After several months of tuning and tweaking in its benchmark labs, IBM has finally released a TPC-C online transaction processing benchmark test result for its new Power7-based servers.
Rather than presenting results on a number of heavily configured machines, IBM has packed a lot of flash storage into a relatively skinny Power 780 server to show that this enterprise-class server (somewhere between midrange and high-end) can deliver big throughput but do so at midrange prices.
That flash storage, which is identified only as Feature Code 4637 "3.5 TB SSD Package" in the official TPC-C report, does not appear to be a currently shipping product - which could be why the system under test presumably is not going to be available until October 13, 2010. The Power 780 server tested started shipping on March 16, and the AIX 6.1 operating system and DB2 V9.5 database running on the box have been shipping for years.
While it's good of IBM to show the benefits of flash storage on database performance, it would have been interesting to see the same machine tested with normal disks and then with flash, with their relative performance and costs. And it would have been useful for IBM to have done some tests showing how the midrange Power 750 did on the test and then others showing how the multi-node Power 770 servers (which can scale from one to four nodes in an SMP setup) and their Power 780 brethren (which have half the cores in the nodes turned off so the clock speeds can be boosted in TurboCore mode) perform as they scale.
Instead, IBM took a single Power 780 node and configured it with two eight-core Power7 processors, with half of the cores deactivated so the clock speed can be bumped from 3.86GHz to 4.14GHz. That's a single node's worth of computing and a mere quarter of the Power 780's full processing capacity, but IBM had all four nodes there just the same so it could load up 512GB of main memory.
The Power 770 and 780 are not going to be able to support their full 2TB of main memory until Big Blue delivers a fatter buffered DDR3 memory card in November, and it seems likely that only when this memory is available with the Power 770 and Power 780 be able to offer balanced performance with their respective 64 or 32 Power7 cores.
That Power 780 was configured with two 300GB 15K RPM disk drives per server node as well as with a DS3400 entry disk array and thee EXP3000 disk expansion units, with a total of 36 disk drives and 28.7TB of capacity. The flash storage was clearly used to cope with most of the I/O necessary to support the DB2 databases behind the TPC-C test.
The Power 780 system tested cost $714,029 at list price, with the flash drives being $294,342 of that. The external disk storage costing $56,289. Three years of maintenance on this iron runs $116,919. AIX, DB2, C compilers, and virtualization management software cost $410,000, with $176,000 of maintenance and support costs. (Almost all of the software costs came from DB2.)
The client servers that drive the TPC-C test - IBM System x boxes running Windows - accounted for another couple ten grand. The Power 780 system cost $1.22m in hardware and software, plus $297,660 for maintenance over three years, which works out to $1.51m. But IBM slapped on a 45.5 per cent discount - which it tells the TPC is the typical discount a customer should see when buying such a system - dropping the price tag to $825,004 for the whole shebang. The box was able to do 1.2 million transactions per minute, which is good performance for a box with eight cores in it to be sure, which works out to 69 cents per TPM.
This is the same performance level and price/performance that current crop of x64 servers that were just announced can deliver. For example, HP has put a ProLiant DL385 G7 server using Advanced Micro Devices' twelve-core Opteron 6174s through the TPC-C paces. That two-socket box, when equipped with 256GB of main memory and a mix of SAS disk and SATA flash drive storage (76 disk arms and 125 SSDs) plus Microsoft's Windows Server 2008 R2 Enterprise Edition and SQL Server 2005 SP3 (both 64-bit versions) was able to crank through 704,652 TPM at a cost of $417,505 - including three years of support and after a more reasonable 13.9 per cent discount. That works out to 60 cents per TPM.
While the Opteron 6100 platform could double-up to four sockets and therefore somewhere between 1.3 and 1.4 million TPM, IBM's relative performance (rPerf) benchmarks suggest that a fully loaded Power 780 should be able to do about 3.7 times the work if memory is not constrained, as it seems to be in the Power 780. That puts IBM's a little north of 4.4 million TPM with the Power 780 come November, something an Opteron 6100 platform will not be able to attain to for many years.
Even if you assume 40 per cent of IBM's Power-AIX performance is due to tuning the TPC-C test so each processor's cache gets preloaded with data in a way that's not possible with truly randomized data, IBM could still do 2.6 million TPM on the 780. And, based on the same rPerf numbers, a Power 770 with 64 cores instead of the 32 in the Power 780, but running at that lower 3.86GHz, should be able to do another 61 per cent more TPC-C work, provided you can get enough memory and I/O in the box and they are not constrained. The Power 770 with 64 cores has a 685.1 rPerf rating, while the Power 780 with 32 cores is rated at 425.5 on this relative OLTP test - which is derived from the TPC-C test, by the way.
What all of this really seems to suggest is that IBM was actually ready to ship the Power 750 midrange boxes and the PS701 and PS702 blade servers this spring, but needed more time to cook the Power 770/780 machines. And there's no way IBM will put the Power 795 big bad box into the field until those fatter DDR3 memory cards are available. The jump from 64 cores in the Power 595 to the 256 cores in the Power 795 will require that fat memory, and IBM won't launch these machines until those memory cards are almost ready to go. It's not like Oracle's Sparc and HP's Itanium platforms are putting much pressure on the IBM high-end. At least not yet. ®