OOW2K7: More about Monday

Bryn Llewellyn's talk overran slightly - there is that much which is new about PL/SQL in 11g - so I only just made it to the Moscone Center for Nitin Vengurlekar's talk on Back-of-the-envelope Database Storage Design. The main thing I took from this talk was that we need to have big envelopes. It was a useful overview for me, as I knew nothing about this area of IT systems. For instance, I learned the difference between throughput and good-put; throughput is the total amount of network traffic whereas good-put is the amount of useful bits transferred i.e. throughput minus network protocol, retransmitted packets, etc. I now know that we can specify our requirements in either MBytes/second or IOPS but they are mutually exclusive. What I don't know is in what scenarios I should be using IOPS rather than MBytes. Transactions/s is a useless metric because it can mean many things to us database practitioners and it means nothing to storage dudes. The other particularly useful part of this presentation was a break down of all the layers of the I/O stack, identifying all of the components which could be a choke-point. A system's I/O is only as fast as its slowest component.

I had a quick wander through the Exhibition hall. I didn't pick up much in the way of loot but AccentureEDS has a really useless but potentially time-consuming geegaw: it's a sticky plasticky bundle of tentacles on a elasticated string - sort of a cross between a yoyo and a primary-coloured Cthulu.

The last session of Monday was Juan Loaiza's talk on Scalability and Performance enhancements in 11g. He opened with some recent benchmarks. He started with the bad boys - 4 million transactions per minute , 1 million physical I/O per second. He said they like these sorts of benchmarks because it's the only time they get to play with such enormous systems; the above mentioned benchmark used 2TB of RAM (which costs about $4m - still a lot of money even at current exchange rates). The interesting benchmarks are the price performance ones: the new ranges of Intel and AMD chips means that Oracle is now outperforming Microsoft in the low-end market as well as the high-end.

The point about the large system benchmarks is that they are the systems of the future. By 2010 Oracle predict there will be at least one 1 PetaByte database out there, at least one 1000 node RAC system and at least one system with 1 TeraByte of RAM. These huge systems are being driven by a number of different trends: government initiatives in healthcare and law enforcement; large internet retailers; and scientific projects in areas like genetic research and particle physics.

The challenge for Oracle is to make such huge resources work as well as smaller resources. For instance, we all know that disks are getting bigger but that their performance is not keeping up. We will soon have 1TB disks, which will be pretty much like tape drives. This is why table compression in 11g is so important. By compressing tables we effectively increase the performance of disks considerably (more data retrieved by each disk read). Furthermore the database can natively interpret the compressed data. Obviously there is a DML overhead but it is a fraction of the gains from improved reads. The problems inherent in backing up a 1PB database are obvious. This is why Oracle has written its own backup utility: the database can use its knowledge of its own state to only backup stuff that needs to be backed up. This can be a lot faster than blindly copying all the system files at the OS level.

It is also obvious why Oracle is still committed to per core licensing. The charges for a 1000 CPU cluster (Enterprise Edition plus RAC plus Partitioning plus all sorts of other 11g goodies) would add up to a hefty chunk of change.

In the Q&A somebody observed that Oracle's own cluster still uses Sun E25K and asked Juan when it is still appropriate to go for SMPs rather than blade clusters. Juan replied that it was a matter of comfort. SMP is a very mature technology but RAC is now mature; leading edge rather than bleeding edge. He did say that he didn't think Oracle would ever purchase another SMP ever again. Larry won't pay for them.

I bumped into Eddie Awad a couple of times yesterday and I kept asking him, "Are you Twittering this?" So at the OTN Night he pulled out his iPhone - Eddie just had to be an earlier adopter - and Twittered Dan Norris. Lo! a few minutes later Dan had located us outside the Jeopardy room. So it can be useful and I will stop being sceptical. OTN night was pretty much the same as last year, only with slightly fewer exotic dancing girls and the introduction of some people from the local D&D guild wandering about on stilts. I'm sure it made sense to somebody. Doug Burns introduced me to Kurt "DUDE" Van Meerbeck and his wife but it was really too noisy to have a proper conversation.