Inside Advanced Scale Challenges|Wednesday, October 24, 2018
  • Subscribe to EnterpriseTech Weekly Updates: Subscribe by email

HP to Port NonStop Clusters from Itanium to Xeon 

Hewlett-Packard is in the process of porting its NonStop fault tolerant database clusters to Intel's Xeon processors after an eight-year run on the Itanium chip. The company is also rolling the "Poulson" Itanium 9500 chips into its existing NonStop systems.

The Xeon port of the NonStop stack is significant because Intel has not formally sunsetted its Itanium server processors. Moreover, HP cannot answer direct questions about the Itanium roadmap other than the Intel party line because of its ongoing lawsuit with Oracle over the fate of the Itanium chips. The growing consensus is that the next iteration of the Itanium chip, due around 2014 or so, will be the last.

The venerable NonStop fault tolerant machines were the epitome of scale-out database clustering before the Internet giants created alternative technologies to suit their purposes. The current NonStop machines scale out to a maximum of 16,320 cores across a shared-nothing relational database; the nodes in the cluster are hooked together by the systems proprietary ServerNet interconnect. Each of the 4,080 nodes in that maxxed out system run their own copy of the NonStop kernel and database, and in the event one node fails, other nodes can seamlessly take over its work. The largest account currently has over 1,000 nodes linked together, according to Randy Meyer, director of product management for the NonStop line within HP's Business Critical Systems division. So no one is as yet pushing those scalability limits.

NonStop systems are designed for online transaction processing and hail from the same era that saw the rise and fall of mainframes and vector supercomputing. Tandem Computers, which created the NonStop systems originally, was founded by a bunch of ex-HPers in 1974 to provide a more reliable alternative to an IBM mainframe for online transaction processing workloads that could not have downtime. IBM clustered pairs of mainframes together to accomplish this, and Tandem used a cluster of servers to spread out the risk and lower the cost of implementing the system. Tandem had to create a parallel database and an interconnect that could farm out the work, and this was not a trivial engineering feat.

The initial targets for Tandem gear were stock exchanges, payment systems, ATM networks, reservation systems, and telecommunications subscriber and location systems. Over the years as first Compaq and then HP took control of the NonStop line, the fault tolerant clustering technology has kept most of its niches in these markets excepting the large stock exchanges in the Western economies. These exchanges have long-since moved to Linux clusters and high frequency trading, and they have designed a kind of fault tolerance into their applications, says Meyer, instead of implementing it in a massively parallel database. However, financial services firms in Asia in particular still like NonStop gear for their backends because they want greater fault tolerance and they are doing a lot of journaling, says Meyer.

HP does not disclose the size of the NonStop business, but it is large enough – and presumably profitable enough – that HP is willing to invest in the port from Itanium to Xeon systems.

Documents released during the lawsuits between HP and Oracle concerning the fate of Itanium showed that HP contemplated porting its HP-UX variant of Unix from Itanium to Xeon and Opteron processors at least twice over the past several years. But for reasons that HP has never explained, the company decided instead to put HP-UX in maintenance mode on Itanium gear and shift to selling Linux and Windows on high-end Xeon machines. The company's "Project Odyssey" effort, in fact, has a main mission of bolstering the reliability, security, and scalability of Linux on Xeon servers and positioning that combination as an alternative to HP-UX on Itanium; this is possible through contributing to the open source Linux effort and creating scalable NUMA machines based on Xeons. (See our interview with Kirk Bresniker, who is chief technologist for HP’s Business Critical Systems division, about the resurgence of big NUMA systems as envisioned by Project Odyssey.)

For whatever reason, HP decided against funding the porting of HP-UX to Xeon processors and has not said why. But think about it. HP-UX generally runs an Oracle database and applications from Oracle or SAP or another big software supplier. Those databases and apps run on either Windows or Linux. The NonStop stack is completely controlled by HP and runs a mix of third-party applications tailored for very precise markets or, more often, homegrown applications. There is no alternative platform.

EnterpriseTech asked about when HP decided to port the NonStop stack to Xeons. Meyer said that HP had "investigated doing such a port a number of times of the past several years," but would not disclose when the decision to do so was actually was made. The company is also not saying when the Xeon port of NonStop will be delivered; Meyer said to expect it "over the next few years."

The way HP is positioning the development effort for NonStop on Xeon is that it is less of a port and more of an integration into a normal cluster environment at enterprise shops.

Not only is HP shifting from Itanium chips to Xeons, but it will also be dropping its proprietary ServerNet interconnect, which along with the database provides the fault tolerance in the clustered system, and shifting to InfiniBand interconnects. The current development systems in the HP labs are based on "Ivy Bridge" Xeon processors from Intel and 40 Gb/sec InfiniBand adapters and switches from an unnamed supplier.

The NonStop software is itself written in C and C++ with some old Tandem Application Language code still lurking down in there in spots, according to Meyer. HP shifted from MIPS engines to Itanium in 2005, so it has the skills to do another port from Itanium to Xeon and from ServerNet to InfiniBand. (Believe it or not, Meyer said that there are still a lot of MIPS-based nodes out in the field, but these will be at the end of their support life in 2013 or 2014, depending on the model.)

The NonStop platform also supports JBoss and Tomcat middleware and can run Java applications, and these will already execute on Xeon iron. As for other applications written using the compiler tools native to NonStop, HP is guaranteeing that they will be 100 percent source code compatible. And there will be binary translators in the system as well to allow code compiled for Itanium to run unchanged on Xeon. If companies are not sensitive to the performance of a piece of code, this will be an option. (Similar translation features were available for HP-UX shops moving from PA-RISC chips to Itanium processors a decade ago.)

The NonStop port to Xeon is no doubt being driven by the unexciting nature of the Itanium roadmap. Officially, there is a kicker to the current eight-core Poulson Itaniums, but the roadmap was quietly changed in January this year.

What was supposed to be a unique "Kittson" design with a common Xeon-Itanium processor socket and a shrink to 22 nanometer processes to give the chip a lot more capability  has become a Kittson chip implemented in 32 nanometer processes that is socket compatible with the Itanium 9300 and 9500 chips but not the Xeons. Intel has not said if there is a chip beyond this modified Kittson processor, which really should be considered a Poulson+ at this point. Intel has said that the "modular development model, which converges on a common Intel Xeon/Intel Itanium socket and motherboard, will be evaluated for future implementation opportunities."

It would be unwise to plan on such a product ever coming to market. That said, it would be very interesting to see future Xeon E7 chips equipped with features that make it easier for code written for the Itanium chip to be run on them. But even that seems unlikely.

In the meantime, customers using the NonStop clusters can boost the performance of their processor nodes by approximately 50 percent by shifting to the new Poulson chips.

The NonStop NB56000c system announced this week at the NonStop customer and partner conference in Santa Clara, California is based on the Integrity BL860c i4 blade server. It slides into the same BladeSystem c7000 chassis that is used for Xeon-based blades and that is also modified to create the high-end Superdome 2 HP-UX servers. The NonStop node is actually based on the four-core Itanium 9550, which runs at 2.4GHz. The prior NB54000c node used the 1.6 GHz Itanium 9340 processor, and that clock speed change accounts for most of the 50 percent performance boost.

Meyer tells EnterpriseTech that the new NB56000c nodes for NonStop database clusters have a price that is similar to the NB54000c. An entry configuration aimed at small banks sells for under $100,000.

Add a Comment

Do NOT follow this link or you will be banned from the site!
Share This