IBM Shows Off Power6+ Performance on SAP, Lawson Apps
May 18, 2009 Timothy Prickett Morgan
IBM and its software business partners are putting the new Power6+ versions of the Power Systems i boxes announced on April 28 through the paces, and benchmark results are starting to come out.
So far, I have been able to get my hands on some benchmarks coming out of SAP and Lawson, and it is debatable how useful these tests are considering the lack of comparative benchmarks available for other platforms. (And yes, IBM does this on purpose just to confound anyone trying to do such comparisons.) The benchmark tests are nonetheless useful if you want to do some correlations between IBM’s Commercial Performance Workload (CPW) and real world data warehousing or order entry workloads, as the SAP and Lawson tests do.
First up is the new SAP tests. IBM’s i platform was the only machine used in the prior Business Intelligence – Datamart (BI-D) benchmark test, which IBM used to test a few System i machines running i5/OS V5R4 and the initial Power6-based Power Systems i machines running i 6.1. (You can see those tests here.) Forget that test. Now IBM is using something called the Business Intelligence Mixed Load (BI-MXL) test, which is run without SAP’s BI Accelerator co-processor, a hardware appliance for speeding up SAP’s Netweaver middleware, and which is different from the old BI-D test in a number of ways that do not make them comparable. You can see the four tests that IBM ran on Power Systems i boxes here.
IBM ran the BI-MXL test on the Power 520, 550, and 570 machines using the Power6+ processors, which came to the Power 520 and 550 on April 28 (and will be shipping on May 22) and which shipped on the Power 570 (as well as in the Power 560 not tested) last October. A Power 520 with a single Power6+ processor with two cores running at 4.7 GHz equipped with 32 GB of memory was able to process 41,090 query navigation steps per hour on the BI-MXL test at 94 percent of CPU utilization on a data warehouse with 300 million records. (That’s the smallest database used in the test, which also has 1 million and 3 million record variants.) A Power 550 with two processors and four cores activated running at 5 GHz and with 64 GB of memory was able to handle 90,635 query navigation steps per hour at 98 percent of CPU; for some reason, IBM ran the test again on this box a few weeks later and got a slightly lower 97 percent CPU utilization and only handled 90,492 query navigation steps per hour. (Go figure.) A Power 570 box with four 5 GHz cores and 96 GB of memory did slightly more work, at 93,468 query navigation steps per hour.
The message here is what I have been telling you for years: Don’t buy into the Power 570 unless you have to. It is more expensive for modest workloads than a Power 550. If you don’t need the expansion that the Power 570 embodies, don’t do it. By the way, those machines had 96 of IBM’s 15K RPM disks attached to them apparently, according to sources familiar with the tests, although the SAP reports don’t detail this.
The lesson from these SAP test results is that the i 6.1 servers scale pretty linearly on the data warehousing workload that SAP and IBM have cooked up. But three points do not make a very confident line. I would like to see how the BI-MXL test scales on each machine as from one to the maximum number of cores and an appropriate amount of memory are scaled up on each box, and then compare how each Power machine–520, 550, 560, 570, and 595–does in terms of linear scalability. And then I would tag prices to specific configurations to see which ones for a given throughput on the BI-MXL test gave the best bang for the buck. This isn’t hard, but it stinks too much like commitment to server makers to do it right.
ERP software maker Lawson is very keen on seeing IBM finally put L3 cache onto the Power6+ chips used in the Power 520 server and the companion JS23 blade server. That’s because its M3 (formerly known as Movex) ERP suite is written in Java, and Java applications really need lots of clocks and lots of L2 and L3 cache to boost performance. According to the documents that I have seen, Lawson has tested three different Power 520 machines–one with the feature 7149 card with a 1.9 GHZ Power5+ chip, another with feature 5633 with a 4.2 GHz Power6 chip, and the last with feature 5587 with a 4.7 GHz Power6 chip with 32 MB of L3 cache–on its M3 Order Entry benchmark. In all cases, only one core was activated in the boxes.
On the M3 tests, Lawson found that the initial Power 520 using that old 1.9 GHz Power5 chip could process about 75,000 invoiced order lines per hour. Moving up to the 4.2 GHz Power6 core boosted performance by 31 percent (not more than 2X as you might expect from the clock speeds because IBM radically changed the instruction pipelines with the Power6 chips), to around 98,250 invoiced order lines per hour. And while the move to the Power6+ chip only boosted the clock speed by 11.9 percent up to 4.7 GHz, the addition of the L3 cache pushed performance up 21 percent to 118,900 invoiced order lines per hour. Which makes you wonder why on Earth IBM ever cropped the L3 cache out of the boxes to artificially crimp performance on the Power 520s and the JS12 and JS22 blades in the first place. That’s not a smart move if you are trying to support Java applications.