Power6+ Blade Performance: IBM’s Competitive Analysis
June 15, 2009 Timothy Prickett Morgan
In the computer racket, sometimes you can learn as much from what IT vendors choose to say as you can from what they don’t choose to say. All I know for sure, in my two decades of watching this complex industry, is that you have to look at as much data as you can stomach if you want to try to come out with something that even remotely approximates the truth. And so it is with the updated Power6+ lineup that came out in late April from IBM.
The first thing I noticed in looking at the announcements was how Big Blue wasn’t bragging so much about performance. More importantly, IBM didn’t even tell people that these were Power6+ machines. I had to do that, and if I hadn’t, we’d still not have known that the new Power 520 and Power 550, as well as last October’s Power 560 and Power 570, machines were based on a kicker chip to the Power6. The two are related, no doubt, and I know from looking at roadmaps that IBM was hoping at one point to have Power6+ pack about twice the wallop of Power6, and depending on who you ask, IBM was able to crank clock speeds higher earlier than expected with Power6+ or wasn’t able to crank them high enough to reach its goals. This may be why there was a dearth of competitive analysis information as part of the October 2008 and April 2009 Power6+ launches.
To be fair, Big Blue offers the most detailed and comprehensive capacity planning benchmarks of any server maker, and always has done better than its peers. Power Systems machines have CPW ratings for every processor configuration and speed running i workloads and rPerf ratings for every possible configuration as well; mainframes have LSPR and MSU ratings, which are used in the same way. No other vendor even tries to give out such numbers any more, but in the early days of the commercial Unix server business, Hewlett-Packard and Sun Microsystems gave out detailed relative performance figures for their respective PA-RISC and Sparc iron, mimicking IBM in a good way. None of the X86 and X64 server makers ever offered anything close to what IBM does for its Power and mainframe server performance metrics, mainly because compute capacity is perceived as being so cheap that buyers don’t have to count every dollar. (Boy, were they every wrong about that once the sprawl took on a life of its own.)
My point is that all vendors rely on using a suite of various benchmarks–some controlled by benchmarking institutions such as the Transaction Processing Council and the Standard Performance Evaluation Corporation, others controlled by application software makers such as SAP, Oracle, or Lawson–to prove how their machines stack up against the competition, and they do their best to stay away from workloads where other architectures have the performance or price advantage (if pricing is part of the metric). And now, with power consumption being added to various benchmarks, electricity-hungry architectures that show poor performance/watt will avoid these tests like the swine flu.
The uptake of Power-based blade servers, particularly those running the i 6.1 operating system, has not been as large as IBM had hoped (that’s my guess based on conversations and hints, not hard data based on IBM’s projections for sales versus its actual sales over the past 18 months). And that is why in all of the IBM presentations I saw from the April 28 announcements, the only competitive analysis bit that I saw was for the JS23 and JS43 blade servers compared to HP and Sun Unix alternatives. (The JS23 is a two-socket, four-core blade server that was announced in April and that I detailed in the May 4 issue, and the JS43 is a four-socket, eight-core blade that is created by snapping two JS23s together.)
To make the sales case for the JS23 and JS43 blade servers, IBM chose to focus on the SPECint_rate2006 integer benchmark for multiprocessor systems. Raw integer performance bears some resemblance to the kind of work that commercial computers do, but it is not the same thing as running a database or application server on a box and seeing how it performance. SPECint_rate2006 is one of a family of tests in the CPU2006 suite of benchmarks administered by SPEC, which was preceded by the CPU92, CPU95, and CPU2000 tests. (You can find out all about SPEC CPU2006 here, and you will discover that the integer test email, data compression, vehicle scheduling, video compression, XML processing, network simulation, gene sequencing, and a few other applications all running at once, as well as a game of chess and another of Go.) The SPECint2006 test measures how quickly a machine processes a single task, while the SPECint_rate2006 test measures how much throughput you can get out of a box performing many tasks. Usually, SPECint_rate2006 benchmarks run one copy of the benchmark suite per core or thread, if the cores are multithreaded.
So here’s how the JS23 and JS43 stacked up against an HP Itanium-based BL860c blade and a Sun T6340 blade:
The JS23 was rated at 110 on the SPECint_rate2006 test using the 4.2 GHz Power6+ processors and with four cores activated. By IBM’s calculation, that blade, when configured with 32 GB of memory and a single 146 GB disk plus AIX 6.1 and the PowerVM Standard Edition hypervisor would cost $11,878, or just under $108 per SPECint_rate2006 unit of performance. (IBM ran the test with Novell‘s SUSE Linux, but priced it with AIX, which is a bit of a no-no.) The JS43 was rated at 219 on the test, almost precisely twice the performance as the JS23 (as you would expect considering it is just two JS23s snapped together), and cost a little more than twice as much at $23,756 for a configuration with one disk and 64 GB of memory, or just over $108 per unit of performance.
Because IBM is trying to position its Power-based blades against other high-end Unix iron, among the myriad machines IBM could compare the JS23 and JS43 blades to, it chose HP’s Integrity BL860c running HP-UX 11i, which is a two-socket blade that is based on Intel’s “Montecito” 1.6 GHz dual-core Itanium 9040 processor. This machine is rated at 53.9 on the SPECint_rate2006 test, and IBM says that with a base HP-UX license (with no virtualization) it costs $13,971, or $259 per unit of performance. That’s pretty awful by comparison, and that’s no surprise given that this machine is more than two and a half years old.
The Sun blade, which is a two-socket blade based on Sun’s “Niagara” Sparc T2+ processors, has chips with eight cores and 64 threads per socket for a total of 128 threads. That is as many threads as IBM puts into a high-end Power 595. Pity the Sun chips only run at 1.4 GHz, compared to 5 GHz for the Power6 chips used in the Power 595s. Anyway, that two-socket Sun box still does pretty well, at a 160 rating on the SPECint_rate2006 test, but with an $83,273 price tag, it is no wonder that Sun is being acquired by Oracle and hardly anyone believes that Oracle will be keeping the Sparc hardware alive for very long. That’s a pitiful $520 per unit of performance, which is deeply embarrassing.
Somewhat embarrassing is IBM’s selective use of the rather huge list of SPECint_rate2006 benchmarks, which is so big that when you load the file you might think your browser is crashing. (Just let the data download and be patient.) Moreover, considering that we are talking about Power Systems versus commercial servers, it might be interesting to see how the cost of a database management system changed the balance and it would have been fair to put i 6.1 and AIX 6.1 both to the test as well and prove that they are just as good as Linux on the test.
Finally, IBM did not go anywhere near making comparisons with the new “Nehalem EP” Xeon 5500 two-socket servers, which are the real competition for the JS23 and JS43 blades. Big Blue’s own BladeCenter HS22 blades configured with these quad-core Nehalem EP Xeon X5570 chips running at 2.93 GHz and 24 GB of memory were rated at 252 on the SPECint_rate2006 test using Red Hat Enterprise Linux. With a single 73 GB SAS drive slapped on it and the same hardware used in the SPEC integer test, costs $8,245 for the blade server and with a standard subscription to RHEL 5 costs $799, for a grand total of $9,044 or $36 per SPECint_rate2006 unit of performance. Even with a premium Linux subscription, the price only rises to $9,544, or $38 per unit of performance.
Like I have been saying for a long time now. IBM can’t wait until Power7 to compete with such X64 iron. The time is now. That gap in price is just too large.