefarious purposes.
Given the commonality of the scalability, power consumption, and cooling issues between Web 2.0 shops and supercomputing centers, Hewlett-Packard has decided to merge its existing High Performance Computing organization and a stealthily operating Web 2.0 organization that has shipped thousands and thousands of custom-built servers already, to create a single Scalable Computing and Infrastructure Organization.
SCIO, which is akin to OSLO (HP's cross-divisional Open Source and Linux Organization, which has spearheaded its drive into open source products and support for them), will be headed up by Christine Martino, who ran the HPC division at the company. And she concedes that the overlap between the requirements of the two customer sets she is chasing don't overlap perfectly. "Massive scale is a common requirement," says Martino, "but Web 2.0 customers are driving a different set of metrics, such as performance per watt or dollars per megawatt of electricity." Of course, I quipped, when you are spending government tax money to build a supercomputer center to redesign nuclear bombs, you don't have to care all that much how much money you spend--not the way public companies like Google, Yahoo, Amazon, and eBay have to worry.
Martino won't say how much funding SCIO has behind it or even how many people are dedicated to the new organization, except to say HP has devoted a three-digit number of people to the unit who work in research, development, marketing, and sales for products particularly geared to these high-end customers and their unique IT needs. The new unit does not include the employees of EYP Mission Critical Facilities, a consulting company with 350 employees and expertise in data center design that HP acquired last November and merged into its HP Services unit. The experts at EYP are, however, providing input into the offerings SCIO will create, and visa versa. Just like OSLO drove services offerings and server bundling deals with Linux and middleware suppliers, SCIO will drive particular product development for Web 2.0 and HPC customers and will also focus HP's server, storage, software, and services offerings and tune them for these customers.
The first product to come out of this newly constituted unit is actually one that was already announced, the StorageWorks 9100 Extreme Data Storage System, or ExDS9100 for short. This network-attached storage array, which was launched in early May and will be available before the end of the year, is based on BladeSystem blade servers with clustered disk arrays, with from three to 10 "storage blocks," which are clusters of disks with 82 TB of capacity each. The file system driving these storage blocks are based on the PolyServe clustered file system that HP got through its acquisition of the company by that name in February 2007. HP says that each blade in a chassis can deliver 200 MB/sec of sustained performance on such workloads, and a chassis of 16 blades with 12.8 cores dedicated to streaming workloads (with the other cores in the chassis running the file system) can deliver 3.2 GB/sec of sustained performance.
The other neat new product that is coming out of SCIO is a blade server that packs two whole servers--each a two-socket Xeon server, in fact--on a single, half-height blade server form factor. The BL2x220c G5 blade server used Intel's dual-core Xeon 5200 and quad-core Xeon 5400 processors, either regular 80-watt parts or low-voltage 50-watt models, and supports up to 16 GB of main memory and two Gigabit Ethernet ports per physical server on the board. The resulting blade, when packed into standard racks, offers three times the density of a two-socket 1U form factor rack server setup. With the two-server blade, HP can put 1,024 CPU cores and 2 TB of main memory in a single rack, which works out to 12.3 teraflops of number-crunching power using the fastest quad-core Xeons that Intel can make. The blade has an optional mezzanine card that provides a PCI-Express x8 slot for 4x InfiniBand links.
A base configuration of the BL2x220c G5 blade with two 2.5 GHz Xeon L5420 quad-core processors per server node (that's four chips per blade and 16 cores) with 4 GB of memory per node with no operating system or storage costs $6,349. Using the faster and hotter E5450 Xeons, which have four cores each running at 3 GHz, boosts the price of the blade to $9,469. A BL260c G5 blade, which has two-sockets and only a single server node, but which can support up to 48 GB of main memory, costs $2,941 with two of the Xeon L5420 processors and 4 GB of memory. So HP is charging a pretty hefty premium for that density. But given the supply-demand issues, the company probably has little choice.
HP isn't saying if it will deliver an Opteron or an Itanium variant of the two-server blade, but it might if customers ask for it and they promise to order enough of them to make it worth the engineering while.
Post this story to del.icio.us
Post this story to Digg
Post this story to Slashdot