Sparc Enterprise Line Competes Well with RISC, Itanium Servers

Published: April 26, 2007

by Timothy Prickett Morgan

In the wake of the rollout last week of the Sparc64 VI-based Sparc Enterprise server line, which is based primarily on processor and system technology developed by Fujitsu but which is also sold under the same brand by Sun Microsystems, the two companies announced performance benchmark results on software maker SAP's mySAP ERP suite. The new machines hold up well with the competition, giving Sun and Fujitsu competitive performance for the first time in a long time.

The two companies chose the two-tier implementation of the SAP Sales and Distribution (SD) test to prove the might of the Sparc Enterprise machines. In a two-tier benchmark test, the SD workload, including both the application tier and the database tier, are running on the same physical server; companies sometimes also run in a three-tier mode, which puts the database servers on one machine and application servers on one or more other boxes.

On the SD test, Sun and Fujitsu did not yet post results on their top-end Sparc Enterprise M9000-64 box, which has 64 processors, 128 cores, and 256 threads in the processor complex. This box also supports up to 2 TB of main memory and 128 PCI-Express slots. The companies did start out with a smaller machine, as high-end box vendors often do; as they learn to tune their operating systems and databases for the SMP scalability inherent in their designs, they test larger configurations.

The M8000 server that was tested by Sun and Fujitsu had 16 of the dual-core 2.4 GHz Sparc64 VI processors, which have 6 MB of L2 cache on chip each. The server was configured with 256 GB of main memory, and brought 32 cores and 64 threads to bear on the SAP SD test; the software on the machine included Solaris 10 and Oracle 10g. On that test, the machine was able to support 7,300 SD users with an average response time of 1.98 seconds on transactions, processing 731,330 orders.

That is only slightly less performance that a Fujitsu PrimePower 2500 server with 64 of Fujitsu's prior generation of Sparc64 V chips running at 1.3 GHz could do; that machine could handle 7,550 users running Solaris 8 and Oracle 9i. And a few years earlier, a PrimePower 2000 using 675 MHz Sparc64 IV processors could support 7,800 users, but needed 128 processors to do it. Obviously, moving to multicore and multithreaded processors has helped the performance of Fujitsu's high-end Sparc servers.

The M8000 machine also compares favorably with current and prior generations of Sun gear. Earlier this month, Sun tested a 24-socket E6900 server using its dual-core UltraSparc-IV+ processors, which were just cranked up to 1.95 GHz and 2.1 GHz in limited quantities ahead of the Sparc Enterprise server announcement last week. That 24-socket box, which has 48 cores and 48 threads, was equipped with the same Solaris 10 operating system and Oracle 10g database, and it was able to support 6,160 SD users with an average response time of 1.99 seconds. That E6900 server had 96 GB of main memory, which is a lot less than the M8000 box had, but it also had 50 percent more processor sockets, which matters to customers who buy processors on a per-socket or per-core basis. If Sun did a test on a 2.1 GHz configuration of this machine, there is not enough memory in the box to push performance up by all that much. In terms of performance, the M8000 is a far better deal. And compared to earlier UltraSparc boxes from Sun in the same power class, the M8000 just blows them away. Back in 2003, Sun did a test on a top-end 104-core Sun Fire 15K server using 1.2 GHz UltraSparc-III processors and running Solaris 9 and Oracle 9i, and that machine could only support 8,000 SD users; that box also had 576 GB of main memory, which helped boost performance considerably, but which certainly was not cheap.

The interesting bit is how the M8000 stacks up against boxes in about the same performance range from IBM and Hewlett-Packard. In December, HP tested a 16-socket Integrity Superdome server using the dual-core "Montecito" Itanium 9050 processors from Intel; these processors run at 1.6 GHz and have 24 MB of L3 cache on chip, and the Integrity machine had 256 GB of main memory. On the two-tier SD test, the box was configured with Microsoft's Windows Server 2003 Datacenter Edition and SQL Server 2005, and it was able to support 5,600 users with a 1.91 second average response time. This is a lot less performance for the same number of sockets, cores, and threads as the M8000 server from Sun and Fujitsu.

In January 2006, IBM tested a System p5 570 server using eight dual-core Power5+ processors running at 2.2 GHz running AIX 5.3 and DB2 8.2. This machine, which had half the memory (128 GB) and half the threads (32) of the M8000 box support 5,520 SD users with a 1.97 second average response time. It is hard to say where a p5 595 with the same number of cores and threads as the M8000 would end up, but it is probably about twice the performance of the p5 570 that IBM tested. Using the fastest 2.3 GHz Power5+ processors and the full load of 64 cores and 128 threads, IBM's top-end p5 595 machine only handled 23,456 users when it was tested last July. That is four times the number of threads, a slightly higher clock speed, and a lot more memory to do about 4.25 times the work. So if p5 performance scales more or less on this line, twice the cores should do twice the work on the SD test.

HP, of course, is the king of the SD test. Last December, as it was launching HP-UX 11i v3, the company showed a 64-socket Integrity Superdome using the Montecito chips with 512 GB of main memory and running Oracle 10g, and this box supported 30,000 SD users.

The lesson here is that each server line scales differently as processors, memory, and I/O are added. So extrapolating from the performance of the initial M8000 box to predict that the M9000-64 will soon be the king of the SD hill is not wise. It could turn out that way if the Sparc Enterprise line scales well as processors are added. You can bet that if this turns out to be true, Sun and Fujitsu will be bragging about it soon.

The other thing to remember is that one benchmark is not a good gauge of performance. Sun and Fujitsu need to test the Sparc Enterprise machines on other benchmarks to test their mettle. The sooner they do this, the better. And after they do that, they have to prove to the enterprise customers they want to sell these boxes to that there is a future in this Sparc Enterprise product line--Fujitsu has committed to moving to a quad-core chip in the future in these "Jupiter" frames--and that upgrades or box swaps will be available to Sun's future "Supernova" servers and their "Rock" 16-core Sparc processors, which are due next year.

RELATED STORIES

Fujitsu, Sun Deliver Joint Sparc Enterprise Server Line

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

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