Today I’d like to share some of what we’ve been working on in our cloud-optimized server offering. Admittedly, I’ve been remiss on blogging about some of the interesting developments going on in the cloud technology space, but if you count yourself in the server architecture geek set: carry on, dear reader.
Early last year, we developed a revolutionary design for the folks building hyper-scaled-out platforms. We’ve evolved that design and recently began shipping a new iteration of this system – the XS-23 II. One of the first customers to receive this new system is Lawrence Livermore National Labs who is utilizing XS-23 II for their Hyperion project. Ashlee Vance of the New York Times got one of the first glimpses of the system during a visit this week.
We've made several changes and improvements to the XS-23 chassis, among them some serviceability features. Our first iteration required some screw-turning for system board swap outs. While the failure/service domain of individual servers isn’t a primary concern for the true cloud applications (from which our initial design was created) – we know that serviceability is still important from a labor-savings standpoint. With that in mind, we’ve added slide-out trays for the individual servers to make removal/insertion an easier task.
On the performance side we’ve added to the unique configuration advantages this chassis offers. In its fully populated incarnation we house four 2S servers with three 3.5-inch HDD each. Over the past year we’ve offered customers the flexibility of removing a server (or two) from the chassis to bring them to a four or six 3.5-inch spindle-per-server ratio if that is what they needed. This reduces density to 1 or 1.5 servers per rack unit but carries a significant local storage subsystem advantage over other designs. Taking this flexibility a step further, in the new version of XS-23 we’ve added support for 2.5-inch HDD. Now we can make even more granular choices on storage performance/capacity/cost tradeoffs. Where spindle count is paramount we can go to six HDD per server at 2-servers-per-U density. Not bad. Compare this to general purpose server blades where we see a max of 2 x 2.5-inch spindles at the nearest density equivalent, which for a mainstream 2S system is 1.6 servers-per-U (16 in a 10U chassis).
Another design idea, “twins,” comes closer with 4 x 2.5-inch HDD per server at that 2-per-U density point. We’ve offered a twins setup for about a year but most of our customers have opted for XS-23 instead.
Back to the XS-23 II, this system will support Intel’s Nehalem family of processors and we’ll also continue to offer other flavors (San Clemente, Seaburg) for specific applications. One of the other interesting things we’ve done recently in the XS-series was a two-server design where we replace two of the four system boards in the chassis with two monster graphics cards which are employed as co-processors. More on that later.
XS-23 is starting to look like a product family unto itself and if imitation is the highest form of flattery I’d say we’re on to something.