Original URL: https://www.theregister.com/2010/05/10/superdome_2_tpc_h/

BI benchmark outs HP Superdome 2 details

Not shipping til October 20

By Timothy Prickett Morgan

Posted in Channel, 10th May 2010 18:31 GMT

More details of HP's forthcoming Tukwila Itanium 9300 servers have come to light thanks to TPC-H data warehousing benchmark test results that HP has just published.

When HP announced the first of its Itanium 9300 servers three weeks ago, the company's techies gave out some of the salient characteristics of the new blade-architected Superdome 2 servers. But a lot of the feeds and speeds were unanswered about the machines, which are due to ship in the second half of this year. The new TPC-H numbers fill some of the gaps.

As El Reg previously reported, the Superdome 2 servers have replaced the four-socket cell board architecture of the three prior generations of Superdome boxes, made possible by three different HP chipsets married to the Itanium architecture. As we've learned from the TPC-H data warehousing benchmark, HP has made a massive two-socket Superdome 2 blade server called the CB900s i2 that slides into a modified c7000 chassis that had previously been used for x64 and Itanium blade servers.

The bottom 10Us of the CB900s i2 blade server looks exactly like any other HP blade server, plugging into the blade enclosure's VirtualConnect or other kinds of switching and its Insight Control management tools. The upper 8Us of the Superdome 2 blade is where the sx3000 chipset made by HP does its work, gluing together 16 sockets - that's eight blades with a total of 64 processor cores - into a single system image.

This time around the Superdome crossbar implemented in the sx3000 is fully redundant and fault tolerant, unlike HP's prior Yosemite, Pinnacles sx1000, and Arches sx2000 chipsets. The CB900s i2 blade has 32 DDR3 memory slots, but HP did not say what capacities of memory it was supporting.

According to the TPC-H test result, the Superdome 2-16s server costs $56,500. Each CB900s i2 blade server is equipped with a top-bin 1.73GHz Itanium 9350, which has a 24MB L3 cache and a Turbo Boost speed of 1.86GHz. In the pricing portion of the test, HP labels this as the real clock speed - which it is not - but elsewhere in the report it's labeled correctly at 1.73GHz. (You can see the feeds and speeds of the five Tukwila Itanium chips, which were launched in early February, here.)

With two Itanium 9350s, that CB900s i2 blade costs a whopping $44,000, and even at list price, the two Itanium chips cost only $7,676 in 1,000-unit quantities. That 31.5-inch blade server, devoid of CPUs and memory, costs $36,324 - not bad for a big ol' motherboard with some chipset magic in it.

HP has cooked up a special 512GB memory bundle for the Superdome 2-16s machine, which costs $112,000 - that's $219 per GB. HP didn't provide the memory capacity of the Superdome 2 machines when it announced them in April, but using 4GB memory sticks, HP should be able to put 1TB of memory on the eight blades in this box.

The Superdome 2-16s system, CB900s i2 blade, and memory bundle won't be available until October 20, according to the full disclosure report from the TPC. Ditto for the HP-UX 11i v3 license for the machine, which will cost $5,450 per socket.

As El Reg reported last week, HP tweaked its HP-UX 11i v3 Unix variant with the March 2010 update, adding support for the Tukwila Itaniums to the operating system along with a number of other features, including live migration of running guests inside of Integrity VM partitions.

HP also said - vaguely - that it had shifted to per-socket pricing for its new round of Tukwila systems, which would result in as much as a 50 per cent price cut compared to per-core licenses on early Itanium-based machines. The Base Operating Environment license on prior Superdome machines, as this story revealed, cost $2,370 per core. Had the pricing remained the same, HP-UX would have cost $151,680 on this 16-socket Superdome 2 box. But the March 2010 update will only run $87,200 on this machine, representing a 42.5 per cent cut in software license costs.

