Entry Server Bang For The Buck, IBM i Versus Windows Server
November 4, 2019 Timothy Prickett Morgan
Some big changes that Microsoft has instituted with its Windows Server platform to make pricing consistent across on premises and public cloud deployments has had the interesting side effect that entry IBM i machinery based on Power9 iron is now more competitive with entry X86 servers using the latest processors from Intel and AMD.
This is not universally true, mind you, but it is certainly true of machinery in the P05 software tier where a lot of the IBM i base hangs out. There is still a large gap on entry iron in the P10 software tier, and we did not have the stomach to price out reasonably configured P20 systems. (We will get to that at some point if we can gather enough pricing information.)
We saw that IBM had posted list pricing on its web site for the Power S914 entry machines, and that got us to thinking that we have no really done a detailed competitive analysis in a while, and certainly not with the usual detail we like, because IBM did not provide full list pricing to the public on the Power9 iron. We did a detailed analysis back in early 2017 when the Power S812 “Mini” system was launched, but did get around to comparing this machine and its peers in the Power Systems line to Windows Server and Linux systems equipped with the same stack of hypervisor, operating system, and database software. We did do a review of IBM’s own competitive analysis pitting its entry Power8 machines against roughly equivalent X86 iron way back in September 2014. This turns out to be pretty useful five years later, believe it or not.
We used the information in that story, which compared Power8 machines from IBM to X86 iron based on Intel’s “Ivy Bridge” Xeon E5 v2 processors. At the time, IBM offered relative performance metrics for SAP Sales and Distribution (SD) and SPECjEnterprise2010 transaction processing benchmarks as well as SPEC integer and floating point rate tests, and the composite of these averaged out to a Power S824 with two twelve-core 3.52 GHz processors having a little less than 2X the performance of a system with two 12-core Xeon E5-2697 v2 processors running at 2.7 GHz. Knowing this data, and the relative performance of the Power and Xeon chips going forward in terms of cores, clock speeds, and instruction per clock for successive architectures, we can get a pretty good approximation of raw OLTP performance as gauged by IBM’s Commercial Processing Workload (CPW) benchmark test, which is heavily inspired by the TPC-C test that Big Blue no longer runs on Power Systems iron.
In our comparisons, we use X86 iron from Dell, and we tried as best as possible to get processors with low core counts like IBM is offering in the Power S914, which comes with four, six, or eight cores. We also tied to stick with single-socket machinery in rack enclosures because this is the form factor that the Power S914 comes in, although to get a comparison with the latest “Cascade Lake” Xeon SP processors from Intel, we had to half populate a two-socket machine.
There is no true single-socket Xeon SP machine, which is odd, although Intel does repurpose its high-end Core processors with integrated graphics processors as single-socket servers and brands these the Xeon E-2100 series, which have the “Coffee Lake” cores in them that are very similar to Skylake cores. These were just updated at the end of May with a revised Coffee Lake core and that have higher core counts (up to eight cores) and slightly faster clock speeds plus some new security features; they are only now appearing in servers, and Dell does not have them as yet. And so, for fun, we used an eight-core server based on AMD’s “Naples” Epyc lineup – specifically the Epyc 7261 clocking in at 2.5 GHz – to run off against the eight-core Power9 chip running at 2.8 GHz.
On the four core Power S914, IBM caps the main memory at 64 GB even though it could, in theory, reach 1 TB, and it does so because many IBM i shops would never buy a P10-class machine otherwise. So I capped the memory at this level on the Cascade Lake system – the PowerEdge R640 – even though this is ludicrous. On the even more entry PowerEdge R340 system, it only has a single processor socket and only four memory slots and the memory tops out at 16 GB capacity per stick, so Dell has a natural 64 GB cap here. (This ain’t much memory for a modern server, but it can still function.) I put 256 GB of capacity on the machines with six and eight cores.
In the monster table describing these machines, which you can see here, you can see all of the feeds and speeds of the configurations. It includes the base and turbo clock speeds for the processors as well as base memory, base disk, and base Ethernet networking that comes with each machine. The idea here is to show how the entry machines are initially configured. For the Dell machines, there is the base configuration price and then there is a second price for customers who have a special code (which is published right there on the page) that gives them 37 percent off on the hardware (except in one case where it was 63 percent for reasons I cannot explain). I have ignored the discount in the price/performance and cost per user comparisons, since we don’t have the comparative discount levels from IBM for the Power S914. This is a list price to list price comparison. There are a wide variety of SAS disk drives that come in these five different machines, and I did my best to keep the disk arms and capacity in the same ballpark. Some machines have a dozen drive bays, some only four, so this is not an apples-to-apples comparison in a lot of cases.
To calculate the number of users on each system, I took the CPW rating for each machine and divided it by 2,100 CPWs. The idea was to show how software costs scaled with users, and to have users scale with the underlying OLTP performance of each machine.
