IBM Tests eServer i5s on SAP Benchmark

by Timothy Prickett Morgan

The new eServer i5s are rolling off the production lines, and now IBM has to start cranking up the marketing machine to show that the "Squadron" Power5-based servers combined with i5/OS V5R3 (which we know as OS/400) is a compelling hardware and software platform on which to build applications affordably. One way to do this is to run benchmark tests, something IBM did a lot during the AS/400 era.

IBM announced last week that it has run the two-tier SAP sales and distribution benchmark test on the i5 Model 520 to show how it compares with other two-way servers in the market. A decade ago, when SAP was just getting rolling as a big ERP software supplier, the idea that you could even load SAP R/3 on a two-way entry machine would have seemed ludicrous. Moore's Law has done much to pack a lot of computing into small cabinets, and now a machine like an i5 Model 520 can do quite a bit of work.

The SD benchmark test simulates users logging into the sales module of the SAP R/3 and mySAP ERP suite and running transactions, and it has become one of the de facto performance tests used in the server industry. It comes in two flavors: a two-tiered implementation, which puts the SAP application servers and databases on the same physical machine, and a three-tiered setup that puts a database server (or a cluster of database servers) in the center of a network with many separate physical application servers. Back in the late 1990s, the only way to support a lot of users was with a three-tiered SAP implementation, but with the advent of more powerful processors, vendors have been using the two-tiered test more frequently because it reflects how many midrange shops really do run applications like SAP. What the SD benchmark does not have is a price metric, so I can only talk about performance at this point.

According to a benchmark test run a few weeks ago, a Model 520 with two 1.65 GHz Power5 processors, 1.9 MB of L2 cache, 36 MB of L3 cache, 8 GB of main memory, and 1.3 TB of disk capacity was able to support 433 SD users with an average dialog response time of 1.98 seconds. (A dialog is SAP's way of saying a portion of a transaction, like logging in, then getting a customer account number and looking up an order. That is three dialog steps, but it may be only a single transaction, to some way of thinking. SAP counts all the steps rather than arguing about what is or is not a transaction.) The i5 Model 520 that IBM tested was running i5/OS V5R3 and its integrated DB2/400 database. That machine, by the way, is the most powerful two-way box IBM sells, whether it runs OS/400 or another operating system. But other two-way machines are right in there, as far as the SD test is concerned.

IBM also recently tested a BladeCenter HS20 blade server equipped with two of Intel's 3.2 GHz Xeon processors, 512 KB of L2 cache, 2 MB of L3 cache, 6 GB of main memory, and 792 GB of disk capacity. This IBM blade server was set up with Microsoft's Windows Server 2003 Enterprise Edition and IBM's DB2 UDB 8.1 database for Windows. When IBM ran the SD test on it, it could support 428 SD users with an average dialog response time of 1.99 seconds. This is very nearly the same result as with the i5 Model 520.

A nearly identically configured X86-based BX600 blade server from Fujitsu-Siemens, running Novell's SuSE Linux Enterprise Server 8 and a variant of the open source MySQL database co-developed by SAP and MySQL, could only handle 385 SAP SD users with an average dialog response time of 1.95 seconds. This particular machine only had 216 GB of disk, so it may have been I/O-constrained.

Sun Microsystems has also recently tested its two-way Sun Fire V20z Opteron-based server on the SD test. Sun put two of the 2.2 GHz Opteron 248 processors in the V20z, which has 1 MB of L2 cache on the chip. It added 16 GB of main memory to the box and 948 GB of disk capacity. This machine was, ironically, not equipped with Sun's own Solaris 9 operating system, which can run in 32-bit mode on the 64-bit Opterons, but opted instead for a 64-bit version of SuSE 8 and the Oracle 9i database. This machine could support 410 SD users with an average dialog response time of 1.84 seconds. If Sun had done a bit more tuning to stretch out the response times a little, it may be been able to throw on a few more users.

While IBM and its partners can use the lower-cost-of-ownership argument as an ace up the sleeve, clearly the performance of the new i5 Model 520 is as good or better than the alternatives in the entry market, as far as the SD test goes. It will take a lot more tests on a lot more benchmarks to prove this to new customers and to excite existing customers to upgrade their machines, of course. But this is an excellent first step for performance-conscious customers.