Today is our last day to talk about infrastructure as a service. We are moving up the stack into platform as a service after this. The higher up the stack we get, the more value it has to the business and end users. It is interesting to talk about storage and compute in the cloud but does this help us treat patients in a medical practice any better, find oil and gas faster, deliver our manufactured product any cheaper? Not really but not having them will negatively effect all of them. We need to make sure that these services are there so that we can perform higher functions without having to worry about triple redundant mirroring of a disk or load balancing compute servers to handle higher loads.
One of the biggest complaints with cloud services is that there are perception problems of security, latency, and governance. Why would I put my data on a computer that I don’t control. There is a noisy neighbor issue where I am renting part of a computer, part of a disk, part of a network. If someone wants to play heavy metal at the highest volume (remember our apartment example) while I am trying to file a monthly report or do some analytics, my resources will suffer and it will take me longer to finish my job due to something out of my control.
Many people have decided to go with private hosted cloud solutions like VCE, VBlock, Cisco UCS clusters, and other products that provide raw compute, raw storage, and “hyper-converged” infrastructure to solve the cloud problem. I can create a golden master on my VMWare server and provision a database to my configuration as many times as I want. Yes, I run into a licensing issue. Yes, it is a really big licensing issue running Oracle on VMWare but Microsoft is not far behind with SQL Server licensing on a per core basis either. Let’s look at the economics of putting together one of these private cloud servers.
It is important to dive into the licensing issue. The Oracle Database and WebLogic servers are licensed either on a processor or named user basis. Database licensing is detailed in a pdf on the Oracle site. The net of the license says that the database is licensed based on the core count of the processor running on the server. There is a multiplication factor (0.5 for X86) based on the chip type that factors into the license cost. A few years ago it was easy to do this calculation. If I have a dual core, dual socket system, this is a four core computer. The license price of the computer would be 4 cores x 0.5 (Intel x86 chip) x $47,500. The total price would be $95K. Suddenly the core count of computers went to 8, 16, or 32 cores per chip. A single system could easily have 64 cores on a single board. If you aggregate multiple boards as is done in a Cisco UCS system you can have 8 board or 256 cores that you can use. There are very few applications that can take advantage of 256 cores so a virtualization engine was placed on top of these cores so that you could sub-divide the system into smaller chunks. If you have a 4 core database problem, you can allocate 4 cores to it. If you need 8 cores, allocate 8 cores. Products like VMWare and HyperV took advantage of this and grew rapidly. These virtualization packages added features like live migration, dynamic sizing, and bursting utilization. If you allocate 4 cores and the processor goes to 90%, two more cores will be made available for a short burst. The big question comes up as to how you now license on a per core basis. If you can flex to more processors without rebooting or live migrate from a 2 core to a 24 core system, which do you license for?
Unfortunately, Oracle took a different position from the rest of the industry. None of the Oracle products contain a license key. None of the products require that you go to a web site and get a token to allow you to run the software. The code is wide open and freely available to load and run on any system that you want. Unfortunately, companies don’t do a good job of tracking utilization. If someone from the sales or purchasing department rolls out a golden master onto a new virtual machine, no one is really tracking that. People outside of IT can easily spin up more licenses. They can provision a database in a cloud service and assume that the company has enough licenses to cover their project. After a while, licensing gets out of control and a license review is done to see what is actually being used and how it is being used. Named user licenses are great but you have to have a ratio of users to cores to meet minimums. You can’t for example, buy a 5 user license and deploy it on a 64 core system. You have to maintain a typical ratio of 25 users to a core of 40 users to a core based on the product that you are using. You also need to make sure that you understand soft partitioning vs hard partitioning. Soft partitioning is the ability to flex or change the core count without having to reconfigure or reboot the system. A hard partition puts hard limits on the core count and does not allow you exceed it. Products like OracleVM, Solaris, and AIX contain hard partition virtualization. Products like HyperV and VMWare contain soft partitions. With soft partitions, you need to pay for all of the cores in the cluster since in theory you can consume all of the cores. To be honest, most people don’t understand this and get in trouble with license reviews.
