Counting The Cost Of IBM i On Power9 Entry Systems
April 9, 2018 Timothy Prickett Morgan
Every time there is a Moore’s Law advance in processing, whether companies charge by the core or by the socket, there is price cut – in theory at least – for systems software. This is true if the chip maker decides to goose the performance of the cores or add more cores to a socket, or both at the same time.
In the past several generations of Power processors used in the Power Systems line, IBM has done both. Because the systems software is based on cores, however, the move from two to four to eight to 12 cores per socket has not meant a consequent reduction in the price of software such as the IBM i platform and its integrated database management system. But the performance of cores, through a combination of clock speeds and architectural improvements, has generally gone up generation to generation, and with the jump from the Power8 to the Power 9, the base improvement, core for core and at clock speeds typically used in Power8 and Power9 configurations is somewhere between 30 percent and 50 percent, as IBM explained to us at the Power9 “ZZ” systems launch back in February. The raw performance of the architecture, normalized for a socket and at a baseline clock speed we reckon is around 4 GHz, was somewhere between 50 percent and 95 percent, depending on the workload, IBM divulged back in the summer of 2016 when the Power9 architecture was previewed to the chip illuminati.
A few weeks ago, we drilled down into the price/performance of the new ZZ systems in base hardware configurations as compared to the prior generation of Power8 entry machinery, which you can review here. The configured machines, which are based on the processor configurations that IBM used to do its Commercial Performance Workload (CPW) benchmark test plus a reasonable amount of memory and disk, cost anywhere from $12,641 for a Power S914 with four cores running at 2.3 GHz, 64 GB of main memory and four 600 GB disk drives to $99,290 for a Power S924 with 24 cores running at 3.4 GHz, 768 GB of memory and those four base 600 GB disks. The bang for the buck on the Power S914, Power S922, and Power S924 machines had a very tight range of between 23 cents and 27 cents per CPW. The price range on Power8 machinery – the Power S814, Power S822, and Power S824 – was virtually the same, but the cost per CPW on the machines had a much higher and broader range, from a low of 25 cents per CPW for a hefty configuration of the Power S822 to a high of 66 cents per CPW for a midline Power S824.
To figure out what the base software costs were on these two sets of entry Power Systems machines, we figured out what OS/400 and IBM i software tier each configuration was in, and then dug around to get the per-core pricing for IBM i. We took out the 90 days of Software Maintenance (SWMA) that is included with the base Power S814 and Power S914 machines to make it equivalent to the rest of the machines, which do not have any SWMA bundled. All Power Systems machines running IBM i require that a year of SWMA, as a minimum, be configured to them, but we are ignoring this for now. We are just trying to isolate the costs of the software. We configured IBM i on all of the cores, which are all activated on the machines by default starting with the Power9 iron. IBM i has a base one-time perpetual license charge, and then there is a per-user fee of $250 a person on top of that. We added this user cost in, reckoning that somewhere around 5,000 CPWs per user is representative of the number of users on a typical machine in these classes. (We do not think that companies actually allocate this much CPW to users, but their database and batch jobs might require that much oomph even if their online transaction processing workloads do not.) I can recall when it sounded like 500 CPWs or 1,000 CPWs per user was a lot.
Here is how the Power8 entry machines look:
Obviously, the Power S812L and Power S822L machines, which are Linux-only boxes, cannot be loaded up with IBM i. We left the Linux-only machines in here for continuity with the hardware pricing we did a few weeks ago. The CPW ratings are our estimates for these configurations, and other estimates are also shown in bold red. On the P05-class Power S814 machine, the cost of the base IBM i license is $7,985 across those four cores, and with the 16 users shown, the price rises to $11,985. This works out to 30 cents per CPW for the software – a little bit less than the cost of the hardware on a cost per CPW basis. This is the only time that happens, so don’t get excited. (And if we added even a few more users to the machine, it would not be true.)
Hardware costs rise as machines get more powerful, but software costs rise even faster, as you can see from the above table. You can also see the effect of using machines with lots of cores running at lower clock speeds on IBM i pricing. Basically, don’t do that. Get the fastest cores you can and get as few of them as you can is our best advice.
How much more expensive is the hardware than the software? That depends on the machine, but the answer to that question is generally: a lot more. Take a look:
This shows the cost per CPW of the hardware configurations we ginned up a few weeks ago against the base IBM i configuration shown in this story. The factor of price difference between software and hardware is anywhere from 4.6X to 12.4X, as you can see.
Now, let’s see how the Power9 entry machines stack up to these. Here is the software costs for the ZZ systems that we profiled back in March:
In general, the software costs, on a per CPW basis, are lower, ranging from a low of 25 cents to a high of $2.95. This is much better. But as you can see in the chart below, the delta factor between hardware and base IBM i pricing is still in the same kind of gap, ranging from a low of 4.7X to a high of 11X:
That said, the price/performance improvements for both base hardware and base software are pretty significant, so that is the good news. There has been Moore’s Law improvements for both the hardware and the software, which there certainly should have been give the four year gap between these two families of machines. The pace has slowed, to be sure, but the deals are still better. The issue is that it is not possible to get the cost of a base machine below $25,000. We still think this sticker price is too high compared to X86 iron, something we will explore in a future issue of The Four Hundred.