Stacking Up IBM i And Windows In The Clouds
August 5, 2019 Timothy Prickett Morgan
So how does the IBM Cloud running IBM i stack up against the main competition in the midrange, which is on premises Windows Server plus SQL Server from Microsoft as well as that same stack running in the Amazon Web Services cloud? We can’t give you a definitive answer, but we can give you some food for thought as you ponder how good of a deal Big Blue is giving customers who want to put IBM i on the cloud.
Making comparisons between on premises iron and cloudy infrastructure is difficult and problematic. The good thing about a public cloud with various services for compute, database, and storage is that you can scale these things independently, creating workflows of services that you stitch together that are isolated from each other. The hypervisor does this for a single system like an IBM i platform, all on the same server. On the cloud, servers are carved up as raw infrastructure or loaded up with platform services and each slice on a machine is scaled for individual customers, who may be sitting side by side with each other as workloads come and go and flit around the cloud. Some of these services are very expensive.
As we previously reported, Big Blue divulged its plans for IBM i and AIX on its own public cloud, called the Power Systems Virtual Server on IBM Cloud, back in February at its annual Think conference. Two months ago, IBM unveiled pricing for these IBM i and AIX slices, and a week later we had a look at how IBM compared buying on premises AIX machines with AIX slices on its cloud, and we modified this to make IBM i comparisons for base machines with minimal compute.
This time around, it is going to get more complicated. The server that is running IBM i in the cloud is a two-socket Power S922 with 20 cores and 384 GB of main memory; only 15 of those cores and 320 GB of that capacity is available after the PowerVM hypervisor and the Virtual I/O Server (VIOS) systems software for virtualizing I/O and storage in Power-based systems is added on top of it. It costs $117.69 per core per month to rent capacity for IBM i or AIX on this Power S922 slice in a shared system where others are also on the box; it costs $470.74 per core per month to have capacity on a dedicated, non-shared host. IBM i costs $1,000 per core per month to rent in the P10 software tier that the Power S922 is in.
Now, IBM is not doing this, but we think it needs to, and that is offer an IBM i Application Server license, which it does for perpetual licensees, for cores on the machine with minimal DB2 work. You don’t want to pay for the database on all of the cores in the box, you want to scale a database server separately from the application server but still remain in the IBM i operating system. We think IBM should charge $100 per month for this IBM i Application Server license, or about a tenth of the full license price. The reason is simple: At $1,000 per core per month or $14,995 per core for a perpetual license plus $9,660 over three years for Software Maintenance on top of that, the cost of IBM i adds up pretty fast across the 15 usable cores on the Power S922. Like close to $370,000, which is a lot. Ditto for the $1,000 per core per month rental fee for IBM i on the IBM Cloud.
If you configure one IBM i slice of a Power S922 on the IBM Cloud with 5 cores with the full Db2 for i version of IBM i, that costs $9,011 per month with 128 GB of memory; adding another 10 cores of IBM i on another instance with 256 GB of memory costs another $5,021 per month. (It will become obvious in a second why I am doing this in this fashion.) That’s $14,032 per month for a setup that has about one third of its capacity dedicated to database and two thirds dedicated to application serving, which works out to $505,152 over three years.
Now, say you want to buy the same setup and run it on one Power S922. The base hardware costs $55,408 running AIX Enterprise Edition, which is essentially free at $988 per core on small machines. That works out to $40,588 after you peel off AIX for just the hardware. Now, we have to add the full IBM i and IBM i Application Server licenses to this box. It is $14,995 per core for IBM i, so that is another $74,975 across 5 cores and another $14,950 for IBM i Application Server on 10 cores (we estimate the price at one-tenth the full IBM i perpetual license price). Now, you have to add in $22,300 per year to this to cover floorspace, admin, power, cooling, and hardware and software maintenance. Add this all up over three years, and the acquisition of this setup would cost $197,413, which means the cloudy version of this same stack with on demand pricing (and no discounts applied) costs 2.5X that of the on premises iron.
This gap is not shocking. That’s the nature of the public cloud. That’s the price you pay to shift from capex to opex and to get flexibility.
Now, let’s take a look at what a Windows Server setup might cost on Amazon Web Services, which is by far the most popular public cloud on earth.
Let’s start out by being lazy and say, for instance, you don’t want to buy a raw virtual machine on AWS and load up and manage SQL Server, but rather you want to use its Relational Database Service, which as the name suggests comes preconfigured with a relational database (but not storage for data). The Relational Database Service can be equipped with a variety of databases, including Oracle, PostgreSQL, MySQL, and SQL Server.
