Oracle Exadata X5: The Road To Ten Billion Dollars
February 4, 2015 21 Comments
Now that the dust has settled on the announcement of Oracle’s new Exadata X5 Database Machine, I’ve been doing some research in order to update my History of Exadata post (it’ll be ready soon). While reviewing the datasheets and other collateral for the X5 I was struck by the meteoric increase in one particular statistic: the number of processor cores on each database server. Oracle is riding that Moore’s Law train all the way to the bank.
The thing is, the number of cores per database server is directly linked to the cost of licensing the Oracle Database for each Exadata machine – and that means trouble if you are the one paying the bills. Assuming you buy a full rack – and that you license every core in every database server (which is the most common choice, since only a very brave minority would consider the Oracle VM Trusted Partitions option), your license cost has been increasing by 50% for each of the last two Exadata releases (X3-2 to X4-2 to X5-2). Let’s have a look at that in graphical form (click on the image to enlarge):
Let’s not forget here that I am only plotting the cost of licensing the database software. We are not taking into account the extra costs associated with paying for the hardware, licensing the storage servers or purchasing any of the pretty-much essential enterprise edition options such as Oracle RAC, partitioning, multitenancy or the diagnostic pack license. Nor are we considering the infamous 22% per annum software support costs. I’m also using the list price – which you would never expect to pay – but even if you used discounted prices the percentage increases would remain the same.
Anyway, after crunching the numbers it turns out there is good news and bad news…
Good News for Oracle
The good news for Oracle is that if Exadata continues to increase the number of cores at the rate of 50% extra per release, the list price for licensing the database software on the future X23-2 model will be $10.1 billion:
There’s no doubt about it – this will pay for a lot of yachts.
Bad News for Oracle
There is a downside though. As Kevin Closson has already shown, Oracle appears to be having trouble balancing the I/O capabilities of the X5 against this tremendously-increased compute power. Even after abandoning previous claims that a memory hierarchy with “low-cost disk” as the bottom tier would bring the “highest performance at the lowest cost” to customers, the new Oracle Exadata X5 “Extreme Flash” model (because apparently flash on its own isn’t enough, it has to be extreme flash) struggles to deliver an improvement in IOPS per flash device, as Kevin has shown.
Excerpt from the Oracle Exadata X5-2 Press Release
You wouldn’t think this from reading the press release, which promises “breakthrough performance and price per I/O” (emphasis added by me). Price per I/O, eh? That sounds like we take the overall price and divide it by the number of IOPS the system can deliver, right?
So let’s do that. And to be generous, let’s only look at the database license cost (ignoring all those other costs again) and take the maximum IOPS number from the datasheets (which in most cases is for unrealistic, 100% read only, fully cached workloads). How will it look? I’ll overlay it as a line (in blue) on top of the first graph:
Well it turns out that the price per I/O is actually falling: it’s down from $1.71 on the X4-2 (HP) model to $1.65 on the X5-2 EF. But three and a half percent is not much of an improvement considering the 50% extra on the price tag, is it? And as for offering a “breakthrough” price per I/O, the X2-2 was better at only $1.52 per I/O per second!
Summary
In my view, something is not right about the balance between compute and storage on the Exadata X5. It feels as though Oracle is bumping up the compute power faster than would be architecturally prudent because this results in a higher purchase price. Maybe I’m wrong – I have no insider knowledge and can only speculate… but when the Exadata X23-2 finally comes out in a couple of decades time, maybe we’ll know for sure.
Footnote
The comments section of this article makes for interesting reading, with responses from a number of Oracle employees – although not necessarily speaking on behalf of the mothership. The noteworthy (not to mention calm and measured) comments come from the Exadata Product Manager, Gurmeet Goindi. In addition I’d like to draw your attention to the following two URLs:
flashdba 
One remark, you can also use COD without using ovm trusted partition option.
The minimum to license is 40% of the machine.
But indeed, Oracle seems to favor processors with high core numbers over processors with higher frequency.
Not only with exadata, but also for their x5 servers themselves.
Capacity On Demand (which I believe Oracle now calls “Oracle IaaS Private Cloud”) wasn’t really designed for that purpose, it is a way of paying for Oracle products via operational expenditure rather than capital expenditure. Thus it allows Oracle to tap into a different pool of customers – including cloud providers themselves. However, I take your point Freek and am happy to acknowledge that it’s a genuine way of “subletting” an Exadata configuration. I just don’t think many customers will use a procurement option to fix a technical problem.
Hi,
I’m not sure if we are talking about the same thing.
During 2 different presentations given by oracle to customers, the COD thing was very explicitly mentioned as a way to limit the amount of processors that needed to be licensed.
Stating that with the X5-2, the minimum amount of processor licenses was now lower than with the X4-2 (40% of an 1/8 rack).
Customers however did need to have either platinum support or have the configuration manager installed (in connected mode).
Is it this?
Click to access engineered-systems-iaas-ds-1897230.pdf
If not, let me know the URL and I’ll link to it.
All the same, my point is about the cost of licensing the whole machine. It goes up 50% per iteration with little discernible benefit in cost per IOPS.
hmm, not sure
I had the impression this Exadata COD was not “dynamic”, but just a way to limit the number of cores with the same mechanisme as with an 1/8th rack (you can enable additional cores, but not disable them).
