HP Tops SAP S&D Benchmark Test with Itanium and HP-UX Combo

Published: February 15, 2007

by Timothy Prickett Morgan

Hewlett-Packard and IBM continue to slug it out to see who offers the best high-end performance supporting SAP's mySAP ERP stack, and as of the latest results on the SAP Sales and Distribution (SD) benchmark, HP is winning with its Integrity servers and HP-UX operating system.

The combination of the new HP-UX 11i v3 operating system, which was announced this week, and the "Arches" chipset inside the Integrity Superdome servers, and Intel's dual-core "Montecito" Itanium 9000 processors have put HP over the top on the two-tier SD benchmark test. In a two-tier mode, the SD workload, including both the application tier and the database tier, are running on the same physical server.

On HP's latest SD benchmark, an Integrity Superdome with 64 of the top-end, dual-core Itanium 9050 processors--these run at 1.6 GHz and have 12 MB of L3 cache per core plus 2.5 MB of L2 cache per core--was configured with 512 GB of main memory. Because these "Montecito" Itanium 9000 series processors have HyperThreading support, this box has a total of 256 threads, which means it can handle database workloads (which eat threads for lunch) very efficiently. This server was configured with HP-UX 11i v3, which HP says can by itself yield 20 percent to 50 percent performance improvements; HP has said in the past that changes in the Arches chipset, including larger main memories and better memory and I/O subsystem design, could by itself boost performance using single-core "Madison" Itanium 2 processors by as much as 30 percent. This server was configured with Oracle's 10g database, and it ran at 93 percent of maximum CPU utilization to support 30,000 simulated SD users.

HP did not run the two-tier SD test on an HP-UX 11i v2 setup using the prior Madison Itanium 2s and its own prior "Pinnacles" chipset, but it did run a test using Windows Server 2003 Server Datacenter Edition and SQL Server 2005 in June 2006, and that machine, with 64 1.6 GHz Madison chips and 1 TB of main memory could only support 8,575 SD users. Historically, HP-UX plus Oracle databases have run about 20 percent faster than the Windows stack on Integrity iron, so hypothetically the same box could have supported 10,300 SD users running HP-UX 11i v2 and Oracle 9i. That implies that HP has been able to almost triple performance by moving to the new chipset, new HP-UX, and new Intel chips. This is consistent with the performance claims HP has made for the Montecito chips, HP-UX improvements, and the Arches chipset. (If you add up all those effects, you get a machine that can do a little more than 32,000 SD users if you assume a 10,300 SD user baseline for the earlier machine.)

IBM has the second highest result on the two-tier SD test, with its System p5 595 with 64 Power5+ cores and 128 threads running at 2.3 GHz supporting 23,456 users at 99 percent CPU utilization. That machine was configured with 512 GB of main memory and IBM's own AIX 5.3 Unix operating system and DB2 9 database. This machine, which was tested last June, only delivered a 17 percent performance increase over the prior Power5-based System p5, which was tested in October 2004 using 64 1.9 GHz Power5 cores, 512 GB of main memory, AIX 5.3, and DB2 8.2. That machine supported 20,000 SD users and ran at 97 percent CPU utilization.

A Fujitsu PrimePower 2500 server with 128 Sparc64 V single core processors running at 2.08 GHz, was able to support 21,000 SD users. This machine was tested way back in March 2005, and ran Solaris 9 and Oracle 9i. That server offered 62 percent more performance than the 128-way PrimePower 2500 box that was tested in April 2003 using 1.3 GHz Sparc64 V processors and running Solaris 8 and Oracle 9i.

Sun Microsystems' best effort to date on the two-tier SD benchmark test was way back in July 2004, when it tested a Sun Fire 25000 server with 72 of its dual-core "Jaguar" UltraSparc-IV processors running Solaris 9 and Oracle 9i. This was Sun's first generation of dual-core Sparc chips, not the "Panther" UltraSparc-IV+ chips, which had about double the raw performance because of the addition of an integrated L2 cache and other tweaks to the chip. That Sun Fire 25000 box used 1.2 GHz UltraSparc-IV processors and had 576 GB of main memory; it was able to support 10,175 SD users at 98 percent CPU utilization. If Sun did a test on its top-end server using most recent 1.8 GHz UltraSparc-IV+ chips, it should be able to handle about 22,000 SD users on the two-tier test, provided the machine is not memory or I/O bound.

That is nowhere near what HP just delivered and is within spitting distance of the best that IBM has been able to do. If IBM can double again with the Power6 servers this year, it will be in the range of 45,000 SD users. Sun has said very generally that the future "Jupiter" servers, which are essentially kickers to the PrimePower 2500s, which are mostly developed by Fujitsu, and which Sun will resell as part of its Advanced Product Line partnership with Fujitsu, would offer about 1.5 times the performance of the current Sun Fire product line using the Panther processors. That should put Sun at around 33,000 SD users on the two-tier test, which means don't hold your breath waiting for Sun or Fujitsu to do this test. If these numbers pan out, neither company will want a direct comparison to Montecito servers from HP or Power6 servers from HP.

What seems clear is that HP has leapfrogged ahead of IBM finally in terms of performance on the SD test (which may be a better indicator of relative performance than the TPC-C test that IBM prefers because it does so well on it), and IBM will leap ahead of HP later this year. To be followed by HP's future Integrity servers based on "Tukwila" quad-core Itaniums in 2008 and maybe, just maybe, Sun's "Rock" processors and "Supernova" servers in 2008 besting them all.

Sun is claiming that Rock will deliver 16 times the performance of the current UltraSparc-IV+ Sun Fire boxes and 10 times the performance of the Jupiter servers on "throughput workloads." It is hard to say if that means this Supernova box can really support 330,000 to 350,000 SD users on a two-tier benchmark test. This seems a bit hard to swallow. But if Sun can do that, Rock will have been a great thing to bet on.

                     Post this story to del.icio.us
               Post this story to Digg
    Post this story to Slashdot