by Timothy Prickett Morgan
The performance statistics of the new Power Systems machines running the i 6.1 operating system (formerly known as i5/OS V6R1) are starting to come out, helping customers who are looking at buying one of these new boxes and moving to the new operating system try to figure out what throughput the new systems have. Performance benchmarks on real-world workloads have been pretty thin for the OS/400 and i5/OS platform for a number of years, and thus far IBM is sticking to the relatively small number of benchmarks plus its own Commercial Performance Workload (CPW) ratings to help customers gauge relative performance.
This stands in stark contrast to the variety and number of benchmarks that IBM has run on its Power-based AIX and Linux boxes as well as on X64 machinery running Windows and Linux. And while I am hopeful that the Power Systems merger of the System i and System p line might result in the same suite of commercial benchmark tests being run on AIX, Linux, and i software and their relevant DB2 database variants, I would not hold my breath waiting for such events to occur--despite the obvious logic of such an approach to benchmarking. In our household, we teach our children a phrase that dates from the 1440s and that was memorialized by Cervantes, Shakespeare, Christopher Marlowe, and John Donne in various forms: "Comparisons are odious." IBM and its sibling server divisions have long-since embraced this phrase, and therefore made it difficult (almost impossible) to compare the performance of its various server platforms. If servers were children, this would make sense, since people have feelings and they develop their skills and discover their innate abilities at different rates. But when you buy a server, it doesn't have feelings and comparisons are necessary. It drives innovation and competition, in fact.
Anyway, even though there is no longer a System i division, the remnants of that organization certainly do live on inside the Power Systems server manufacturing and Business Systems marketing divisions, and that is why IBM has once again run an SAP data warehousing benchmark test on a Power-based platform running an OS/400-style operating system. As I have commented each time IBM has published SAP BI-Data Mart benchmarks since April 2007, i5/OS and OS/400 shops tend to put data warehouses and data marts onto Windows-based servers running Microsoft's SQL Server, which has basic OLAP functionality built in, although some OS/400 platform customers certainly do use data marts available from New Generation Software. My point in criticizing this is not that IBM should not run this test, but that the BI Data Mart benchmark seems to be exactly the wrong test to run on i5/OS if you want to show its ERP muscle. It would be much more useful to also run the more standard SAP Sales and Distribution ERP benchmark test, which is a de facto system benchmark. By not doing the SAP SD test, IBM looks like it doesn't understand its own customers or that it has something to hide. Maybe both. But, as I said last year, any benchmark test result on any AS/400, iSeries, or System i machine is to be welcomed because more data is better.
IBM ran the SAP BI Data Mart test on the Power Systems 520 server, which is going to be the volume machine in what would have been called the System i space until April. The Power System 520 was a so-called Express i Edition, and it was configured with a dual-core Power6 processor running at 4.2 GHz, which has 128 KB of L1 cache and 4 MB of L2 cache per core; this machine had 32 GB of main memory and was running the new i 6.1 operating system and made use of the integrated DB2 for i 6.1 database management system. When all the software, including SAP's data mart code and its NetWeaver 7.0 middleware, was loaded up on the box, it could handle 40,445 query navigation steps per hour with the processors running at 97 percent utilization. (Interestingly, another IBM document I found shows this same machine handling 41,297 query steps per hour, but the official SAP document has the lower number.)
Last September IBM ran the BI-DM test on a System i 520 with two 1.9 GHz Power5+ cores activated with 16 GB of main memory and using i5/OS V5R4, DB2/400, and NetWeaver 7.0. At the same 97 percent CPU utilization, this System i 520 box was able to handle 26,224 query navigation steps per hour. So the jump from Power5+ running V5R4 to Power6 running i 6.1 resulted in a 54.2 percent increase in throughput.
Earlier in June 2007, IBM ran the BI Data Mart test on a System i 570 server based on the Power6 chip and running i5/OS V5R4M5. This machine had four 4.7 Power6 cores that had the same L1 and L2 caches but also added 32 MB of L3 cache for every dual-core Power6 chip. IBM crammed 64 GB of main memory on this box, and guess what? With almost precisely twice as much iron as the new Power 520 machine had, it did a little more than twice the work: 92,716 query navigation steps per hour at 98 percent CPU utilization, to be precise.
I'll be gathering up all the feeds and speeds for the Power Systems to help you get a handle on performance and get started on your capacity planning. Sit tight.