Server maker IBM is touting its ability to run ERP software from SAP as a means to promote its Power Systems servers running AIX, and it is using an SAP Sales and Distribution (SD) benchmark test to try to prove the value of its Power-based machinery.
For whatever reason - and we all have our guesses on this - IBM has not seen fit to run the SD test on Power-based servers running OS/400 or i operating systems for a number of years. This is odd, considering that OS/400 shops seem to be natural pairs to SAP software, given their desire for control and tight integration.
But it still does run the popular SD test - which is one of a handful of de facto server benchmarks for gauging the relative performance of servers within an architecture and across it - on Power iron using its AIX operating system and DB2 database variant for Unix, Windows, and Linux platforms. We are left to infer how well or poorly the i/DB2 for i combo does or how Linux on Power might do.
Last week, IBM announced the results of a two-tier run on the SAP SD test on a Power 550 server using the Power6+ chips. The machine was configured with four 5 GHz Power6+ chips, for a total of eight cores and 16 threads, and 64 GB of main memory. (The Power6+ chip has 128 KB of L1 cache and 4 MB of L2 cache per core and 32 MB of L3 cache shared by each dual-core chip.)
With AIX 6.1 and DB2 V9.5, and the SAP ERP 6.0 code, the Power 550, working as a database server, was able to push up to 99 per cent CPU utilization and drive 1.23 million dialog steps per hour, supporting 3,752 users with an average response time of 0.97 seconds.
IBM is touting this as the fastest eight-core result, but there are not very many eight-core results out there at this point that matter, except two-socket "Gainestown" Xeon 5500 machines, which come pretty close to IBM's performance. Bigger x64 servers, using the six-core "Dunnington" machines - and having four sockets just like the Power 550 - will smoke the Power 550, and have the same number of processor sockets.
If software vendors price their applications based on processor core count (as IBM does), then the fact that IBM can cram more performance into fewer cores (thanks to the higher clock speed of the Power6+ chips relative to most x64, Itanium, or RISC chips) matters a whole lot. If vendors price based on the performance inherent in the server, or on the basis of processor sockets, then the Power line can have no advantage or may even be at a disadvantage.
The SAP tests don't incorporate any economics, so you can't do any price/performance analysis. In the next few days I will fix that, considering that raw performance and price/performance have to both be taken into consideration when comparing across server architectures.
The full list of SAP SD benchmark tests can be found here, and for the sake of this discussion, I am just going to compare the Power 550 to its midrange peers.
Hewlett-Packard has just done its own SAP SD test on a four-socket ProLaint DL585 machine configured with Advanced Micro Devices' new "Istanbul" six-core Opteron 8439 SE processors, which technically don't debut until today. AMD doesn't have simultaneous multithreading on its chips, so the core count is the thread count, and this box brings 24 cores running at 2.8 GHz and 64 GB of DDR3 main memory to bear.
Using Windows Server 2008 Enterprise Edition and SQL Server 2008 as well as the SAP ERP 6.0 code, this HP ProLiant box was able to process 1.53 million dialog steps per hour on the SD test while running at 99 per cent CPU utilization (as a database server, not running the apps in a single tier with the database) and delivered 0.96 second average response time. This four-socket box supported 4,665 users. Now, that's 24.3 per cent more users per machine or per socket compared to the Power 550 that IBM is bragging about. However, the HP box only supports 194 users per core, compared to 469 users per core on the Power 550.
See how software pricing strategies on multicore machines can really make a difference? Or conversely, how difficult it is to pin a software price to a system, a processor socket, or a core?
By the way, a similarly configured ProLiant BL685c four-socket blade server, configured with slightly slower (and much less hot) 2.6 GHz six-core Opteron 8435s and 64 GB of memory, was able to do 1.45 million dialog steps at peak utilization and sub-second average response, supporting 4,422 users.
Coming at the Power 550 from below are the new quad-core Xeon 5500 servers, which are only two-socket boxes, but are in the same performance class as the Power 550 if you do it by core count. Fujitsu just tested its Primergy BX920 S1 blade server on the SAP SD test using the Windows stack, and an eight-core, 16 thread box with 48 GB of main memory running full-tilt with sub-second response time could do 1.07 million dialog steps per hour and could support 3,260 end users.
