This AzuzA machine, called the NEC TX7, was using 1GHz McKinley Itanium chips and was running the 64-bit versions of Windows Server 2003 Datacenter Edition and SQL Server 2000 Enterprise Edition. It posted TPC-C results that were consistent with the expected performance of Itanium servers. A 16-way AzuzA machine using 833MHz Merced Itaniums back in 2001 was able to handle about 140,000 TPM (although there never were tests to prove it), we were told by HP and NEC. The doubling the number processors should have added another 50% or so in performance and moving from Merced to McKinley should have added another 60% or so of throughput. So when NEC last week announced at the Spring IDF last week that it had been able to post TPC-C results of 433,108 TPM on a 32-way McKinley box running the same Windows software – an increase of 40% on seemingly the same iron – more than a few heads in the server business were getting scratched.
Exactly how NEC did it is unclear, but what is clear is that these results firmly establish the Itanium 2 server as a member of the upper echelon of the server market, right next to the biggest, baddest RISC/Unix servers that IBM Corp, Hewlett Packard Co, Sun Microsystems Inc, and Fujitsu Siemens can bring out. IBM doubled the main memory to 512GB on its pSeries 690 Regatta-H 32-way server, which uses its 1.3GHz Power4 processors, to post a result of 427,761 TPM on the TPC-C test back in early January. That was an improvement of only about 6% compared to the results on the same machine running AIX and Oracle, which handled 403,255 TPM. IBM seems to have done that test merely to squeak by HP’s much-improved benchmark results for its Superdome machines. Last August, HP cranked through 423,414 TPM on a 64-way Superdome machine running HP-UX and Oracle. Early HP test results were dramatically lower for the Superdomes, and HP eventually figured out that someone was sabotaging its tests. (That’s not a joke, but rather a pending lawsuit.) HP’s January 2001 tests showed the Superdomes using an earlier vintage of PA-RISC processors could only handle just under 200,000 TPM, but in theory they should have been able to hit around 250,000 TPM at the time.
Maybe NEC has had a saboteur in the ranks, and has rooted him or her out? Or, maybe NEC has changed to a completely different iteration of the AzuzA server, or a different chipset entirely. It could turn out that NEC, which has long relationships with HP, has picked up HP’s sx1000 Pinnacles chipset. That would explain the June 30, 2003 availability date on the box. HP was selling AzuzA machines for a while as it was waiting to get Pinnacles out the door, and one hand always washes the other in the computer business. Neither HP nor NEC were able to return our calls at press time about this.
In any event, the NEC Express5800/1320Xc server with 32 Itanium 2 processors running at 1GHz used the same L3 cache sizes – 3MB. However, the updated configuration used 15K RPM disk drives and improved S2300 RAID arrays, with 1GB of cache storage instead of the 512MB in the S2100 tests used with the TX7. This could account for a substantial improvement in I/O bandwidth, and if the earlier TX7 server was I/O constrained, that might be enough to explain the big leap in performance. Moreover, the 64-bit versions of Windows and SQL Server are not exactly fully cooked yet, so there could have been a lot of work done here, too. The latest 32-way NEC box was configured with faster 2 Gigabit switches linking client machines to the TPC-C database servers as well, which probably helped a lot. The NEC Express5800 was configured with 512GB of main memory and 40.9TB of disk capacity. The list price of the entire client/server setup came to $5.6m and three years’ of maintenance on the whole shebang came to $1.2m. After discounts of 18%, the street price of configuration came back down to $5.6m, yielding a price/performance of $13 per TPM.
This is excellent price/performance, even by midrange Unix standards, and shaves a few bucks per TPM over the Unix system results in the same price class. Fujitsu Siemens, with its 128-way Kaiser PrimePower 2000 servers, is still the reigning champion in the TPC-C race with a 455,818 TPM on its 128-way Solaris-compatible machine using 563MHz Sparc64-GP clones of Sun’s UltraSparc chips. But that PrimePower machine had four times as many processors as the NEC Itanium box, and it was also was running an obscure database called SymfoWare. Fujitsu Siemens is getting ready to ship 1.3GHz Sparc64-GP processors in these big boxes, which should be able to at least double the performance compared to the older PrimePower box. What the PrimePower machines can do running Oracle is completely unknown, and Fujitsu Siemens has had no compunction to clarify that.
What is safe to proclaim, unless the TPC-C reviews the NEC results and pulls them, is that an Itanium 2 server is hands-down the fastest 32-way in the West – and East.
Source: Computerwire
