Where else are performance improvements seen?
Shetler: Anything that sort of processes large amounts data like I described where they are reading a whole table to do a report or maybe multiple table joins or things like backups. Backups touch a lot of data. We can optimize the performance of backup. We have had situations where a customer used to take many hours to do a backup an they have reduced it now down to minutes and they actually think it is not running correctly because it ran so fast. A lot of folks went from once-a-week backups on the weekends to every-night backups so they kind of change the dimensions of how they are doing database processing because of the performance they are getting.
Other folks are doing ad hoc queries or analytics during the day, and they used to launch something and go off to lunch and they'd come back and look at the results and they'd do that maybe one more time during the day. Now they iterate many, many times, even within a single hour, and so that is how that performance really has impacted what they are doing.
What else does Exadata do differently?
Shetler: The other big impact is in the advances we have made in compressing data so tightly that they can now store and analyze so much more data than they were able to store previously — and save a lot of money on future disk purchases as well. We have this technology that we call Hybrid Columnar Compression. On net, instead of being able to compress data by a factor of 2x or 4x versus non-compressed data, we are now able to see 10x or 15x compression.
Exadata also can process this data without decompressing it, so once you compress it down and you do a query against that table for example, we can analyze the data in its compressed state — and only when we get to the point where we say, this is the data I want, this is the stuff I'll return, then we do decompression. And so with Hybrid Columnar Compression we actually see a performance improvement because what used to be 10 block I/Os to get uncompressed data, now is a single block I/O, so we are doing much less I/O than before because the data is not only really compressed but it stays compressed for most of the time. Most people think that if you are doing compression you must lose performance because you have to burn all this CPU to decompress it, and although that is true, you do have to burn CPU to decompress it, what you gain in performance on I/O savings actually overwhelms what you lose in extra CPU that you are burning. And so, compression not only saves you a lot in disk storage, which translates directly to hardware investment, but it actually gives you a performance boost.
Performance and compression are two major innovations.
Shetler: And then the third thing that is innovative is our use of flash storage. Flash is roughly 30 times faster than a spinning disk to find a record. That translates directly to faster response times particularly for things like transactional applications.
With flash what we have done in Exadata is we have again integrated that into the database and, in an intelligent fashion — we call it Smart Flash Cache. With flash since it is more expensive — a lot more expensive than disk — you want to make sure that when you use it to store data, you are putting data in there that is very likely to be accessed with an I/O request right after that, or subsequently.
We say, we know why you have asked for this data and we can see the patterns that are happening on the system right now and so we are only going to put the data that we think is likely to be accessed by applications after this into flash. Because we only put data into flash that we think is likely to be accessed subsequently, we get more benefit out of flash than just the mindless dumping of data from disk into flash.
Those are the three things: The Smart Scan for the good performance underneath; the compression for better performance and better capacity on the disk; and then the integrated flash into the database for very speedy response times and OLTP support.
What do you expect going forward in terms of the market for appliances?
Shetler: We think that the model around the integrated systems — we call them engineered systems; we don't really call them appliances — that that model is now well-proven and it is a good model going forward for other areas. You can imagine a lot of different kinds of system usages that would be attractive in this appliance-like model where you have prebuilt everything and you have got the right software and the right hardware and you have tuned it all to work well as a combined engineered system. We see that going forward. It's hard for me to predict what other categories that will show up in, but I know there are other efforts here and probably elsewhere to figure that out, and come out with these integrated systems that work better because of that. And they will all be attractive because the customers haven't had to figure all that out on their own.