I have not received the presentation slides and can’t directly find an offical document describing it.
I will try to get some more (official) information about this and update later on.
No C-o-D is C-o-D not “Oracle IaaS Private Cloud” they are different things. you are wrong – in stating ‘) wasn’t really designed for that purpose’. You seem to have been away from Oracle for too long.
C-o-D explicitly exists to allow Exadata customers to license just the CPU they need.
As Freek says – the Capacity-on-Demand feature kinda blows your making-money-through-increased-cores argument out of the window. Maybe a little more research is required on your behalf. Waiting to read the next installment from a storage vendor.
You are absolutely right. I’ve been such a fool. Thank you for helping me to see the light. If it wasn’t for clear-minded, unbiased readers like you I’d be …
Oh wait. No, wait, hang on…
Your IP address resolves to: Oracle Corp UK Ltd, Thames Valley Park, Reading, UK.
In a world of opinions, none of us are unbiased. But at least I display my company logo in a big box on the top right so that everyone knows where my allegiance lies.
everyone has a bias – even you
Now, the capacity on demand will limit the number of processor licenses, but I’m still crossed about the tendency to pick processors with high number of cores vs processors with high frequency.
> “the tendency to pick processors with high number of cores vs processors with high frequency”
Bingo.
I think Ive figured out what “IOPS” really means in [comment truncated]
No Phil, I’m not publishing one of your infamously-long comments on this site.
Hi Flash,
Database licensing on Exadata is a lot more versatile than stated in this post. With Capacity on Demand and Trusted Partitions, customers have a lot of options at their disposal as to how they deploy their Oracle assets on Exadata Database Machine.
This post discusses all these options in some detail https://blogs.oracle.com/dr/entry/capacity_on_demand_and_trusted
Gurmeet, in the light of comments by other Oracle employees, may I say thank you for your calm and measured response.
Hi there,
As an Oracle employee (although my comment is my own and not that of my employer) I would be thrilled if we could charge $10 billion for a single machine. But do you really find it realistic that a single CPU would have 26602 cores in 18 releases?
I’ve always appreciated reading thoughts, ideas and reviews from other people, biased or not.
I find the various accusations of bias (some insinuated, others more direct) very interesting given that they tend to come from Oracle employees.
I work for Violin Memory, a flash storage vendor who could be considered a competitor to Oracle’s Exadata product. I believe – as I think my entire blog conveys – that flash storage on open systems is (in general) a better solution that Oracle Exadata. This obviously makes me biased – apart from anything else I am subject to the dreaded “confirmation bias”. Likewise, employees of Oracle are to be considered biased towards the hand that feeds them.
Surely it is the natural way of things that all of us vendors do our best to explain the reasoning behind our beliefs and biases in the most articulate way we can; highlighting our claims and, when necessary, shining a light on the claims of others if we believe them to be false. Only then can the customers decide whose truth they wish to believe.
So Martin, if that makes me biased then I’m happy to agree with you and your colleagues. I just don’t think we can agree on whether it’s a positive or a negative thing.
Hi there,
It is obvious that new generation brings performance improvements, sometimes more and sometimes less such as in case of 18-core CPU is less than it was before. Also depends how are performance measured: end-to-end or components (compute, storage, network) based…
I am interested into a new capabilities and features of X5 that makes this product a bit more faster or competitive on a market. Is it marketing or it could be beneficial to customers use of features such as: CoD, NVMe, proprietary protocols for direct access to infiniband, software features such as pure columnar flash cache.
Hello Sreten my old friend 🙂
CoD isn’t a feature, it’s a way of offsetting Oracle’s licensing requirements to try and make entry level Engineered Systems more attainable. NVMe is a protocol that offers benefits over older methods of communicating with solid state drives, but it isn’t unique to Exadata. I don’t really know much about this new “Exabus” connectivity that Oracle is promoting with the X5 models, or whether it’s actually a new thing rather than just a marketing label for connecting Exadata to Exalogic over the Infiniband network. Nor do I know much about the columnar flash cache – sorry.
What I do know is that for open systems (i.e. non-Exadata) compute power is continuously increasing as it tracks Moore’s Law. Memory is getting denser. Storage is getting faster and denser as well as closer to compute. And networking bandwidth is increasing: 40GE is already in town and 100GE is not far behind. All of these things, in my opinion, combine to make the design of Exadata less and less relevant going forward. Sure, I buy into the idea of converged infrastructure and the unified point of support, but the actual architecture of having separate compute and storage grids? I don’t see where the need is… and I certainly don’t see it justifying the cost.
But that’s just me. An army of Oracle employees are out there just waiting to disagree.
Hope to see you again soon Sreten.
HI Flash,
I can’t agree more mate, you hit the …….. what they say on the head, i remember times before ADSL all the telecom gaints focused on fiber optics and laid cable and once ADSL1 , 2 came along and lot of this fiber optics lines where free and so evolved the paid telephone cards.
i feel the same is going to happen this time with the INMEMORY database .
What do you think ????
Cheers,
Raj
Pingback: Oracle didn’t buy Sun just to sue Google but to keep it away from IBM, CEO says (ORCL, GOOG, GOOGL) | IOT POST