On the latest generation db.m5 instances for the Relational Database Service, to get something in the same rough performance class as the Power S922 IBM has put into its cloud running IBM i and AIX, you might pick the db.m5.12xlarge instance, which has 48 virtual CPUs and 192 GB of memory with a single 10 Gb/sec Ethernet port linking it to the outside world and to the AWS backbone. Each virtual CPU is a thread on an Intel Xeon SP processor, which means this is equivalent to a two socket machine using 12 cores in each processor. The same core count as the two-socket Power S924 we found relevant integer benchmarks for to do our comparison and obviously more than the 20 cores that the Power S922 on the IBM Cloud has.
We base this rough equivalence of compute on raw integer performance from pouring over the SPEC CPU2017 integer benchmarks. A two-socket Power S924 running Power9 cores at a base 3.4 GHz and having two dozen cores in total has a SPECrate2017_int_base rating of 213. The Power S922 has cores running at 2.9 GHz and if it were tested would probably have somewhere around a 152 rating on the same SPEC integer throughput test. If you dig around on the SPEC site, you can find a Dell PowerEdge R740xd system with a SPECrate2017_int_base rating of 152, which has a pair of “Skylake” Xeon SP 6136 Gold processors, each with 12 cores running at 3 GHz, and 384 GB of main memory.
If you price this machine up with the Dell online configurator and put Windows Server Datacenter Edition and SQL Server Enterprise Edition on it, with Pro Support on the hardware and the Microsoft software for three years, then this machine sells for $32,300, which is obviously very inexpensive. If you add in the same $15,000 per year that IBM figures floor space, administration, and other costs eat, that’s another $45,000 over three years, for a total of $77,300. The Microsoft stack and its support are heavily discounted by Dell, which helps that Windows-X86 platform price come down quite a bit. The Power S922 running IBM i as we configured it above costs 2.5X as much as the Windows Server system to acquire. This is consistent with the pricing we have seen between OS/400 and IBM i versus Windows Server for the past two decades. This gap is nothing new, even if it is annoying.
But mind this gap, which is even bigger.
If you fire up the price configurator on AWS and plug this in with a full Microsoft SQL Server Enterprise Edition license with a mere 20 GB of data storage, you better keep your heart medication close because this costs a whopping $22,255 per month for an on-demand instance. If you cut that back to the SQL Server Standard Edition license, you are talking about $11,804 per month. You still have to buy infrastructure to support the applications on top of this, plus storage. And if you want to move data off AWS, you will have to pay for that, too.
For the sake of argument, let’s assume that a third of the workload is for the database and two-thirds is for the application. You can step down to a db.m5.4xlarge instance for the database workload on AWS, and that will run you $4,064 per month for SQL Server Standard Edition and $7,525 per month for SQL Server Enterprise Edition. That’s a lot less dough, but now you are beginning to see the value of the self-optimizing, integrated relational database management system that IBM i has and has had since 1988. Anyway, that leaves an m5.8xlarge instance, with 32 vCPUs and 128 GB of memory, as the application server. That will cost you $2,422 per month running Windows Server 2016 as the application serving platform. Add that up, and you are looking at basically just a hair under $10,000 per month for the combination of a Windows Server application server and a SQL Server Enterprise Edition database server with no storage at roughly the same combined performance as that of the Power S922 slice setup outlined above. If you go with reserved instances for the application tier with three years prepaid – the lowest price you can be charged on AWS – you can shave $1,000 a month off that price. (Dropping down to SQL Server Standard Edition is not a fair comparison to the Db2 for i database inside IBM i, I think.) There doesn’t seem to be reserved instance discounts on the database service on AWS. So call it $9,000 a month or $324,000 for a cloudy system that can do about as much database and application work as a Power S922.
That’s a factor of 4.2X between the on premises Dell PowerEdge R740xd compute equipped with software but no storage and the combination of the Relational Database Service for SQL Server and Windows Server on the EC2 infrastructure service. That gap is huge, and IBM deserves some credit for keeping the gap between on premises and cloud lower – 2.5X for the IBM i stack versus 4.2X for the Windows stack. We wish the gap was smaller between IBM i on the cloud and AWS configured for database and application serving on the Windows stack at roughly the same performance, but the IBM i setup we describe is 1.55X as expensive as the AWS Windows setup over three years. But IBM can close that gap substantially by offering reserved instance discounts for one year or three year terms, which we think it will do soon. A 30 percent discount for reserved instances for IBM i would absolutely close that gap, cloud to cloud.
One last thought: With that big gap for X86 iron running on premises and in the cloud, you begin to see why AWS has such high margins, and moreover, customers can argue hard for discounts from AWS or find another cloud. It’s not your job to make Jeff Bezos richer, after all. Bezos gets enough of your retail business already.