Big iron makers test their metal on SAP
Rites of passage in the server room
The server makers of the world have been busy little benchmarketeers in the past few weeks and have released a slew of test results that show the prowess their machines have supporting SAP's ERP software.
Even IT shops that don't run SAP software take a hard look at the Sales & Distribution (SD) benchmark tests because a wide variety of machines have been put through the paces to gauge their performance relative to other boxes. That's because the SD test, like the processor and Java benchmarks that come out of the Systems Performance Evaluation Corp (SPEC) and the TPC-C and TPC-H benchmarks from the Transaction Processing Council, is a rite of passage for a server. With a mix of benchmarks, you get a sense of what a box can - and cannot - do.
This week, IBM was going on about how a Power 795 with 16 sockets and sporting 128 cores running at 4GHz was the top of the SD charts now. (You can see the SAP SD test results here.) But what seems as important was the fact that the Power 795 that IBM tested was only half of the largest machine that IBM says it can actually deliver.
The only plausible explanation for Big Blue not running the SD test on a fully loaded 32-socket, 256-core Power 795 is that the scalability of the hardware, the AIX 7.1 operating system, or the DB2 9.7 database - or perhaps all three - is a whole lot less than a factor of two more than the box that IBM did test. Of course, the only way to know that is to have run the test and to have decided to spike the results and only show results on half the box, where performance per core is understandably better.
On the SD test that IBM ran on the Power 795 (or rather, what we might call a Power 747.5 to be a smart-alec), the machine was configured with 2TB of memory, which was only a quarter of its maximum. The Power7 chips used in the machine have four threads per core, so 512 threads were supporting the DB2 9.7 database that drove the SAP SD test, which was based on SAP ERP 6.0 with Enhancement Pack 4.
With 99 per cent CPU utilization, the Power 795 was able to process 7.69 million items per hour and so 23.1 million dialog steps per hour with an average dialog response time of 0.93 seconds. This works out to supporting 70,032 simulated SD users.
That's about twice the performance of IBM's prior generation Power 595 machine running full out with 32 sockets and 64 cores with a mere 512GB of main memory, which was able to handle 35,400 SD users when it was tested several years ago. IBM has said that the Power 795s offer roughly four times the oomph of the Power 595s when fully loaded, so this is in that ballpark if the jump from 16 to 32 sockets on the Power 795 yields something close to 2X more bang.
The performance of IBM's Power 780, which is the younger brother to the Power 795, is consistent as well, with an eight-socket, 64 core configuration using 3.8GHz processors and 1TB of main memory being able to support 37,000 SD users running AIX 6.1 and DB2 9.7.
IBM is, of course, positioning the Power 795 against the Sparc Enterprise M9000 box designed and built by Fujitsu and sold by both Oracle and Fujitsu as well as x64 clusters running Oracle's Real Application Clusters (RAC) database clustering.
Last November, before the Oracle acquisition of Sun Microsystems was done, Sun and Fujitsu put the Sparc Enterprise M9000 through the SD paces. That machine was tested using Fujitsu's quad-core Sparc64-VII processors running at 2.88GHz, with 1.1TB of main memory and 64 processors and the same 512 threads that IBM is bragging about.
The Fujitsu/Sun box ran Solaris 10 and Oracle 10g and with sub-second response time Fujitsu was able to support 32,000 SD users on this configuration, processing 3.5 million order line items per hour and doing 10.5 million dialog steps per hour. The machine was ramped up to 94 per cent CPU utilization using the latest 2.88GHz Sparc64-VII+ chips using the same SAP software stack.
Hewlett-Packard has not tested its HP-UX Unix boxes on the SD test in a very long time, but last week the company put out a benchmark result for its eight-socket ProLiant DL980 G7 server using Intel's eight-core X7560 processors running at 2.26 GHz.
That server was set up using 512GB of main memory running Windows Server 2008 R2 Datacenter Edition and SQL Server 2008, and it could chunk through a little more than two million SD order line items per hour and 6.1 million dialog steps per hour, simulating 18,635 SD users. That ProLiant DL980 G7 server was humming along at 97 per cent of CPU utilization, and it costs a lot less money than either an Oracle/Fujitsu Sparc Enterprise M9000 or an IBM Power 780 or 795.
IBM this week also put its System x3850 X5, also an eight-socket box based on the eight-core Xeon X7560 processors, through the SD course.
Running Windows Server 2008 R2 Enterprise Edition and its own DB2 9.7 database on a machine with 512GB of main memory, IBM was able to support 19,700 SD users on this machine at 99 per cent utilization, besting HP's ProLiant DL980 G7 by a tad.
Fujitsu's PrimeQuest 1800E, also an eight-socket Xeon X7560 machine, was tested last week, and was able to support 18,310 SD users running the same Windows Server 2008 R2 Datacenter Edition and SQL Server 2008 stack HP used. The Fujitsu machine did a little less work, but was only using 97 per cent of CPU and had an average response time of 0.8 seconds. Back in August, NEC tested a similar eight-socket Xeon 7500 box, with the catchy name of the Express5800 Model A1080a, that ran Windows Server 2008 R2 Enterprise Edition and SQL Server 2008 and that was able to support 18,185 SD users with 96 per cent of CPU dedicated to the task.
Oracle, as you might expect, would no doubt counter all this talk about big iron that the smart thing to do is to use clusters of even cheaper four-socket x64 boxes and slap RAC on top of the cluster to scale up the performance while also providing some resilience to the SAP applications.
In June, Oracle ran a parallel implementation of the SAP SD test on a two-node cluster comprised of four-socket Sun Fire X4470 servers that are, in essence, half of the Xeon 7500 boxes tested by the other vendors above.
Using the same X7560 processors running at 2.26GHz with 256GB per node of main memory - and equipped with the Solaris 10 operating system and the 11g database with RAC extensions - Oracle was able to handle 21,000 SD users at 97 per cent of CPU on the nodes and a response time of 0.93 seconds on the dialog steps in the SD test. And in late September, to show off the near-linear scalability of the RAC clustering, Oracle glued together four of these Sun Fire X4470 nodes and was able to support 40,000 end users.
The word on the street is that RAC peters out after eight nodes, but it can probably match the performance of IBM's half Power 795 with an eight-node cluster, and if Oracle used its eight-socket server, an eight-node cluster might keep pace with a full Power 795. IBM might actually have to prove that its PureScale database clustering for Power Systems machines and its AIX and DB2 stack can keep pace with RAC, to the detriment perhaps of its Power 780 and 795 server sales.
Any way you cut it, those RAC and PureScale clusters are big iron. And that perhaps explains why Oracle seems to be in no hurry to complete a server roadmap and distribution deal based on future eight-core Sparc64-VIII processors from Fujitsu. ®