HP Scales Down NonStop Servers to Chase New Customers
Published: June 8, 2006
by Timothy Prickett Morgan
Having made the jump from MIPS R Series processors to Itanium chips from Intel in its NonStop line of fault tolerant servers last summer after years of coding, Hewlett-Packard has quickly repackaged its entry rx Series servers and added some NonStop functionality to create a smaller NonStop server line with better bang for the buck in an effort to expand its customers base for these exotic Unix machines.
When the NonStop NS 16000 servers were announced in June, they were about a year behind schedule. The NonStop server platform consists of a special Unix variant and a homegrown clustered database management system that made Tandem famous decades ago as the only real highly available alternative to IBM mainframes. Tandem, which was founded by ex-HPers, was acquired by Compaq in 1996, which was in turn bought by HP in 2001, bringing the Tandem line back home and giving HP a fairly large and sophisticated installed base of customers in financial services, government, and healthcare who had chosen fault tolerant architectures over other kinds of clustered Unix or proprietary solutions. The problem with the Tandem machines, of course, is that they were based on the relatively feeble MIPS RISC processors, which used to be top-notch a decade ago, but which started falling behind other architectures around that time. Moreover, as Unix servers became wickedly popular in the early 1990s and then absolutely mainstream by the late 1990s, application vendors supported Solaris, HP-UX, and AIX, but viewed the much more costly--at least in terms of initial sticker price--Tandem NonStop as a niche platform.
This has limited its appeal in the market to customers who absolutely must have seven 9s of availability--that means the server is up and running 99.99999 percent of the time, which equates to a mere three seconds of downtime a year--for their applications. Think ATM networks, stock exchanges, mobile phone networks, large hospital networks and you'll get the right idea of who buys NonStop servers.
The NS 16000 servers that HP announced last June and began shipping last August cram 16 single-core, 1.6 GHz "Madison" Itanium processors into a single node, and the machine has enough ServerNet interconnection fabric between the nodes, which run the NonStop Unix kernel and the NonStop database, to scale to 4,096 processors in a single system image. (ServerNet is the grand daddy of the InfiniBand interconnect, by the way.) Two months ago, HP quietly announced the NS 14000, a trimmed down version of the machine that uses a slightly slower 1.4 GHz Madison processor and also trims back on the I/O capability to provide a slightly lower price point.
But, HP thinks that it can chase a lower price point by creating a system based more on stock Integrity servers and missing some of the ServerNet electronics and triple redundancy that push availability from five 9s to seven 9s a year. That's about five minutes of downtime a year for five 9s availability. That extra 177 seconds of uptime comes at a pretty high price, thanks in part to the sophisticated triple redundancy, "voter logic" that the ServerNet infrastructure in the NS 16000 and NS 14000 NonStops machines has.
That's what the NS 1000 is all about, explains Craig Wagner, director of solutions and marketing for the NonStop Enterprise Division at HP. The NS 1000 line is built from HP's existing rx2620 Integrity servers, which are two-socket machines that the company normally sells running HP-UX, Windows, Linux, and OpenVMS. The NS 1000 is built using 1.3 GHz Madison processors, and the NonStop's ServerNet interconnect scales only to four nodes, or a total of eight processor sockets. This limits the ultimate scalability of the NS 1000 to 2,048 total processors in a single system image, which, if you think about it, is not much of a limit, especially when you consider that the largest RISC/Unix boxes only scale to 64 or 128 processor cores in a single system image. The NS 1000 can have 8 GB of main memory per processor socket, and Wagner says that when the future dual-core "Montecito" chips are ready from Intel, these will plug right into the Itanium-based NonStop line; so will the future four-core "Tukwila" Itaniums, whenever Intel gets them out the door.
A base NS 1000 with two processors installed, including the NonStop operating system and the NonStop SQL database plus ServerNet links and outbound Gigabit Ethernet networking costs $170,000, says Wagner. That's a lot cheaper than the base NS 16000 node with two processors, which costs $400,000. While this might sound like a lot of money to charge for a two-socket server, when you are talking about high availability in the Unix or mainframe markets, it certainly does not come free, since more mainstream clustered databases require customers to double up on server capacity and do failovers from one machine to another--for instance, using HP-UX servers and HP's MC ServiceGuard clustering software and Veritas clustered file systems from Symantec. With the Tandem approach, you simply have an extra node in the database cluster for redundancy, you do all processing in parallel, and if one node fails, the cluster heals around the failure and keeps going.
Why this fault tolerant approach is not the dominant Unix architecture today just shows how stupid the IT business can be.
Still, if fault tolerant processing cannot be dominant in the market, it surely can be more appealing to existing and new customers. Wagner says that it has been peddling a high-end patient care system to hospitals in conjunction with General Electric for years, but the pricing on the prior NonStop machines meant that the solution was only affordable for hospitals with 500 beds or more. With the NS 1000s, Wagner says that it can make compelling economic arguments to hospitals with as few as 150 beds. What holds true for hospitals holds equally true for smaller telecom, financial services, and governments that need uptime, but do not need seven 9s of it. "There are places in the world where ATMs and point of sale networks are now coming to the market," explains Wagner, "but customers in these emerging markets are extremely price sensitive."
Perhaps most significantly, the NS 1000s give HP something to peddle against its old adversary: Big Blue. Wagner says that somewhere between 75 and 80 percent of the NonStop deals that HP does have NonStop gear being peddled against IBM mainframes. "Mainframes can make a high availability claim reasonably and rationally," says Wagner. "But with the NS 1000s, we can deliver fault tolerance in the same price range of regular clustered Unix servers." He believes that even with the new System z9 Business Class announcements from May, where IBM delivered a smaller z9 mainframe with substantially lower prices, HP can still out-compete IBM for dollars in data centers in established customers in North America and Europe and in fast-growing, emerging markets in Asia and Eastern Europe. Moreover, with about one out of five of its deals displacing a clustered Unix platform, HP is also hoping that it can take a bit out of the Unix businesses of Sun Microsystems and IBM.
HP has been pretty bullish about Itanium in recent months, and the NonStop business is probably the fastest-growing Itanium platform in the world right now; it certainly has that title within HP. In the span of nine months, Itanium-based NonStop machines now comprise 25 percent of NonStop sales. "This is the fastest adoption rate of any new processor that we have ever seen with the NonStops," says Wagner. With the Itanium-based machines yielding a factor of three price/performance improvement compared to the MIPS-based NonStop servers, moving to these new NonStop nodes is a no-brainer. But, then again, customers who have working clusters are loathe to touch them, too. And because customers can mix and match old and new nodes in a NonStop cluster, customers can take their time upgrading.
Which is another thing that is great about the NonStop architecture.
NonStop Fault Tolerant Servers Jump to Itanium