Advanced Computing in the Age of AI | Thursday, March 28, 2024

HP’s Itanium Odyssey Ends With Superdome X Xeon Iron 

Companies that have been looking for a capacious shared memory system from Hewlett-Packard and who have been waiting for years for the company to deliver a variant of its Superdome machines based on Intel's Xeon processors do not have to wait any longer. The day has come, and the "DragonHawk" Superdome X systems have landed – at the HP Discover conference in Barcelona, Spain, to be precise.

HP is also previewing a variant of its NonStop fault tolerant clusters that are based on Xeon processors and use InfiniBand networks to link machines together. The prior generations of Superdome and NonStop machines were based on Itanium processors from Intel, and the launch of these new systems marks the end of an era even if HP will support Itanium-based systems out to the end of the decade.

Back when "Project Odyssey" was announced by HP in November 2011, the company did a number of different things. First, HP anointed Linux and Windows running on enterprise-class Xeon machines as the future of mission critical computing. And second, then-new CEO Meg Whitman put the kibosh on the idea that the HP-UX variant of the Unix operating system would be ported to X86 iron.

This HP-UX porting effort was one that HP was encouraged to take on by many in the Unix community and a number of key executives in its Business Critical Systems division over the past decade. We know of two HP-UX porting projects that were canceled, but there were likely more. Project Redwood was a plan from early 2010 that proposed to port HP-UX to Xeon E7-class processors in early 2011. Project Kinetic followed later that year, after Redwood was shot down, and included a converged hardware platform supporting both Itanium and Xeon servers in a common socket, with Itanium and Xeon versions running two different flavors of HP-UX – one tuned for each processor. This plan also did not come to pass, and neither did the common Itanium-Xeon E7 socket, which was spiked in early 2012 along with a plan to have the next generation of Itanium, called "Kittson," etched in Intel's 22 nanometer processes. Intel has not said much about the future of Itanium for nearly three years now.

What became clear as Hewlett-Packard and Oracle were fighting in the courts over the fate of the Itanium processor a few years back (and the source for the internal documents that describe Project Redwood and Project Kinetic) is that the independent software vendors who had ported some 5,000 applications from PA-RISC to Itanium systems running HP-UX were in no mood to do such a port again to HP-UX running on Xeon servers. At that point, software vendors would simply prefer to port to Linux or Windows and have code that can run on any system that supports a Xeon processor. And thus, Project Odyssey was born, with the idea of making Xeon-based Superdome systems with large numbers of cores and terabytes of main memory for applications to play upon.

The plan when Project Odyssey was announced three years ago was to get the DragonHawk systems out the door within two years or so. Or, to say it another way, about a year ago. But as is usually the case, it takes longer to get the processors, chipsets, and operating systems all lined up, and the Superdome X machines are only now available. (It probably did not help HP's timing that Intel skipped the "Sandy Bridge-EX" generation of the Xeon E7 processors and waited for the "Ivy Bridge-EX" chips to get all of the Xeon family a little closer together architecturally.)

A variant of this DragonHawk machine, tuned up to run SAP's HANA in-memory database, was announced back in June, but was not available for running generic Linux or Windows workloads. This HANA box, code-named "Kraken" just so you can keep track of the code-names and sold as the ConvergedSystem 900, was made available with either eight or sixteen sockets and it was only made available with one type of processor, the fifteen-core "Ivy Bridge-EX" Xeon E7-2890 v2 that runs at 2.8 GHz and has 37.5 MB of cache and a thermal design point of 155 watts.

The DragonHawk Superdome X is a more generic machine and supports a number of different processors in addition to the E7-2890 v2. Customers can also add the 15-core E7-2880 v2, which runs at a slower 2.5 GHz, which only burns at 130 watts, and which costs a bit less. HP is also allowing for the high-end Xeon E7-8891 v2 processor, which has the fastest QuickPath Interconnect port speeds for its 15-cores to play with, to be used in the Superdome X system, and those who want to use the Xeon E7-4830 v2, which has ten cores running at 2.2 GHz and only burns at 105 watts, that is an option, too.

The Superdome X system has an extended BladeSystem c7000 chassis that has been made tall to support the interconnect that links the nodes together and that could, in theory, allow HP to build a system that supports up to 64 sockets in a single system image. Each blade has four Xeon E7 sockets on it, with 24 memory slots per socket. Each pair of sockets is glued together into a two-socket subsystem using HP's own "Orion 2" chipset to lash them together into two-way nodes. This Orion 2 chipset is part of what HP calls its PREMA architecture, which is short for Performance, Resiliency, Efficiency, Manageability, and Availability. The first generation of Orion chipsets were used in the ProLiant DL980 G7 server, which scaled up to eight sockets using the "Sandy Bridge-EX" Xeon E7 v1 processors from Intel.

In the Superdome X, the Orion 2 chipset hooks into the QuickPath Interconnect (QPI) point-to-point interconnect ports on the Xeon E7 chips on one side and then out to a variant of the sx3000 chipset that is employed in the Itanium Superdome 2 machines. The sx3000 is what is often called a node controller or a crossbar switch, and it links multiple two-way nodes together so they can present a single memory space to the operating system. Intel's own "Patsburg" C602J chipset does the same thing for up to eight sockets if the system is using the E7-8800 variant of the processor. The difference is that the E7-2800 processors are a lot less expensive than the E7-8800 chips, so by making its own NUMA machine, presumably HP can shift more of the hardware profit margin from Intel's books to its own. (But again, for those customers who need more QPI bandwidth, HP has left them the option of using E7-4800 or E7-8800 series chips.)

