One Power 750 Matches Two Xeon Servers On SAP BW Test

August 26, 2013 Timothy Prickett Morgan

For many years now, IBM has been very keen on using various data warehousing benchmark tests based on applications from German software giant SAP to show the relative scalability of Power Systems running OS/400, i5/OS, and IBM i, and to make comparisons to other systems. Now, IBM has perhaps won a throwdown with Hewlett-Packard.

Last year, when HP was trying to show off the performance of its ProLiant DL580 G7 servers running SAP code and using the HANA in-memory database from SAP to support its Business Warehouse data warehouse, the company chose the Business Warehouse Enhanced Mixed Load benchmark, or BW-EML for short, as a metric. (You can read all about the test here.) This is a new benchmark test, and therefore cannot be directly compared to earlier SAP tests such as the Business Intelligence Mixed Load (BI-MXL) or Data Mart (BI-D) tests that have been run on earlier generations of Power-based systems running OS/400, i5/OS, and IBM i.

The interesting bit about the BW-EML test, explains Ron Schmerbauch, who is an IBM i and DB2 performance guru at the Rochester Labs, is that it has both a load component and a query component running at the same time, like earlier tests. “But the data is changing as you are doing your queries, so it is very hard to tune it.” This test, says Schmerbauch, was specifically designed to show off the HANA in-memory database, but as you can see from the results below, a regular DB2 for i database running on a mix of flash and disk drives held its own against a pair of Xeon E7 servers, one acting as a HANA in-memory database engine and the other as an application server:

The Power 750+ versus two HP ProLiant DL580 servers.
(Click graphic to enlarge.)

Schmerbauch says that IBM, which did its tests more than a year after HP did, went with a 32-core Power 750 machine for its tests because the typical IBM i shop has maybe four or eight, and sometimes 16, cores, and that way a 32-core machine would show some headroom over the typical customer’s needs. That Power 750+ machine had four Power7+ processors, each with eight cores and 32 threads, running at 4.06 GHz. The system was configured with 512 GB of main memory. While the SAP benchmark report doesn’t show this, IBM configured that Power 750+ with two dozen disk drives and a dozen SSDs to boost the performance of IBM i 7.1 and its integrated DB2 for i database.

The database IBM tested had 500 million records, which could fit into a machine with 32 cores; I am guessing here, but it would probably take a 64-core machine to do 1 billion records as HP tested with its pair of ProLiant DL580 G7 servers. The database was much larger than 512 GB, according to Schmerbauch, and would not fit in main memory.

In any event, the Power 750+ was configured with IBM i 7.1, DB2 for i 7.1, and SAP’s NetWeaver 7.30 application server, and it was able to process 66,900 ad-hoc navigation steps per hour while ingesting new data and updating that database. Both the database and the application server ran on the same physical machine without logical partitioning–something you can do on IBM i thanks to the operating system’s subsystems architecture. The system ran at 96 percent of maximum CPU capacity. About half of the CPU capacity in the Power 750+ was allocated to the database and about half to the application server, says Schmerbauch.

The pair of HP machines had one DL580 G7 configured with 512 GB of main memory with four of Intel‘s 10-core Xeon E7-4870 processors. These spin at 2.4 GHz. Each node therefore had 40 cores and 80 threads. The servers ran SUSE Linux Enterprise Server 11. One server ran the HANA 1.0 in-memory database, and the other ran NetWeaver 7.30 and the benchmark code. The database server did not have any disks in it, but did have some flash drives for database and app logging. The HANA database has data compression and columnar data store, which is why it can get a fat database into the 512 GB of main memory.

With the ProLiant DL580 G7 database server running at 88 percent of CPU and the app server linked to it running at 28 percent of CPU (which suggests some bad sizing if you are worried about wasting money), the cluster of Xeon E7 machines could process 65,990 ad-hoc navigation steps in an hour while still ingesting randomized data.

What is not clear from these benchmark tests is what these two machines cost. SAP should require detailed configuration information and hardware and systems software pricing as part of its test reports. I can figure out Power 750+ server pricing and also get HP server pricing, but I have no idea–yet–what the HANA database costs. But I aim to find out and will report back with bang for the buck metrics as soon as I do.