When we talk about cloud services, licensing is also important to understand. Oracle published cloud license rules to detail limits and restrictions. The database is still licensed on a per core basis. The Linux operating system is licensed on per server instance and is limited to 8 virtual cores. If you deploy the Oracle database or WebLogic server in AWS or Azure or any other cloud vendor, you have to own a perpetual license for the database using the formulas above. The license must correlate to the high water mark for the core count that you provision. If you provision a 4 core system, you need a 2 processor license. If you run the database for six months and shut it off, you still need to own the perpetual license. The only way to work around this is to purchase database as a service in the Oracle cloud. You can pay for the database license on an hourly or monthly basis with metered services or on an annual basis with non-metered services. This provides a great cost savings because if we only need a database for 6 months we only need to pay for 6 months x the number of cores x the database edition type. If, for example, we want just the Database Enterprise Edition, it is $3K/core/month. If we want 4 cores that is $12K per month. If we want 6 months then we get it for $72K. We can walk away from the license and not have to pay the 22% annual maintenance on the $95K. We save $23K the first year and $20K annually by only using the database in the cloud for six months. If we wanted to use the database for 9 months, it is cheaper to own the license and lease processor and storage. If we go to the next higher level of database, Database High Performance Edition at $4K/core/month, it becomes cheaper to use the cloud service because it contains so many options that cost $15K/processor. Features like partitioning, compression, diagnostics, tuning, real application testing, and encryption are part of this cloud service. Suddenly the economics is in favor of cloud hosting a database rather than hosting in a data center.
Let’s go back to the Cisco UCS and network attached storage discussion. If we purchase a UCS 8 blade server with 32 cores per blade we are looking at $150K or higher for the server. We then want to attach a 100 TB disk array to it at about $300K (remember the $3K/TB). We will then have to pay $300K for VMWare. If we add 10% for hardware and 20% for software we are at just over $1M for a base solution. With a three year amortization we are looking at about $330K per year just to have a compute and storage infrastructure. We have to have a VMWare admin who doles out processors and storage, loads operating systems, creates golden masters, and acts as a traffic cop to manage and allocate resources. We still need to pay for the Oracle database license which is licensed on a per core basis. Unfortunately, with VMWare we must license all of the cores in the cluster so we either have to sub-divide the cluster into one blade and license all 32 cores or end up paying for all 256 cores. At roughly $25K/core that gets expensive quickly. Yes, you can run OracleVM or Solaris on one of the blades and subdivide the database into two cores and only pay for two cores since they both support hard partitioning but you would be amazed at how many people fight this solution. You now have two virtualization engines that you need to support with two different file formats. No one in mass wants two solutions just to solve a licensing issue.
Oracle has taken a radically different approach to this. Rather than purchasing hardware, storage, and a virtualization platform, run everything in the cloud and pay for it on a monthly basis. The biggest objection is that the cloud is in another city and security, latency, … you get the picture. The new solution is to run this hardware in your data center with the Oracle Public Cloud Machine. The cost of this solution is roughly $260K/year with a three year commit. You get 200 plus cores and 100 ish TB of storage to use as you want. You don’t manage it with VSphere but manage it with the same web page that you manage the public cloud services. If you want to script everything then you can manage it with REST apis or perl/java/insert your lanaguage scripts. The key benefit to this is that you no longer need to worry about what virtualization engine you are using. You manage the higher level and lease the database or weblogic or SOA license on an hourly or monthly basis.
Next week we will move up the stack and look at database hosting. Today we talked about infrastructure choices and how it impacts database license cost. Going with AWS or Azure still requires that you purchase the database license. Going with the Oracle public cloud or public cloud machine allows you to not own the database license but effectively lease it on an hourly or monthly basis. It might not be the right solution for 7x24x365 operation but it might be. It really is the right solution for bursty needs list holiday peak periods, student registration systems, development and testing where you only need a large footprint for a few weeks and don’t need to buy for your highwater mark and run at 20% the rest of the year.