Moreover, the Superdome X machine can out twice as much compute under the same maximum 12 TB memory footprint, which is important for some workloads. With the two variants of the E7-2800, the system can bring 240 cores to bear in a single image, which quite frankly is larger than a lot of production clusters in the HPC space. With the E7-4800 and E7-8800 processors, the machine tops out at 160 cores.

As we have pointed out before, there is no reason why HP cannot support a single 24 TB of memory space if it wants to employ 64 GB memory sticks in the various Superdome X machines. And if HP shifted up to a 64-socket system, which the sx3000 node controller can do because it does this already in the Itanium Superdome 2 machines, HP could in theory support 96 TB of memory in a single address space. The only problem is that the current crop of Xeon processors have 46-bit physical memory addressing, which means that a Superdome X system actually tops out at 64 TB of memory, no matter how many sockets it has. Still, HP has options to expand the Superdome X line further than these initial DragonHawk machines if customers want larger core counts, memory capacity, or both.

One of the things that has made the jump from PA-RISC and Itanium Superdome machines to the Superdome X is HP's nPar hardware partitions, which offer electrical isolation between chunks of processing and memory. The number of nPars has not been divulged in the specs, but in general HP's architecture allows for an nPar partition to be as small as a single cell board (in this case, a two-socket segment) and as large as the entire system. HP says that nPar hard partitions have about 5 percent of the single points of failure of software-based partitions, and therefore are about 20 times more reliable, and adds that modeling of applications running in the two kinds of environments and field data from customers support this data. HP adds that the Superdome X will have about 60 percent less downtime overall than other X86 platforms, again based on modeling and field data from customers, but the specifics on the compared machines were not revealed.

As for performance, Jeff Kyle, director of product management for the Mission Critical Systems division at HP, tells EnterpriseTech that the machine exhibits "near linear scalability" as gauged by the SPECjbb2013 Java benchmarks, and the documentation for the new Superdome X machine says that the system delivers up to nine times the performance as the ProLiant DL980 G7 system, and eight-socket machine based on the "Westmere-EX" Xeon E7 v1 generation of processors. A fully loaded ProLiant DL980 G7 would have 80 cores running at a maximum of 2.4 GHz with 4 TB of main memory; the top-end Superdome X would have 240 cores running at 2.8 GHz with 12 TB of memory. That's four times the cores and three times the memory, and yet nine times the performance. HP adds that transaction response time on the Superdome X will be up to four times that of other high-end X86 systems, again gauged by the SPECjbb2013 Java benchmarks.

HP is not providing pricing on the Superdome X machines, but does say in its data sheets that a Superdome X will deliver 32 percent lower total cost of ownership than a similarly configured Power8 system from IBM. HP did provide preliminary pricing to EnterpriseTech for the Kraken HANA-only ConvergedSystem 900 machine back in June, and there is every reason to believe that the pricing will be similar for the more generic Superdome X system. A loaded 16-socket ConvergedSystem 900 with 12 TB of memory had a list price of $1.56 million.

HP's data sheets did not say what variants of Linux and Windows will be supported on the Superdome X systems, and it looks like Linux is coming out first because Windows Server is not even mentioned in any of the material that HP provided. HP confirmed to EnterpriseTech that Red Hat Enterprise Linux 6.5, 6.6, and 7 and SUSE Linux Enterprise Server 11 SP3 and 12 will be the first operating systems supported on the machine, since these are the latest-greatest versions from the two big Linux distros. Kyle said that there would be a version of the ConvergedSystem 900 that would support Microsoft's SQL Server 2014 and "Hekaton" in-memory database technology, so we know that Windows can run on the DragonHawk/Kraken machinery. And in fact, the plan is to certify Windows Server 2012 on the Superdome X in 2015.

Working NonStop

In addition to the new Superdome X machine, HP is also previewing its Xeon-based Integrity NonStop X server line.

NonStop X server image

The NonStop systems are fault tolerant clusters that are currently based on Intel's "Poulson" Itanium 9500 processors and that use a proprietary interconnect called ServerNet to lash machines together to run the NonStop kernel and a distributed database. NonStop machines are the backbone of many financial exchanges and banking networks where downtime is not permissible. Because most of the applications running on the NonStop platform are homegrown – unlike the case for the HP-UX variant of Unix – HP can port the NonStop environment to the X86 architecture and customers can recompile their own code, and given the price/performance benefits of shifting from Itanium to Xeon, many must be so inclined or HP would not have bothered.

Beginning in March next year, HP will be shipping a new lineup of NonStop X systems. Kyle says that the new NonStop X machines will replace Itaniums with the a future "Ivy Bridge" Xeon E5 processor and that it has ported the ServerNet fault tolerant clustering to run atop of an InfiniBand network running at 40 Gb/sec speeds (that's called FDR in the InfiniBand lingo). The NonStop X will use the BladeSystem c7000 chassis and ProLiant BL460-class blade servers. HP tells EnterpriseTech that it will be able to put 64 cores and 3 TB of memory inside of a single NonStop node, which will be comprised of multiple motherboards.

The InfiniBand network will offer 25X the interconnect bandwidth as the ServerNet it replaces, and by switching to Xeon processors HP will be able to jack up the compute density on the NonStop clusters by a factor of two. Also, adds Kyle, HP will be able to provide about 50 percent more performance in like-for-like configurations of the Itanium-based NonStop and Xeon-based NonStop X machines.

EnterpriseAI