Today HP announced its new Extreme Scale-Out (ExSO) portfolio, a modular architecture designed to address Web 2.0-type horizontal scalability needs.  Of course every major vendor announcement touts ground-breaking cost savings (absolute lowest CapEx and OpEx!), energy savings (100% fewer watts consumed!), and density (more sockets, less space!). But we think this really is a very interesting product, and a particularly well-timed strategic move for HP.

Instead of putting a slightly new spin on blades or their existing rack-mount servers, HP took a clean-sheet approach. The result is an extremely lightweight rail and tray design that replaces the usual chassis + rack – a significant advantage over existing gear. HP also is offering servers in three different flavors to handle different needs. The first is the SL 160z, which sports two sockets, 18 memory slots, and room for two drives. The 170z trades memory slots (it has 16 rather than 18) for storage, upping the drive count to a max of 6 3.5” SATA or SAS spindles. The third server is the 2x170z, which is the most computationally dense configuration – two 170z servers on a single tray, each with 16 DIMM slots and a single drive.

Extreme x86 computing, marked by very lower power systems packed into very small spaces, is the low-cost answer for large-scale Web 2.0 (and ‘old’ Web 1.0) customers. More and more traditional corporate customers will find systems like these to be the low-cost solution for their scale-out needs as well. Another interesting market is HPC, where cycles per sq/ft and cycles per watt are increasingly important – and the HPC market is the highest growth computing segment right now. It’s obvious that HP did quite a bit of work devoted to really understanding the unique needs of this market. The impulse for most vendors is to put more and more stuff into their servers to, for example, increase hardware RAS. However, the usage model for these customers is rip and replace, meaning that their solution to a HW failure is to simply replace the entire server.

They get their RAS through having massive redundancy and capacity; the availability of any individual server doesn’t really matter when you have 100,000 of them and stateless applications that can be run on any of the other 99,000 boxes. So, for these folks, acquisition and operating costs are king, meaning that every square foot, watt consumed, and btu is measured and used to calculate TCO. When you have hundreds of thousands of servers, minute differences add up to real money. The same thing applies to the service model. As one vendor told me when they introduced their massive scale-out systems, “These guys handle service a little bit differently. They want to send a high school kid up and down the rows with a wheelbarrow and have him pull the servers that are lights-out and slide in a new one. They’d have two pallets on the loading dock, one with new servers from us and another with busted stuff to send back to us…it’s a simple as that.” It looks to us like HP is offering the same kind of service model with this new box – or it can certainly be negotiated, since customers who buy boxes in 10,000-unit orders have significant leverage, even when margins are very low.

Bottom line:  this is HP’s best foray into the massive scale-out system space, and a great first effort. Like IBM’s iDataplex, these systems are aimed squarely at the space that Rackable (and others) pioneered a few years ago. HP looks like it will make some waves in this market.