IBM whips out its TPC-C...cluster

OLTP pedal to the metal

graph up

At the last minute, as it was rolling out the low-end and high-end of its Power7-based server lineup on Tuesday, IBM rushed out a new TPC-C online transaction processing benchmark. No surprises there. And there are no surprises that Big Blue has pushed the OLTP benchmark, showing it could put a box into the field that can do more than 10.3 million transactions per minute.

The surprise is that the box IBM was showing off today was not a Power 795, but rather a cluster of three Power 780 servers that finally have those fat memory cards that were missing back in February when that machine was launched. These fat memory cards – which boost the memory of the Power 780 server to 2TB, rather than the512 GB they were restricted to with the skinnier memory cards from February – mean that IBM can turn on all 64 cores in the Power 780 and put the OLTP pedal to the Power7 metal.

Back in April, IBM was showing off that a Power 780 with only two eight-core Power7 processors running at 3.86 GHz, backed by 512 GB of memory and a slew of flash-based drives using controllers made by SandForce. Specifically, IBM has packaged up five PCI-Express controllers and four 177 GB flash drives to create a 3.5 TB SSD offering lovingly known as feature 4637, which costs $99,650. The single Power 780 tested in the spring running AIX 6.1 and DB2 V9.5 had three of these SSD bundles units, which made up 41 percent of the base hardware cost at list price. And contributed mightily to the 1.2 million transactions per minute the machine showed on the TPC-C test.

With the cluster of Power 780s that IBM is talking up today, each server has 64 cores spinning at 3.86 GHz, 2 TB of memory, and eleven of these 3.5 TB SSD bundles, for a total of 38.9 TB of flash storage at a list price of just under $1.1m, or about 34 percent of the hardware cost. IBM charges a lot for memory capacity on those denser memory cards, with the 2 TB of memory in one of the Power 780 costing $748,800, or $366 per gigabyte when you add up the base memory card and memory activation fees. The two-socket Power7 processor card used in the Power 780 server is also not cheap, at $57,429 a pop; each machine has four of these. And it costs another $8,375 per core to activate each Power7 core. So the basic processors in the Power 780s used in the three-way cluster cost $765,716.

Including some disk storage, the chassis, frames, interconnects, controllers, and such that make the Power 780 three-way cluster into a system, the hardware has a list price of $9.61m with $1.6m of maintenance costs over three years. (The first year is free, so that is really two years of cost.) With 360 2 TB disks used in external DS3400 arrays and EXP3000 expansion drawers, storage costs add another $634,921 to the hardware bill and $27,660 to the maintenance.

As for software, the cluster has 192 cores, and AIX 6.1 costs $499,200 at list price and three years of 24x7 maintenance on AIX costs an additional $425,472. DB2 licenses for the three-machine Power 780 cluster have a list price of $11.52m including a year of maintenance; two more years of support runs a cool $4.61m at list, and the total software and support bill for the cluster under test comes to $17.1m.

Toss in a few hundred grand for the clients to simulate the TPC-C end users, and you are talking about $29.5m for this Power 780 setup, and then IBM lops off $15.2m of that, or 51.5 per cent of list price, and drives the price down to $14.3m. Divide by 10.3 million TPM, you get $1.39 per TPM.

The only reason why that number matters to IBM is because it is lower than the $2.36 per TPM that an Oracle T5440 cluster with a dozen machines running Solaris 10 and Oracle 11g with Real Application Clusters extensions. That Oracle Sparc cluster, which El Reg told you all about last October, was originally able to do a little over 6 million TPM at a cost of $2.81 per TPM after a 54.2 per cent discount. Oracle has subsequently tuned the software stack and pricing and was able to push 7.65 million TPM on the same setup running the TPC-C test at a cost of $2.36 per TPM after hefty discounts.

To scale to the 256 cores and 1,024 threads on the Power 795, IBM needs AIX 7.1, which has been radically altered down in its plumbing so it can span the Power7 hardware. AIX 7.1 is not shipping until September 17, but it was not expected to be ready until October. And it is my guess that Big Blue has lots and lots of tuning and tweaking to do to get decent performance across its largest system. And hence it decided to do a Power 780 cluster to prove it had the biggest OLTP system and the best price. A single Power 780 would not have beat the Sparc T5440 cluster, nor would have two boxes.

So where is the Power 795 likely to end up on the TPC-C test? That's a good question, and it would also be very interesting to see what an Oracle RAC cluster of four or eight Sparc Enterprise M9000s would do.

Based on IBM's relative performance (rPerf) metrics for the Power 780 and Power 795 scale OLTP performance whether customers use a few SSDs or lots of disk arms (the system bus doesn't know the difference, really, since either way it gets saturated with data flooding the database), then a single Power 780 with 2 TB of memory and all of its 64 cores should be able to handle 3.45 million TPM assuming very little overhead from the partitioning of the databases on the TPC-C test.

The Power 780 is rated at a little over 685 on the rPerf scale for AIX workloads, and the top-end Power 795 launched today has an rPerf rating of 2,978. That implies that a Power 795 loaded up with 256 cores, 8 TB of memory, and probably something like 156 TB of flash storage should be able to handle around 15 million TPM.

Very few companies in the world will need such a behemoth, of course. But unless something radical happens with the Oracle, Fujitsu, and Hewlett-Packard Unix server lines, none of these Unix stalwarts will have anything that comes even close to this. How much this matters in terms of revenues, profits, and competitive takeouts remains to be seen. But clearly IBM is engineering its machines to do a massive amount of AIX, HP-UX, and Solaris workload consolidation.

If I had to guess, I would say that IBM is saving its TPC-C test results for the Power 795 for the Oracle OpenWorld, which runs from September 19 through 23. ®

Biting the hand that feeds IT © 1998–2018