IBM's Power7 pitch deconstructed
Big Blue polishes UNIX crown
Some of IBM's Power7 machines have been shipping for several weeks, and the high-end Power 770 and 780 boxes start shipping this coming week. Now, the sales pitching and smooth talking by IBM and its local business partners will begin. What, exactly, will that sales pitch be?
That all depends on what gear you have installed, how old it is, who made it, and what applications it runs. One thing is for sure: IBM is trying to get out in front of a whole lot of upcoming server iron to show some good numbers.
Business partners make their own pitches to their customers, of course, but they take their cues from IBM. And as you well know since even before the Power Systems division was created from the merger of the formerly independent System i and System p divisions, IBM is keen on dividing and conquering the server market and bringing as many workloads as it can to the Power7 lineup.
This chart, culled from a presentation IBM gave to partners a few days ahead of the launch of the Power7 machines, succinctly sums up how Big Blue wants to push its Power7 boxes this year:
First and foremost, IBM wants to remind everyone that it has dominant Unix market share (in terms of revenues) and that compared to its $5bn in sales, Hewlett-Packard and Oracle (now that it has assumed control of Sun Microsystems) have a combined $8bn in sales, split about evenly between the two. IBM was able to make $600m in sales (presumably including software and services as well as hardware) last year on HP and Sun takeouts, and it wants to accelerate these. If a Power Systems migration is not in the cards, IBM is perfectly happy to bolster its share of the $30bn X64 server space by moving HP-UX and Solaris workloads to Linux on X64 iron.
While the chart doesn't say this, there is also some migration to Windows going on here and there. IBM also wants to go into HP, Dell, and Fujitsu shops using X64 iron and consolidate multiple workloads onto Power-based servers. This is a familiar plan, and one that IBM's AS/400 faithful have seen since PC-based server cards were first slapped into AS/400s back in the early 1990s and that was augmented as logical partitioning debuted back in the late 1990s on the AS/400. (And yes, well ahead of LPAR support for AIX that was worth a damn. But don't expect the Unix-centric Power Systems people to mention that).
As you can see, IBM wants to "minimize leakage" from Power Systems to X64 boxes, and it doesn't have any arrows that would show Power or X64 workloads moving to HP or Oracle Unix iron. Like that never happens. Of course it happens, and there are hundreds of millions of dollars in consolidation at stake there too. Interestingly, there is nothing in this chart that mentions i/OS specifically and that says there is a goal to accelerate migrations from X64 iron to Power-i/OS combinations or to minimize leakage of i/OS workloads to X64 iron. But there should be, and hopefully, there is even though the presenters briefing business partners did not get into the i/OS strategy.
AMD and Intel on the way
There are a whole lot of new chips coming down the pike from Intel and Advanced Micro Devices, and if IBM had waited until May for the initial Power7 launch, it would not be able to get business partners psyched up with a chart like this:
This chart compares the performance of various servers on SAP's two-tier Sales & Distribution (SD) ERP benchmark test. A four-socket Opteron 8400 box using the latest six-core chips does about the same amount of work as an Oracle/Sun T5440 server (also four sockets) using the eight-core Sparc T2+ processors. With 15,600 SD users supported, the four-socket, 32-core Power 750 server was able to do more than three times the work. (That machine is configured with AIX 6.1 and DB2 9.7 and uses the 3.55 GHz Power7 cores with all the threads turned on).
Both AMD and Oracle/Sun are on track to more or less double the performance of these machines by the middle of this year, and interestingly, AMD is not planning on offering an eight-socket version of the future 12-core "Magny-Cours" Opteron 6000 systems, and Sun will not push beyond four sockets with the Sparc T3s later this year, either. So the Power 750 will likely have a sustainable advantage here in terms of performance.
It is unclear how HP Integrity machines (what IBM still calls Superdomes in its presentations) using the new quad-core Itanium 9300s (formerly known as "Tukwila") will perform on the SAP SD benchmark, but HP did test what I presume was an I/O and memory constrained BL860c Itanium 9140M blade (with two 1.66 GHz cores in each of its two sockets) and was able to support 1,165 SD users on the same test running HP-UX 11i v3 and Oracle 10g, but this machine had average response times of just under two seconds (the norm back then) instead of sub-second (which is the norm now).