To complete the Superdome 2 system tested using the TPC-H data warehousing benchmark, HP tossed in four Superdome 2 external I/O enclosures, a ProCurve 6120XG blade switch, and a rack, bringing the hardware price to $579,138. Toss in another $321,742 for three years of maintenance, another $566,665 for systems software (HP-UX plus Oracle 11g), and $709,617 for MSA 2324fc Fibre Channel disk arrays with 41TB of disk capacity spread across 580 disks, and the whole shebang costs $2.18m.

Then shave off 21.8 per cent in discounts, bringing it down to just over $1.7m.

Battle of the bangs for the bucks

The Superdome 2-16s machine was able to crank through 140,181 queries per hour (QPH) on the 1TB version of the TPC-H benchmark suite, yield a bang for the buck of $12.15 per QPH after discounts.

Compared to the prior generation of Integrity Superdome servers using the Montvale Itanium machines, the new Tukwila box looks pretty good. But x64 machines from HP are smaller, less expensive, and offer better bang for the buck.

An Integrity Superdome with 32 sockets (that's 64 cores) was able to handle 123,323 QPH when it was tested last April running HP-UX 11i v3 and Oracle 11g (with 384GB of main memory and 110TB of disk), at a cost of $20.54 per QPH. That means the Tukwila box with the same number of cores had 13.7 per cent more oomph on the 1TB TPC-H test.

That Superdome machine was wickedly more expensive at $2.4m. Also, HP-UX cost twice as much, so HP had to slash the prices in half to get the price down to $20.54 per QPH. So the price delta at list price between last year's Superdome and this year's Superdome 2 is a lot wider than the 41 per cent improvement in cost per query per minute implies.

HP is talking up how the new Superdome 2-16s has 4 per cent better performance per core and 17 per cent better bang for the buck than an Oracle/Fujitsu Sparc Enterprise M9000 server with 64 cores. But that machine was tested back in May 2008 using dual-core Sparc64-VI processors running at 2.4GHz, and Oracle and Fujitsu have since started shipping quad-core Sparc64-VII processors running at 2.88GHz, and will soon push that up to 3GHz. (Well, "soon" in the Fujitsu sense of that word, meaning late 2010 or early 2011).

Core for core, the newer quad-core chip yielded about 25 per cent more oomph, according to Fujitsu, so if a current M9000 server were tested, it should be able to best the Superdome 2. And scale to twice as many cores.

The Superdome 2 had better scale to four of those blade chassis and deliver 256 cores of aggregate oomph or HP is going to have competitive issues on many big iron workloads.

But that's not the real threat to the new Tukwila-based Superdomes. Standalone x64 servers are packing serious punch, and clusters of them are doing a pretty good job, as well.

In February, ahead of the twelve-core Opteron launch, HP tested an eight-socket DL785 G6 server using Advanced Micro Devices' six-core Opteron 8439 processors (2.8GHz) with 384GB of main memory and 15TB of disk; this 7U machine was able to handle 102,375 QPH on the same 1TB TPC-H test, at a cost of $3.63 per QPH. This server was configured with Red Hat Enterprise Linux 5.3 and Sybase IQ Single Application Server Edition v15.1.

And in an even more stunning comparison, server virtualization juggernaut VMware partnered up with parallel database provider ParAccel to lash 40 virtualized HP DL380 G6 servers using quad-core Xeon 5560 processors (two per server running at 2.9GHz and with 72GB of memory in each node). This 320-core cluster had HyperThreading turned off and an aggregate of 2.8TB of main memory and 87TB of disk, and was able to crank through 1.32 million QPH at a cost of 70 cents per QPH.

Ouch.

IBM has not yet run the TPC-H test using the 1TB data-set size (or any other) on its new Power7-based machines. IBM and HP tend to stay away from each other in the TPC-H test, which has data warehouses that come in 100GB, 300GB, 1TB, 3TB, 10TB, and 30TB sizes. It seems likely that the two will avoid going head-to-head in future comparisons. ®