A Primergy TX300 S5 tower server and a RX300 S5 rack server from Fujitsu running the same software with 48 GB of memory was able to support 3,328 users. The Power 550 beat these Fujitsu Gainestown machines, which were all configured with two of Intel's quad-core X5570 processors running at 2.93 GHz, to be sure. But without even doing the math, I have a sneaking suspicion that the Intel box is going to cost a lot less than the Power 550. HP's similar Proliant DL380 G6 rack servers and BL460c blade servers using the same X5570s supported 3,300 and 3,310 SD users on the Windows stack.
By the way, a Gainestown-based X4270 server from Sun Microsystems, running its Solaris 10 Unix and the Oracle 10g database using two of the same quad-core Xeon X5570 chips and 48 GB of memory, was able to support 3,700 SAP SD end users. That is within spitting distance of the 3,752 users that IBM is boasting about with the Power 550.
Long ahead of the March launch of the Intel Gainestown chips, back in December, IBM apparently ran the SD benchmark on its System x3650 two-socket box with two X5570 processors and using Windows Server 2003 Enterprise Edition (not Windows Server 2008), coupled with its DB2 V9.5 database and with 48 GB of memory, was able to support 5,100 SD users. However, driving up that user number could also push the average response time on the SD transaction up to 1.98 seconds. So this is not comparable to the other sub-second results.
A two-socket HP ProLiant BL465c blade server using six-core Opteron 2435 chips running at 2.6 GHz could support 2,355 SD users, and a rack-style ProLiant DL385 using the same chips came in at 2,350 users. I am not sure why these Opteron boxes did so much more poorly than the Xeon 5500 box. They should be in roughly the same performance class.
A four-socket Dunnington Xeon X7460 box from NEC is in roughly the same performance band as the Power 550 and the two-socket Xeon X5570 machines. And in fact, the two-socket Gainestown machines offer better performance than the quad-core X7400 series chips, which is why Intel really needed to get the future Nehalem family of "Beckton" eight-core Xeon 7500 chips to market soon, rather than late this year.
Anyway, the NEC Express5800 server tested back at the end of May using the six-core X7460 chips running at 2.66 GHz (for a total of 24 cores) and 64 GB of memory running the Windows stack was able to handle 970,000 dialog steps per hour with 0.97 second average response time at 98 per cent CPU utilization, supporting a total of 2,957 users on the SAP SD test. This NEC machine can scale as far as 96 cores, delivering 7,500 users, all in a single architecture. You have to jump to a Power 570 to scale higher than the Power 550, but if you do, a Power 570 using 16 4.2 GHz Power6 chips (that's 32 cores) can handle 14,432 SD users.
To help make a broader case for Power Systems for supporting SAP enterprise software, Big Blue commissioned long-time consulting partner International Technology Group (which doesn't seem to have an identifiable Website, but is located in Los Gatos, California) to do an analysis comparing the cost of supporting SAP workloads on Power Systems versus x64 platforms.
In that study, which is based on using Oracle 10g Real Application Clusters in various scenarios (and based on real-world configuration information gathered by ITG) on Windows x64 and AIX Power iron, the Power Systems showed a three-year cost of ownership that ranged from 25 per cent to 33 per cent lower than x64 gear. You can read the ITG report, which is called Value Proposition for IBM Power Systems - Platform Choices for the Enterprise SAP Infrastructure, at this link.
As you can see on page 27 of the ITG report, the IBM Power iron can be a lot more expensive than the x64 alternatives in the scenario, but systems software can be a bit more expensive, and personnel costs are a lot higher over three years - two or three times the cost of the iron, by ITG's math.
I have no idea how typical this spread in personnel costs is out there in the real world, especially considering that these estimates are based on data from a handful of customers. What I do know - and I say this with a peach-eating grin - is that the merger of the RS/6000 and AS/400 is truly complete. IBM may have given the RS/6000 all the glory and price cuts in the past several years to gain market share, but now, when the room to slash prices and compete head-to-head on hardware price with x64 iron is not really practical, it has to use the AS/400 playbook and its TCO arguments even to sell AIX boxes. ®