HP in NonStop rack server chase
Blading (not quite) everything
Hewlett-Packard today is rolling out an entry configuration of its NonStop fault tolerant machines based on rack-mounted servers instead of the more expandable and more expensive blade-style NonStop iron the company has been peddling recently.
The new rack NonStop machine, the NS2000, is a bit of a return to the past for HP, which has delivered two rack-style, Itanium-based NonStops so far.
HP ported the NonStop platform to the Itanium processor in June 2005, launching the NS16000 based on single-core "Madison" Itaniums running at 1.5 GHz. The NonStop environment, which includes its own operating system kernel and relational database, runs on a fault tolerant cluster and is intended for online transaction processing. NonStop was created by some ex-HPers who founded Tandem Computers in 1974 to chase the IBM mainframe business. (Compaq ate Tandem in 1997 and HP ate Compaq in 2001).
Anyway, the first Itanium-based NonStop machines, the NS16000s, were nodes in a NonStop cluster, nodes that were based on multiple servers using HP's own zx1 chipset. The NS16000 scaled from 2 to 16 Itanium processors and from 4 GB to 32 GB of main memory, with each node having from 10 to 60 ServerNet I/O connections. (ServerNet is the secret fault tolerant sauce in the NonStop design). These nodes can in turn be clustered to create a very large fault tolerant infrastructure with thousands of processors. A typical NS16000 node had between 4 and 8 Itanium processors and sold for around $1 million, according to HP at the time.
The following June, HP launched the NS1000, an entry machine based on the Integrity rx2620 server that used 1.3 GHz single-core "Madison" Itaniums and scaled up to four server nodes. (The rx2620 is a two-socket box, but the NonStop derivative only allowed one processor per server board).
Last summer, HP announced a blade implementation of the NonStop platform, as part of its "blade everything" strategy. HP also wanted to get the NonStop platform running on standard HP servers, and in this case, the NS50000c blade server was itself based on the Integrity BL860c blade server that HP put into the field in February 2008. The NonStop kernel was tweaked to take advantage of dual-core "Montvale" Itanium processors running at 1.66 GHz, but the NonStop blade only used one of the processor sockets in the dual-socket BL860c blade. (The point of the NonStop design is to have server redundancy, so cramming too many cores on one board defeats the purpose).
Anyway, the NB50000c node is one logical processor in the NonStop cluster, and supports from 8 GB to 48 GB of main memory. Up to 4,080 logical processors (or 8,160 cores) can be clustered together in the blade-style NonStop boxes. An entry-level NB50000c configuration with two blades, a c7000 blade chassis, I/O controllers, and SAS drives for the blades ran about $300,000, and a typical entry configuration cost around $700,000, according to HP.
According to Randy Meyer, director of product management for the NonStop line within HP's Business Critical Systems division, the NB50000c has a lot more scalability than some NonStop customers need, particularly customers with older NonStop setups based on MIPS processors from years ago or new customers who want fault tolerance in Europe or emerging markets where their OLTP workloads are fairly modest. The MIPS-based NonStop S Series machines were sunsetted at the end of 2008, so customers with these boxes have to start thinking about an Itanium future, whether they like it or not. And with the new NS2000, HP is offering them a different and less expensive option than buying an entry blade configuration.
Blading (not quite) everything
So, the strategy of blading everything does not mean - and never did mean - only selling blades. But it could have. HP could have put a two-blade or four-blade machine together if it wanted to. But apparently, that is not what customers want, or HP can't do it at a compelling price.
The NS2000 rack-based NonStop node announced today is comprised of two or four pizza box servers, each with a single-socket, dual-core Montvale Itanium processor running at 1.4 GHz using 12 MB of L3 cache. Each node in the cluster can have 8 GB or 16 GB of memory per processor, for a maximum of 64 GB. Compared to the entry S Series MIPS-based NonStop iron, customers can get anywhere from four to eight times the performance in one-third of the floor space.
A base NS2000 configuration with two logical processors (that's four cores) and 16 GB of memory (8 GB per socket) and three mirrored pairs of 73 GB, 15K RPM disks runs about $125,000, according to Meyer. Throwing the NonStop kernel, database, middleware, and development tools adds another $100,000 to $150,000 to the price tag, depending on options. Still, at $225,000, the NS2000 has a much lower entry point for the whole NonStop stack than the NB50000c, which in a four logical processor configuration cost something closer to $1m once the software was added in. (Oh, you wanted the software, too?)
Meyer says that HP has no plans to return to the rack form factor except for this entry configuration, although he did say that going forward, HP intended to offer an entry rack server based on future "Tukwila," "Poulson," and "Kittson" Itanium processors as well as more expandable blade configurations for larger production workloads. (The NonStop kernel will have to be tweaked so the four cores in the Tukwila chip look like one logical processor to NonStop applications, and the same work will have to be done for Poulson and Kittson, whatever their core counts will be).
Meyer also expects that companies enamored of NonStop iron will be picking up NS2000 rack variants to use as development and test environments, since they are relatively less expensive and provide an isolated machine on which developers can be let loose and do whatever it is they do all day.
Customers sitting there with NonStop S Series MIPS machines have a number of ways they can move their code over to the Itanium-based NonStops. The NonStop kernel at the J Series level (yes, that is confusing) that runs on Itanium iron has an interpreter that is resident in the OS that can run MIPS-based applications in emulation mode. The software also includes an object code translator, which takes MIPS object code and converts it to Itanium object code, which allows it to run in native mode (of a sort) on the Itaniums. And then, for the full performance bang, they can actually recompile their MIPS applications to run on Itanium iron using the compilers inside the NonStop stack.
HP does not break the NonStop line out separately from other lines of products in the BCS division, so Meyer cannot provide any indication of how much revenues the NonStop products drive at HP. All that Meyer could say was that the blade products launched last year were ahead of plan in terms of the percentage of sales HP expected these products to rake in relation to total NonStop sales. Customers are moving applications off mainframes and vintage gear made by fault tolerant rival, Stratus Technologies, he says, as well as upgrading older NonStop gear. "This has remained a steady, profitable business for HP," Meyer added. "And we have been gaining share in the high-end of the server market."
It would be interesting to know just how different the NonStop business is today than it was when Compaq bought it. In 1997, the Tandem line accounted for about $2bn in sales, and in 2000, before HP bought Compaq, the business was still kicking out around $2bn in sales, but profits were under pressure. It is hard to imagine NonStop sales being anywhere near $2bn here in 2009. But Meyer can't say anything about it and still keep his job at HP. ®