Exadata vs. IBM P-Series
Earlier this year I participated in a Total Cost of Ownership (TCO) study that was run by a company called the FactPoint Group. The goal was to compare the cost of purchasing and running Oracle on Exadata vs. the cost of purchasing and running Oracle on IBM P-Series hardware. The findings are published here:
Fair warning, the study was funded by Oracle and the first 10 minutes are a sales pitch for Exadata, but the data collected by FactPoint for the TCO study fills the rest of the presentation.. That part was very interesting (to me anyway). They interviewed people from 5 companies which had been running in production on Exadata for at least 6 months and 5 companies that had production systems running on IBM P-Series machines. As a bonus, 2 of the Exadata customers had previously run their systems on IBM P-Series hardware or currently had other systems running on P-Series machines.
The TCO calculations were based on equivalent systems from a CPU count basis. That is, they priced out an IBM P7 system that had the same number of cores as an Exadata Half Rack and used that for the calculations. I didn’t think that was really a fair comparison since the performance of the two systems was vastly different. In fact, one of the customers that had moved from IBM to an Exadata had the following to say:
“This data warehouse was originally on 12 a P570 and the nightly load took 5 hours. With Exadata it takes 30 minutes with the Exadata 1⁄4 rack running at only 5% CPU utilization. Weekly stats updating went from 60 hours to 35 minutes. Backups reduced from 14 hours to 45 minutes.”
This is only one of many quotes from the customers included in the presentation. The quotes were quite revealing, and to my way of thinking, were the most interesting part of the presentation. Another oddity of the cost calculations was that they included RAC licenses in the Exadata costs while not including RAC in the IBM costs. Again I felt that this was an unfair comparison as RAC provided HA capabilities to the Exadata platform that the IBM platform simple didn’t have. Nevertheless, the study found that Exadata was about 35% less expensive. I wonder what it would have been if they had tried to cost an IBM system that performed as well as an Exadata and provided the same HA capabilities.
Updated 11/23/2012 – Note that the official white paper produced as a result of this study is now available here: Cost Comparison – Oracle Exadata Database Machine vs. IBM Power Systems
Hi Kerry
When I hear such quote the only question to ask is “how this improvement was possible” – with the breakdown of what kind of work is done (serial/parallel) in the job, breakdown of IO/CPU parts and activity profile.
I have a strong feeling such improvements are possible either on the same or non-Exadata platform. Wait, I’m not an Exadata Expert. I’m not an IBM expert either. What do I know? 🙂
They had 12 boxes and no RAC? Sorry, I didn’t read the doc.
Hi Timur,
We have had a long term relationship with this particular customer but they are pretty self sufficient. We have done project based work for them over a 4 or 5 year time frame. They basically do a top to bottom infrastructure refresh every three years. This time around they decided to go with Exadata for the DW component. We only helped install and configure the machine and then pretty much walked away. So there wasn’t any “tuning” by us and I don’t think they made any significant changes either. We did do a POC with them before they made the decision so they had an idea what kind of benefit they would get – but I get the impression they were still very surprised with how well it worked (being skeptical like all of us are about marketing fluff). I find it interesting to hear these kinds of quotes because we often aren’t around long enough to hear how the story really ends.
With regard to the RAC stuff, the quote was referring to the system they migrated from. (by the way, I think 12 is either a typo or a misquote. I have an email into the customer to validate what he actually said there) Anyway, my comment about RAC was just me complaining that for the cost comparison calculations in the TCO study, the FactPoint guys decided to spec an “equivalent” IBM P7 with the same number of CPUS as are present on the compute nodes of an Exadata Half Rack and without RAC. I don’t think that is really very close to “equivalent”. We’ve worked with a few companies going through hardware refresh cycles that did the same sort of exercise for themselves and they usually have the same options on both sides (RAC, partitioning, etc…) and they usually try to come up with systems that perform similarly (i.e. they usually add more CPU’s to non-Exadata platforms in an effort to make up for the work that they expect to be offloaded to the storage tier on Exadata). This could be almost nothing in a pure OLTP workload or a lot in a more DW type environment.
Yeah that was supposed to be a single P570. I’ve notified the FactPoint guys and modified it on my post.
Hello Kerry, it seems to me that you are a software guy that does not pay to much attention on the hardware part. Maybe you should in some way. P570 are old machines now, and one machine is not Twelve (now corrected) !
Lately I have discussed with Oracle consultants in offline conversations, they always told me that power systems are the best machines to run Oracle RDBMS and when I see where they look to get inspiration (Mainframe Coupling Facility and DB2 PureScale, based on mainframe CF too) plus LPAR facility, may be you should take a look too. Anyway I really appreciate your Oracle Expertise. Regards
Hi pcadiille,
I am definitely a software guy and I don’t really consider myself to be a hardware guy. However, I believe that everyone has their set of colored glasses that they look through. The color is made up of the experiences they have had over their lifetime. I’ve never been employed to design or build hardware. But I have spent my whole career doing work on machines that run Oracle so I don’t consider myself to be completely ignorant of the hardware side of our business either. I am well aware that P570’s are not the most current version of platform. I might also point out that the current generation of Exadata (X2-2) is not exactly bleeding edge at this point either. Westmere CPU’s are a couple of years old and the F20 flash cards are even older. The new X3 versions (which should be announced in the next few minutes) will be more in line with currently available hardware, sporting Xeon E5-2690 (Sandy Bridge) CPUs. By the way, we have several customers running Oracle on P7’s and several that have converted from P7’s to Exadata as well.
The particular customer story that I mentioned from the FactPoint TCO Study was doing a hardware refresh (which they do every 3 years). As a matter of fact, we helped them with the migration to the IBM boxes in the first place (off of Sun servers). The main point of the post was to notify people that the study exists (and I hope you’ll listen to the whole thing). I also mentioned that I find the quotes from real customers to be quite revealing. When someone says things like “My 5 hour job now runs in 30 minutes” and “My 14 hour backup now runs in 45 minutes” it gives a sense of how much difference the platform has made to them. I’ve personally never seen a hardware upgrade create that kind of difference and I’ve been working on Oracle systems since the mid 80’s. (that’s prior to Exadata of course).
Which brings me to my last comment. Moving to Exadata is not the same thing as just moving to faster hardware. There are some fundamental changes in how Oracle behaves on Exadata. That’s where the improvements come from, not the hardware. We frequently do a demonstration where we turn the Exadata software optimizations (offloading) off and run baselines. We then turn the features on one at a time (column projection, then filtering, then storage indexes) and watch the execution times drop like a rock. If you are at OOW you can come by the Enkitec booth and see Tim Fox do the demonstration in person. If you ask nicely, he may even let you log on to one of our Exadata’s and try some queries on your own. 😉
For DW, you may be right,But for OLTP,the main improvement is come from the flash and the Infiniband, the flash radically reduced the db file sequential/scattered read latency and the db file parallel write latency, and associated CPU cycles for read/write data,the Infiniband radically reduced log file parallel write ,so improved commit efficiency intensively. Otherwise,the CPU processing power of Intel X5675 and E2690 is a little powerful than the IBM power 5 series and Power 7 Series. Just compare the CPU power is a little flaw. If compare the performance, I recommend your to suggest your customer with the new Flash Storage with the old architecture for OLTP. New Flash vendors list : Kaminario, Nimbus storage,Texas Memory System,..
[…] point out that the official white paper version of the TCO study I talked about in a previous post (Exadata vs. IBM P-Series) is now available on the Oracle web site […]
[…] Kerry Osborne’s Oracle Blog for more details about the Exadata vs. IBM P-Series […]