IBM i Shops Pay The Power8 Hardware Premium
May 19, 2014 Timothy Prickett Morgan
The initial entry Power8 systems were announced a month ago, and we have been analyzing the IBM i 7.1 Technology Refresh 8 and IBM i 7.2 releases of the operating system over the past few weeks because that software runs on lots of different hardware and the Power8 systems were not going to be available until June. We are getting close to the date when the machines will be available and so I figured I had better get the lead out and do some configuring and pricing.
IBM i customers are not going to be happy, I think. Once again, we are back to the days when the IBM i shops are paying a premium over Linux shops. The two machines that run IBM i can also run AIX, but the single machine that can run AIX or Linux has a price/performance that is a little better than these two IBM i/AIX boxes.
In my initial pass on the analysis of the Power8 systems, I grabbed the entry configuration and pricing information off the IBM online store. The single-socket Power S812L (the entry Linux-only machine) is not going to be available until August 29, but the other four machines are ready to rock on June 10. Here are the base configurations of the five Power8 systems in their processor variations:
Here are some observations to begin. First, these machines are all using what are called dual-chip modules, and IBM is taking two half slices of a Power8 chip and putting them together into one socket to make the chip. The full-on Power8 chip has 12 cores and two memory controllers, and if half of it is messed up because of yield issues on IBM’s 22 nanometer processes, it slices the chip in half and throws away the bad half. The remaining chip has three, four, five, or six cores that work, and two of these are used in a single Power8 socket to make a socket that has six, eight, 10, or 12 cores. This sounds like a lot of work, but IBM did this with several machines in both the Power5+ and Power7+ generations. Presumably, we will get an entry Power8 machine with only four cores to satisfy the modest processing needs of small and midrange businesses running the IBM i operating system, and presumably this will also be created using DCMs.
As you can also see, the two Linux-only machines have slower but fatter SAS disks than the remainder of the line and have more cores per socket. I have no idea how much this changes the price on the disks. Eventually, I will dig through the feature prices and configure up machines with Linux, AIX, and IBM i with proper configurations suitable for OLTP workloads and get a system-level comparison. This all takes a lot of time. I am still trying to figure out what IBM i costs on these machines, which are in the P10 software group, just to make sure this has not changed. The IBM online store used to allow for IBM i to be configured on the machines, but it has stopped that practice.
The other interesting thing about the data IBM is providing about the pricing is what it costs to get a 36 month lease on the entry configurations for each machine. I could get a two-socket Power S822 machine with two six-core Power8 DCMs (for a total of a dozen cores in two sockets) for about what I would pay per month for the 2014 Dodge Challenger RT with the 5.7 liter V8 Hemi and the 20-inch rims–bright white, silver gas cap, silver rims, leather seats, sport model with the sunroof, and a license plate that says YIELD. If I wanted to really go for it, a two-socket Power8 S824 box would be in the same monthly price as the Challenger SRT with the 6.4 liter V8 Hemi. That one has the same styling in my version, except the license plate now says DEFER.
The point is, when you think about it that way, a Power8 box that can run IBM i doesn’t seem all that expensive. Ah, but you know me better than that. I never, ever want the IBM i customers to pay one penny more than Linux or AIX customers, or for that matter, than a competitive Xeon or Opteron box running Windows or Linux and a database and application serving software stack. So you can just about imagine my reaction when I put this table together:
This is the kind of thing that makes me want to hack into IBM’s systems and change its pricing. (Fear not, Big Blue. I don’t have the skills for such an outlandish maneuver.) I understand that the Power Systems business has to make more profits than it has and that the Power8 machines have to take on X86 iron running Linux and meet or beat them on the field. I understand everything, intellectually. But emotionally, I just don’t plain like it.
In the table above, I had to guesstimate the performance metrics for IBM i workloads on the two Linux-only boxes, the Power S812L and Power S822L as well as the Power S822, none of which are allowed to run IBM i precisely because they have lower prices per unit of performance. The performance for IBM i workloads is reckoned using IBM’s own Commercial Performance Workload (CPW) test, which is an I/O unconstrained version of the TPC-C online transaction processing benchmark test that Big Blue has used to show relative performance of processors for a very long time. On AIX, IBM uses a different test that is also based on the TPC-C test called the Relative Performance benchmark, or rPerf. The two Linux machines also did not have CPW or rPerf ratings, but there is nothing special about the hardware that doesn’t allow me to get a pretty good guess. So I did that.
And lo and behold, as we all expected even before the entry Power8 machines launched, on entry configurations, the Linux-only machines are considerably less expensive than what IBM i and AIX customers are paying. This being The Four Hundred, we will focus on the CPW price/performance of the five machines in their various incarnations.
The Linux-only machines cost somewhere 8 or 10 cents per CPW in the entry configurations, as you can see. The base Power S814 that is most like the PowerS812L Linux-only machine costs 19 cents or 21 cents per CPW, by comparison. This is in perfect keeping with a rule I have observed for two decades now: RISC/Unix and AS/400 capacity costs half of what mainframe capacity does, and X86 capacity costs half of what the RISC/Unix and AS/400 capacity does. And so, the Power-Linux machines are priced like X86 iron. They have to be to compete. Period.
The Power S824 is the bigger box that can run IBM i and AIX (and Linux if you want), and it is also in the P10 software group like the Power S814. And as you can see, the processing capacity on the entry Power S824 is considerably more expensive than on its single-socket baby brother, ranging from 27 to 31 cents per CPW.
The Power S822 has a little less performance than the Power S824, and many IBM i shops might be inclined to go with that one, but alas, it cannot run IBM i. Only AIX and Linux are allowed, and this is clearly a box that is aimed at AIX shops who would otherwise grouse about the premium that IBM i shops are being asked to pay. It is hard to imagine this Power S822 machine being chosen to run Linux unless the Linux customer needed the faster cores than the Power S822L machine offers.
None of this seems fair, of course. It is possible that on configured systems, including operating systems, hypervisors, and databases, the differences in hardware prices for entry configurations are largely irrelevant. I will start cooking them up so we can find out. Stay tuned, and press for Linux or at least AIX pricing on those Power8 boxes that run IBM i. Those are price ceilings listed above, not price floors.