A 32-processor, 64-core Fujitsu PrimeQuest 580A server tested in April 2008 was able to handle 10,400 SD users, but with a nearly two-second response time. The Tukwila processors and their new memory architecture, based on the QuickPath Interconnect, are going to have to do a whole lot better to compete. And if they cannot compete on performance, they are going to have to make it up with price/performance. Ditto for the Sparc T3s and any servers from Oracle or Fujitsu supporting Solaris on quad-core Sparc64-VII chips.
Claims to fame
As all vendors accentuate the positive, IBM is throwing around a lot of claims. For instance, business partners are being told that a single Power 740 with four sockets can replace 92 Sparc T2000 servers, cutting back the number of cores by 95 per cent, the square footage of floor space by 97 per cent, and the energy used by 95 per cent. IBM also has another slide that uses the SPECint_rate2006 benchmark to "prove" a Power 750 with four-sockets (32 cores) can do 29 per cent more work than an Integrity Superdome with 64 1.6 GHz cores, but it burns only 1,950 watts instead of 11,586.
If you read the fine print, these are not the wattages of the systems as they are running the test, but the maximum power usage for data center site planning, which as far as I am concerned is playing it a bit loose. But the SPEC tests don't all track power consumption, and if you were trying to get a feeling for performance per watt, you would have to do exactly what IBM did.
The Power 750 apparently smokes four-socket Sparc, X64, and Itanium systems in terms of performance and performance per watt on the SPECint_rate2006 test. With a SPECint_rate2006 of 1,060, that smoked a Sparc M4000 server with 2.52 GHz Sparc64-VII chips (152 rating with 2,350 watts), a Sparc T5440 with 1.6 GHz Sparc T2+ chips (360 rating with 2,700 watts), an HP rx6600 with Itanium2 chips (102 rating with 1,600 watts), and an HP ProLiant DL580 G5 using older six-core Xeon 7400 chips (291 rating with 1,412 watts).
A two-socket Xeon 5500 box can do almost as much work (254 rating using the 2.93 GHz Xeon X5570, and a Westmere-EP kicker will probably be on the order of a 380 rating on the test). HP has four-socket Opteron boxes rated at nearly 400 on the SPECint_rate2006 test, but IBM didn't bring this up. And HP will be pushing up to around 800 or so with the Magny-Cours later this quarter. That Magny-Cours box will have 48 cores and no threading compared to the Power 750 with 32 cores. IBM will still be able to show more oomph per core on this test, but AMD and its partners are going to cram more cores into the box.
The other thing we noticed in the sales pitch is that IBM is assuming that a replaced machine has lower utilization (but is running at maximum electrical consumption) and that when you consolidate onto a Power7 box, you can run it at higher utilization and thereby save big bucks. Obviously, many i/OS, AIX, HP-UX, and Solaris shops are running at high utilization, so the savings IBM shows may not materialize. But just to show you one example so you can be careful of this.
In one presentation, IBM says that a Power 780 can consolidate nine 64-core Integrity (well, it said Superdome) boxes into a single Power 780 and save 91 per cent on floor space and cut the number of cores by 88 per cent. Those nine HP Integrity boxes have 576 cores running at 1.6 GHz, but IBM's comparison assumes that they are only running at 25 per cent utilization (you gotta read the fine print), while the single Power 780 with its 64 cores running at 3.8 GHz (a full rack, taking up only 7.6 square feet of floor space) hum along at 75 per cent utilization.
While IBM is offering three times the performance per core compared to the dual-core Itanium 9000/9100 series, the utilization levels it set in the comparison account for another factor of three in compression. If you assume the machines are running at the same utilization (which they would be on a test like SPECint_rate2006), then IBM has only a factor of three compression, and if you go with the full-out 128-core HP Integrity box, then the Power 780 has only a 54 per cent performance advantage if you look at the actual SPEC tests. And by the way, a Sparc Enterprise M9000, rated at 2,586 on this SPEC test, beats the 780, although by IBM's reading of the Oracle/Sun manuals it burns seven times the juice at maximum loading.
Here's another comparison IBM cooked up, which kills the ProLiant and Integrity birds with one comparison stone:
These comparisons above are based on the SPECint_rate2006 test as well, which seems to be IBM's favorite with the Power7 launch. And the CPU utilization on the unvirtualized ProLiant boxes is assumed to be low (19 per cent according to the fine print at the back of the presentation) and is at 60 per cent on the Power 750.
Ultimately, what customers need to look at to do comparisons is raw data on the systems themselves, then they need to examine benchmarks and other performance data, gather up pricing information, and finally try to weave together a picture that has more depth and breadth than the IBM sales pitch. ®