Bang For The Buck On Power9 Entry Hardware
March 19, 2018 Timothy Prickett Morgan
When it comes right down to it, there are three things that drive a server upgrade or new system purchase among IBM midrange shops: They have an aging system that has to be replaced, workloads are expanding and driving new capacity, or new workloads are being added to the system. In many cases, the situation involves two or even all three of these factors. But what gets the deal done is the value that the new system brings above and beyond satisfying those needs. You have to get a good deal so everyone looks like a hero.
The good news is that, at least as far as the basic compute and storage hardware is concerned, the new Power9 “ZZ” systems that were announced in February and that are the star of IBM’s Think 2018 conference in Las Vegas this week, the Power9 machines are providing considerably better performance, core for core and socket for socket, and thanks to price decreases on memory and disk drives, a base Power9 ZZ system – that’s the Power S914, Power S922, and Power S924 if you want to run IBM i as your primary operating system – is considerably less expensive than the Power8 systems they replace – that would be the Power S814, Power S822, and Power S824.
I have been waiting for a long time for Big Blue to publish the relative performance of the new Power9 systems using its Commercial Performance Workload (CPW) online transaction benchmark test, which is a variant of the TPC-C benchmark that IBM has tweaked long since and has used for so many generations of Power iron that I can’t remember when RAMP-C was phased out. Last week, we told you about the performance impact that the Spectre and Meltdown patches to plug security vulnerabilities in speculative execution processor and memory management in Power8 and Power9 processors. I figured that this was more important than doing the initial pass on price/performance analysis for the Power9 iron compared to Power8 machinery, particularly with the Power9 ZZ machines not yet shipping. But shipment day is March 20, so this is perfect timing to talk about this.
You can look at the latest performance specs on Power Systems machines, including the Spectre/Meltdown impacts, in IBM’s Power Systems Performance Report, which you can download here and which was published on February 27. For some reason, IBM did not test every Power8 or Power9 machine in every possible configuration, so I have had to fill in the blanks a bit here in there. In the tables below, I have made my best-guess estimates at what performance should look like given the core number and socket counts of the systems. This is the same estimation that IBM engineers, business partners, and customers have to do. If you want to do your own, look to the Related Stories section at the bottom of this story, which has links to our analysis of both the Power8 and Power9 entry servers. The performance shown for each configuration is for a machine that is neither memory nor I/O bound (except intentionally as in the four-core variants of the Power8 S812 and the Power9 S912 machines, which have a 64 GB memory cap), and estimated performance is shown in red, bold italics.
Back then the Power8 entry machines were launched in April 2014, I did price/performance analysis on the configurations that IBM was selling on its Web site and used the pricing from those machines. With this comparison, I am going back to the basics and configuring a bare bones machine with the chassis, the processor, core activations where needed, 32 GB of memory per core, and four disk drives for operating system storage. The Power8 machines use 300 GB drives spinning at 15K RPM, and the Power9 machines use 600 GB units at the same speed. The machines do not have disk controllers, network interface cards, flash drives, or other peripheral controllers or expansion units on them. The idea is to give a sense of the relative bang for the buck for the core compute complex and enough memory for it to be useful, plus some local storage for IBM i. This is not necessarily the configuration that would be necessary to drive that performance on the real test for either TPC-C or CPW. In fact, certainly not for TPC-C, which has a very high I/O requirement. The CPW test is all about stressing the CPU and assuming there is not a lot of I/O or contention for memory.
After revamping the Power8 configurations to be more consistent with the Power9 base configurations I used in my initial pricing for the machines, here is what they look like:
As you can see, the price/performance has the typical curve we have seen in the IBM midrange since at least the launch of the AS/400 back in 1988. The entry machine has the lowest price, but not the best bang for the buck, while the middle of the line machines have the best price/performance but cost quite a bit more. The larger machines cost more and have price/performance that is not quite as good as the middle, and in some cases lower than on the entry machines. (Lower price/performance means a higher number, remember.)
With the Power8 machines, if you could run IBM i on the Linux-only Power S812L and Power S822L machines, you would get pretty good bang for the buck – from 22 cents to 26 cents per CPW for that base configuration shown. On the Power S814 that actually could run IBM i, the list pricing back in April 2014 pegged it at around 31 cents to 33 cents per CPW, and with the beefier and more expensive Power S822 (which also came in a higher IBM i software pricing tier) had better value, at around 25 cents to 27 cents per CPW. These are very tight ranges, you can see. That is probably not an accident on the part of Big Blue. For the Power S824, which has more scalability on every front, the pricing is quite a bit higher, at between 37 cents to 66 cents per CPW. The biggest outlier was the Power S824 machine that had two Power8 chips with twelve cores running at 3.89 GHz. We suspect this was not a very popular machine, but if you need a combination of single-threaded performance and lots of multi-threaded throughput, this may be a machine you had to buy.
Jump ahead four years, and here is what the value situation looks like with the Power9 entry servers:
The entry Power S914 machines have about the same price or slightly higher as the entry Power S814 machines, but offer somewhere between 33 percent and 43 percent more performance, and that means the bang for the buck improves to around 24 cents per CPW, roughly a third to a quarter better bang for the buck depending on the comparison.
For the more midrange (and denser) Power S922, the systems cost somewhere between 17 cents and 19 cents per CPW, and you can buy fewer cores to get a certain level of performance. (In fact, that is true of all of the machines, now that I think of it, but it is obvious here.) That’s about 35 percent better bang for the buck, and you can tune the configuration to have fewer cores and therefore lower IBM i and third party software licenses if they are priced per core. (Existing customers probably just want to be able to slide their OSes over from one machine to the other, which is event cheaper.)
On the larger Power S924 machines, which are the workhorse of the IBM midrange, pricing works out to 27 cents per CPW, the price/performance has improved by about 30 percent or so at the averages of the cost per CPW ranges, and this is absolutely consistent with what IBM said when it announced the Power9 entry machines back in February.
Having settled what the iron costs, next we have to consider the cost of systems software and Software Maintenance support for these respective machines.