Stacking Up IBM i On Entry Power10 Iron Against Windows Servers

April 3, 2023 Timothy Prickett Morgan

Because of the dearth of commercial benchmarks that are run on IBM’s Power Systems machines these days, it is hard to make direct comparisons between machines based on Big Blue’s Power10 processor and the Windows Server and Linux platforms – generally an X86 system based on a processor from Intel or AMD. But it is not impossible to do.

And it is an important exercise not because you are thinking of moving off the platform, which of course you are not contemplating, but because by doing this math you can show that it is not really much cheaper to buy hardware and systems software for X86 machines than it is for the Power Systems machines.

When we sat down to do the math on this for an entry Power10 machine, we had no idea how the math was going to work out. In years gone by, sometimes the OS/400 and IBM i platform was competitive with Windows Server or Linux iron configured to run online transaction processing workloads against a relational database, and at other times the premium that Big Blue was trying to charge for OS/400 or IBM i was embarrassingly high compared to the alternatives.

In this first in a series of economic analysis of the Power10 and alternative platforms, we picked what we think will be a typical machine for IBM i shops with fairly intense workloads (relatively speaking against the 120,000-strong IBM i base, not necessarily compared to the worldwide server base). Since most IBM i shops have one, two, or maybe four cores on their production machines, it is important to start at the low end with this discussion. And interestingly, X86 iron does not typically scale down far enough to make a reasonable comparison.

In fact, the kind of X86 processors from AMD and Intel that might be used for high frequency trading and high performance computing in many guises (think simulation and modeling, particularly electronic design automation), with a relatively small number of cores and comparatively high clock speeds, is the most appropriate configuration to stack up against a Power S1014, Power S1022, Power S1022s, or Power S1024. Which is ironic.

Because of the mix of low core count and high memory capacity and I/O capacity, we decided to see how the Power S1022s, which we profiled in detail here, running IBM i stacks up against the Windows Server stack. It is a bit tricky to figure out what X86 machine to compare it to.

We are well aware that Intel is shipping its new “Sapphire Rapids” Xeon SP v4 processors and that AMD is also shipping its “Genoa” Epyc 9004 processors. But both of these are overkill compared to an entry Power10 machine. Sapphire Rapids has 60 cores and Genoa has 96 cores, which is a lot of compute. We poked around on the Dell site to find AMD Epyc machines, since AMD is setting the performance pace in the X86 market right now, and tried to fine one with a fairly limited core count and relative clock speed. No such luck. But, we did find a machine that had the prior generation “Milan” Epyc 7002 series, which did have some variations that had low core count and high clock speeds akin to what is available in the Power10 entry server lines. To be specific, we configured up a Dell PowerEdge R7515 machine with the AMD Epyc 72F3 processor, which has eight cores, and configured this up against a Power S1022s also configured up with eight cores. Take a look:

We have not tried to do witchcraft to try to figure out how many CPWs of performance this AMD Epyc 72F3 processor might have to run OLTP jobs. This was just attempting to get two eight-core systems configured up with roughly the same sustained clock speed. The IBM machine has a base clock speed of 3 GHz and can be run as high as 3.9 GHz depending on the thermal load on the system; the Dell machine using this specific X86 server chip has a 3.7 GHz clock speed. They are in spitting distance of each other. I put 256 GB of main memory running at a reasonable speed on both machines, four 1.6 TB/sec enterprise-grade NVM-Express flash drives for primary storage, two 10 GB/sec Ethernet ports, and two 16 Gb/sec Fibre Channel ports to link out to SAN storage. The costs for these hardware platforms is shown.

Then, on top of this, I added the cost of server virtualization hypervisors, operating systems, and a relational database, including any perpetual licenses for the software and any per-user additional fees. (Both IBM and Microsoft use a mix of server and user pricing.) The way IBM and Microsoft package up these stacks is quite a bit different, but the end result is the same.

I attempted to add support costs to this basic hardware and software, and I am not convinced there are not additional support costs for Windows Server and SQL Server lurking in there somewhere that I do not know about. But these are the hardware and software support costs that I can find.

On a core for core basis, as you can see from the table above, the IBM i platform is running at a 30.9 percent premium for the hardware and a 36.7 percent premium for the systems software to support 250 users. IBM’s Software Maintenance, at $6,600 per core per year, is crazy expensive compared to Dell ProSupport Plus. I can’t find any additional support costs for the Microsoft software stack, and if there are ones that you are aware of, let me know. Because this is way out of whack.

But here is the thing: Historically, a Power core does about twice the processing work of an X86 core. This has been true as long as we can remember, and despite all of the technical advances in the X86 cores from Intel and AMD, I have no reason to believe this generality is no longer true. So, the table below cuts the number of Power10 cores in half to four to match up against – in a very rough sense – the eight cores in the Epyc 72F3 processor:

When you do that, there is a slight break on the hardware cost, with the Power S1022s as configured only being 19.6 percent more expensive than the Dell PowerEdge R7515. And, the IBM i operating system and database stack was 6 percent cheaper than the Windows Server and SQL Server stack, and PowerVM is, as you can see, around a quarter of the price of VMware vSphere 8. And thus, the four-core Power S1022s with IBM i is a tiny bit cheaper at the hardware and software level compared to Dell-AMD-Microsoft setup.

Once again, IBM Software Maintenance makes the IBM i setup very expensive compared to Dell, even with the cores cut in half. But even still, it is a tolerable premium considering the level of support. But as I said, I think there are Windows Server and SQL Server maintenance fees that we are not finding to add in here.

These numbers are not bad, and frankly, I expected worse.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search