On the software front, the Power S914 is configured with the latest IBM i 7.4, which includes a license to the PowerVM hypervisor for carving up the system into logical partitions as well as to the Db2 for i relational database. The P05 license costs $1,996 per core after you back out the three months of Software Maintenance in the license to IBM i. I did not add software support to any of these configurations – I just didn’t have all of the data to do so, it wasn’t laziness. On the Windows Systems, I did the Cadillac implementation, using VMware’s ESXi hypervisor and vSphere tools to control server virtualization, and then putting Microsoft Windows Server 2019 Datacenter on it as the operating system and then SQL Server 2017 as the database engine.
A funny thing happened on the way to the cloud: Microsoft changed its Windows Server licensing, and it is a bit crazy as is always true of a company with a legacy platform with a revenue stream to protect. Microsoft shifted away from per-socket pricing to per-core pricing for Windows Server and SQL Server a while back, but it sorta didn’t at the same time. Windows Server has a minimum charge for eight cores per socket and a minimum charge of two sockets per server, which is the same thing as saying anything between 1 and 15 cores requires a license for 16 cores. For the Datacenter edition, which includes full-on virtualization and private cloud deployments, that base license for 16 cores costs $6,155 – even if it runs on machines with four, six, or eight cores. That is the same thing as saying customers with small machines with limited cores will pay anywhere from 2X to 4X as much per core for the same software.
Ditto for Standard edition, which has a license fee of $972 for a 16-core base pack. Both of these require another $38 per user Client Access License (or CAL), which is not the same thing as Client Access for IBM i. There is a Windows Server 2019 Essentials edition, which is for servers that have a 25-user cap that costs $501 across 16 cores. I do not think this is a real operating system as IBM i shops know it, but many probably could get away with it. I do think that IBM i is an enterprise-grade platform, so the comparisons are with Windows Server 2019 Datacenter edition, which costs a lot more. Standard edition only allows two Windows Server instances, which is very limiting in a world where servers have multiple partitions.
For the database on the Windows Server platform, it costs $7,128 per core for the Enterprise edition of SQL Server 2017, which is way overkill in that it bundles in Spark, Hadoop, and various data warehousing stuff that is not needed for OLTP workloads. That’s why I chose SQL Server 2017 Standard edition for the comparisons outlined here. That costs $3,717 in a two-core pack using per-core licensing. (There are generally not odd numbers of cores in modern processors, the 11-core Power9 chip excepted. Hence the two-pack. Not to be confused with Tupac.)
For virtualization, the latest ESXi hypervisor and vSphere Enterprise Plus tools to make use of it, which costs $3,595 per socket, is added to the box.
These are not perfect comparisons, I realize. But there are lots of datapoints to consider, which is what I am trying to give to you. The thing to remember is this: You can make a cheaper Windows Server platform, but you can’t make a cheaper IBM i box. You could cut out VMware entirely and use the integrated Hyper-V hypervisor, and shift down to Windows Server 2019 Standard edition. And doing so would cut anywhere from 38 percent to 53 percent of the software cost out of those PowerEdge systems, which would significantly boost the bang for the buck for the Windows Server machines shown. But we don’t think this is how enterprises deploy software for running mission critical workloads. We could have thrown in the 37 percent discounts on hardware, too. But the point is to show you what you need to argue for if you are negotiating your own deal for a new Power S914 – or indeed any entry Power9 machine in a world where Intel and AMD are locked into a bang for the buck battle.
There is a lot in that table, we realize, but here is a summary chart that brings it all home:
The stacked columns show the costs of hardware and software for the configurations, which are on the left Y axis, and the black columns show the price/performance for the configurations, on the right Y axis.
As you can see, the four-core Power S914 beats out all of the Windows Server configurations when it comes to price/performance, and core for core, the Power9 chip does about 1.6X the work of the current Cascade Lake cores at a much lower clock speed. Clock for clock, the Power9 does about 2.2X the work of a Cascade Lake chip, as best as I can figure – so the gap is widening a little here, but still around 2X. (It has been around 2X for a very long time – as long as I can remember, in fact.)
When you add cores to the Power S914 and it jumps up into the P10 software tier, then Windows Server comes out on top, even when I keep it at the Cadillac comparison with Datacenter edition of Windows Server and using VMware for virtualization. As I have said before, the pricing gap between P05 and P10 is too large, and this shows you by about how much it needs to come down. Cutting it in half puts IBM i and the Windows Server stack on configured hardware within spitting distance of each other.
Just for fun, I calculated price/performance in cost per CPW and cost per user, so you can think of it either way. Cost per user is a lot easier to get our heads wrapped around, and if you are amortizing the cost over five years, then you can see that a new system is fairly inexpensive to one way of thinking about it. It costs about as much per year to give an employee a slice of a small OLTP system as it does to buy them a very cheapass laptop at Wal-Mart every year. If IBM changed the P10 pricing as I suggested, it would be on the order of $400 a year. I am not saying you should spend that much on laptops, of course. But that a server slice on a mission-critical system costs 5X what a cheap laptop does and maybe 2X what an expensive one does is not unreasonable, considering that they are both tools. We want IBM i shops to continue to invest in this tool.
Next up, we will see how shifting to Linux on the X86 iron changes the cost, and I will try to get comparisons for the Power S924 and Power E980 together that do the same thing. It is very hard to get server pricing information, so if you know something and want to share, you know where to send it.