Now, IBM Corp has leapfrogged HP with test results on its new 32-way Regatta-H server using the new 1.7 GHz Power4+ processors that IBM announced last week. Both IBM and HP are gearing up to significantly improve upon the results they just posted, too.

The 32-way pSeries 690 server was configured with AIX 5L V5.2 and IBM’s DB2 Universal Database 8.1 software. This is significant. IBM has not shown TPC-C test results using DB2 to date. Database maker Oracle Corp has been making fun of IBM because it uses Oracle to run its Unix server benchmarks – and making money because of the performance advantages that the Oracle database seems to have based on the benchmark tests.

There is a tremendous amount of pressure for IBM to demonstrate that DB2 is a viable alternative to Oracle, and this benchmark is IBM’s first step in that direction. However, given the way that benchmarks work, when HP breaks through the 700,000 TPM barrier within the next couple of months – if not sooner – on the Pinnacles version of the Superdome that it showed performance results for at the Windows Server 2003 launch, IBM will have to put its preferences aside and show a better benchmark result using the Oracle database.

There is headroom for this. The benchmark tests results IBM showed last week on the TPC-C test yielded about 59% more throughout than its best results on the Regatta machine using 1.3 GHz Power4 processors; but IBM’s relative performance (rPerf) metrics, which are based on TPC-C simulations running Oracle, say that there is another 6% or more of performance in the new Regatta box than the DB2 benchmark shows.

That Regatta machine, configured with 1.7 GHz processors (with a 1.5 MB shared, on-chip L2 cache), 512MB of L3 cache, 512GB of main memory and 43TB of IBM’s 7103 SSA disk arrays (which are very fast even though they are not exactly a standard product), was able to crank through 680,613 transactions per minute (TPM).

The pSeries 690 server cost $3.27m, with $1.38m going for main memory alone and $1.4m going for processors. The SSA disk subsystems cost a whopping $6.86m, and if IBM had chosen its Shark ESS arrays, it might have lowered the cost somewhat, but the performance might have not been as good as that provided by the SSA arrays. AIX and DB2 for the server cost $632,725. With application servers and three years of maintenance thrown in, the whole TPC-C configuration had a list price of $14.57m, but after a stunning 48% discount, IBM cut the price to $7.57m.

When you do the math, IBM was able to show a price/performance of $11.13 per TPM after the discounts, which is not as good as HP is delivering with its Superdome, but then again, it’s on a box with half as many processors. The processor count is important for application pricing. And on single threaded jobs, like old monolithic financial batch applications, the speed of a single processor is often as important than aggregate OLTP throughput in a box.

The evening before Windows Server 2003 was launched and HP was getting ready its benchmark tests on the Superdomes, NEC showed off its benchmark results on the 32-way AzuzA Express5800/1320Xc using the future Madisons. This server was equipped with a whopping 512GB of main memory and a stunning 41TB of disk capacity.

The core server cost $2.2m (including the cost of the 64-bit version of Windows Server 2003 Datacenter Edition), with disk subsystems costing another $2.9m. Including a 64-bit SQL Server 2000 license for all the processors and three years of maintenance, the NEC setup cost $7.2m, and an NEC large systems discount of 20% dropped the total price to just under $6m. With 514,035 TPM of throughput, that gave the Madison-based AzuzA machine a price/performance of $11.50 per TPM. The Superdome that HP tested most recently had 512GB of main memory. The central electronics complex of this Superdome cost around $2.2m, the main memory another $3.3m, and $2m of the core server coming from the 34.6TB of disk storage and related peripherals. Operating systems and database costs came to $1.3m on this Superdome. The whole enchilada, including application servers and three years of maintenance, cost $10.5m. But then HP invoked a large systems discount of 38.5%, and dropped the price tag down to $6.5m, yielding a price/performance of $9.80 per TPM.

So where is the pSeries 690 machine running Oracle going to end up? When you look at rPerf metrics and extrapolate them, the Regatta machine should be able to break 720,000 TPM. If IBM moved to Shark arrays – provided the performance penalties are not too great – it should be able to get below $10 per TPM, and maybe even under HP’s most recent numbers for the Superdomes. HP is, of course, trying to beat these expected IBM numbers. The game’s afoot.

Source: Computerwire