Power 750 Servers Running i Get SAP Benchmarks
March 15, 2010 Timothy Prickett Morgan
Comparisons are odious, the old saying goes. But, that’s what we do whenever we are buying any kind of electronics, and servers are no exception. IBM, like every other server maker with a new product, wants to show off the performance and scalability of its new Power7-based servers. But it doesn’t want to make the comparisons too easy between the i and AIX platforms, or between the i versions of the boxes and other machines like X64 servers running Windows, as has been the case for many years.
But to IBM’s credit, the company provides its rPerf and CPW benchmark ratings for each possible SKU of processor speed and count, across the entire Power-based system generations, which certainly helps with back-of-the-envelope capacity planning. No other server vendor does this, or if they do, the documents are certainly not public or widely distributed behind their backs. With a mix of CPW and rPerf performance stats and a splash of industry benchmarks, you can start making some estimates about performance across different product lines, despite the best efforts of vendors to obfuscate.
IBM used to do the Transaction Processing Council‘s TPC-C and SAP‘s Sales & Distribution online transaction processing tests on OS/400 and DB2/400, where you could make explicit comparisons of the platform to other boxes. But when the AS/400 and its iSeries successors started getting out of whack with the entry and midrange X86 and then X64 servers running Windows–the primary competitor to the AS/400 since the late 1990s–IBM chickened out and stopped doing TPC and SAP SD benchmarks. (I make my own estimates with each processor generation, and I will do so soon for the Power7-based Power 750, 770, and 780 machines.)
For the past several years, mostly to remind people that SAP’s ERP software does run on OS/400 and i and to give another proof point of system scalability, IBM has run various SAP data warehousing benchmark tests, including the obsolete Business Information Warehouse (BW) test and the more current BI Data Mart test. Up until last October, IBM’s Power5+ and Power6 platforms running the i5/OS V5R4 and i 6.1 operating systems, respectively, were the only boxes using this test. And only a handful of machines were tested. For whatever reason, IBM moved on to yet another SAP data warehousing test and is unlikely to return to this one, and here is why.
Last fall, a two-socket Xeon X5570 server (that’s 8 cores and 16 threads running at 2.93 GHz) running SUSE Linux and Oracle 10g was able to process 320,363 query navigation steps per hour, absolutely cleaning the clocks of a four-socket Power 570 (that’s 8 cores and 16 threads running at 5 GHz) running the i 6.1/DB2 combo, which could only handle 182,112 query navigation steps per hour. None of the SAP tests require pricing for the system under test to be divulged, but there is a huge disparity in price/performance. IBM could have, had it chosen to, demonstrated a lot more scalability in the Power 570 line, since this was only half of the maximum number of cores in the original Power 570. And in October 2008, when IBM slowed down the clocks with the Power6+ chips to 4.2 GHz in the Power 570 and then doubled up the cores on each processor card, it could have shown even more scalability, perhaps pushing up into the realm of 310,000 query navigation steps per hour on the SAP BI test. However, that it would take a 16-socket Power 570 to just barely match a two-socket Xeon box was, to put it bluntly, embarrassing.
No matter. IBM had already moved on to yet another SAP data warehousing benchmark with the Power7 lineup, this one called Business Intelligence-Mixed Load, or BI-MXL, test anyway.
In preparation for the February 8 launch of the Power 750 servers, IBM ran the BI-MXL test on a Power 750 machine configured with a single processor, which is eight cores and 32 threads, running at 3.3 GHz. Now, considering that the Power 750 can scale to four of these eight-core Power7 puppies, you might be asking yourself, why didn’t Big Blue put the pedal the metal and go with a full-bore Power 750 with 32 cores and 128 threads? I dunno. It makes no sense to me. As I told you a few weeks ago, when running i 6.1 in what IBM calls Power6 mode, the i 6.1 operating system only supports 32 cores and 64 threads, but supports 32 cores and 128 threads in Power7 mode; there is special support available to push the Power6 mode to 64 cores and 128 threads, apparently, and with the future and impending i 7.1, IBM will offer PTF patches on a special bid basis that allow the operating system (and presumably the integrated database) to scale to 64 cores and 256 threads.
Clearly, IBM could have pushed the Power 750 all the way, but chose not to. I don’t understand why, either.
Anyway, on the BI-MXL test, a quarter of a Power 750 was able to process 241,526 query navigation steps per hour against a database of 300 million records running i 6.1, its integrated DB2, and SAP’s NetWeaver 7.0 middleware. That Power 750 machine was configured with 128 GB of main memory and ran at 99 percent CU utilization on the test. By comparison, a Power 570 machine with four Power6+ processors running at 4.2 GHz (that’s 8 cores like the Power 750 above, but only 16 threads instead of 32) was able to handle 154,447 query navigation steps per hour on the same data warehouse. The test results on prior Power 520 and Power 550 machines, as well as on a two-processor Power 570, show linear scalability on the SAP test. This is, of course, what you would expect.
Here’s how the SAP BI-MXL workload compares to IBM’s Commercial Performance Workload (CPW):
As you can see, there is a rough correlation between SAP BI-MXL performance on the real benchmark tests and IBM’s CPW ratings.
Next week, I will take a gander at the SPECjbb2005 benchmark test results on the Power 750 and Power 770 running i, AIX, and Linux.