Original URL: http://www.theregister.co.uk/2009/12/09/sun_sap_tests/
Sun-Oracle x86 server combo tops the SAP charts
AMD Istanbuls crank through the transactions
Sun Microsystems is the quietest server maker in the world these days, but its PR machine - which just pumps out the occasional press release like clues for those trying to figure out its plans since the Oracle gag went on in April - said today that Sun's recent top-end x64 server has smoked the competition in a popular commercial transaction processing test and another number-crunching test.
Sun pitted its Sun Fire X4640, which has eight of Advanced Micro Devices' top-end "Istanbul" six-core Opteron 8435 processors, against a bunch of other iron. These Istanbul chips clock in at 2.6 GHz, and each one inside the X4640 server is mounted on a single uniboard (a mezzanine card of sorts) with a max of 32 GB of memory (this is half of the maximum memory for the uniboard).
Rather than put four sockets on a single board and then try to glue two boards together, Sun's top-end Opteron boxes put a single processor and its memory on a single board, eight of which plug into the server backplane; the whole shebang takes up only 4U of rack space, making it one of the most compact eight-socket servers around.
The X4640 that Sun tested was equipped with Solaris 10 and the Oracle 10g (not the more current 11g) database, and it was able to support 10,000 users on the two-tier SAP sales and distribution (SD) benchmark test. The two-tier version of the test puts the application and the database that the applications pound on inside a single box, rather than breaking app and database servers into separate machines as three-tier setups do. With an average response time of 0.9 seconds and at a CPU utilization of 99 per cent, the X4640 was able to process a little more than 1.1 million line items per hour.
By way of comparison, Hewlett-Packard put its ProLiant DL785 G6 Istanbul box through the same SAP SD paces back in November, and an eight-socket, 48-core box using slightly faster 2.8 GHz Opteron 8439 SE processors but with only 128 GB of memory was able to do only 907,000 items per hour and support only 8,280 users. This HP box ran the SAP ERP code on top of Windows Server 2008 Enterprise Edition and SQL Server 2008.
It is not clear how much of the performance difference is due to the operating systems and databases and how much is due to the memory. This HP box was tuned up, by the way. Back in August, the same machine with 192 GB of memory was only able to support 7,716 simulated SD users. A four-socket ProLiant DL585 G6 with the same processors that was tested in July using 64 GB of main memory and the same Windows stack could do 4,665 SD users with sub-second response times.
IBM's four-socket Power 550 machine using 5 GHz Power6+ chips, sporting eight cores and 64 GB of memory, could handle 3,752 SD users with sub-second response time. (This machine ran IBM's own AIX 6.1 Unix variant and its DB2 9.5 database when it was tested in June.)
AMD's packing six cores in a socket compared to Big Blue's two is giving it a performance advantage, despite the lower clock speed on the x64 chips. (You can see now why IBM is cramming eight cores into the future Power7 chip and turning down the clocks. It needs to boost the oomph per chip and still stay in a thermal envelope.)
Older benchmark SD results from all the server makers are not applicable the current ones because the testers all tried to get response time as close to two seconds per transaction as possible, and now they are shooting for sub-second response. The difference between one and two seconds can nearly double throughput (which makes sense when you think about it).
The odious comparison that Sun probably doesn't want its customers to make is this: Back in July, a Sparc Enterprise T5440 server using 1.6 GHz Sparc T2 processors and equipped with 256 GB of main memory (the same as on the Opteron-based X4640) was able to handle only 4,720 users and process 516,670 transactions with sub-second response. A T5440 with 128 GB and two disks runs to $115,695 at list price from Sun.
The X4640 with 64 GB and two disks costs only $49,578 and it can do more than twice the work. Assuming the cost for memory is $525 per GB as Sun is charging on the X4640, then to get the T5440 up to 256 GB would cost another $67,200, bringing the base hardware cost for the tested SAP system up to $182,892. To get the X4640 up to 256 GB of main memory is certainly not cheap, at an additional $100,800, but the eight-socket, 48-core box would cost only $150,378 - and again, it would do more than twice the work.
Basically, the Opteron box from Sun costs $15 per user for the hardware, while the Sparc box costs just under $39 per user for the base hardware. And with the T5440 having more cores, even with the Oracle fudge factors, the database price is going to go through the roof on the T5440.
To show that it has scale on SAP workloads, Sun and Fujitsu also tested the top-end Sparc Enterprise M9000 servers with 32 and 64 sockets loaded up with its most recent 2.88 GHz Sparc64-VII processors using the same Solaris 10/Oracle 10g software stack. In October, a 32-socket, 128 core M9000 with 1 TB of main memory was able to support 17,340 SD users on the SAP test running Solaris/10g, and a 64-socket box with 256 cores and was able to nearly double its performance to 32,000 SD users with 1.1 TB of memory (yes, it is a weird configuration).
Neither IBM nor HP have put their big iron Power and Itanium boxes through the paces using the sub-second response limit and the latest SAP ERP 6.0 code base, so it is not clear how these boxes stack up.
On the SPEC OMPL2001 high performance computing benchmark, which is designed to stress the processor, memory, compilers, and OpenMP implementation like physics, weather modeling, design, and other simulations do, Sun says that X4640 was able to hit a composite peak rating of 381,017 on the suite of nine tests that comprise this benchmark.
Last year, an X4440 server with four "Shanghai" quad-core Opteron 8384s running at 2.7 GHz peaked out at 175,648. That's a pretty decent jump, but a lot of it has to do with moving from 16 to 48 cores, and given the core count you might have expected more.
As you can see from the relatively skinny set of SPEC OMPL2001 results, the Sun and Fujitsu iron offers nearly four times as much more oomph on this test - provided you want to buy an M9000 with 256 cores.
HP and IBM have not run this test on their Unix iron in years, so the comparison Sun is making with this SPEC test is not as useful as it could be. But every benchmark is another data point, and is therefore to be welcomed. ®