On the subject of I/O

One the things that constantly surprises me when talking with clients about hardware for a new database server is that I/O is always at the bottom of the list. Typically the list will look something like this (listed in order of perceived importance)
  • CPUs, have we enough. Fast as possible.

  • Memory, as much as we can put in the box. Oracle don't charge us for that

  • SAN, big as possible.

At this stage the purchase order is usually given the nod and the hardware supplier will ship yet another run of the mill box. Don't get me wrong. Many experienced DBAs have been through this process many times before and realise that not only is the list in the wrong order but its missing some critical components.
  • HBAs, need to specify these in proportion to the CPUs and attached storage

  • NICs, might need a lot of these i.e public, cluster interconnect, storage, management, backup. And typically in multiples for resillience or performance.

  • Backup, are we using the existing backup infrastructure?

I don't blame anyone for this way of thinking, its the way its always been. When discussing a new server the first question that people tend to ask is "So whats this monster packing? 16 CPUs!!!" followed by lots of very macho grunts and hollering. The standard licensing model (not just Oracles) doesn't help. It starts with premise of a CPU describing the power of a server, and to a large degree it does but misses the point of what a database is all about and that's information. Typically that information is held in ones and zeros on a bunch of spinning scrap metal. The real power of a database comes from its ability to aggregate, analyze and process those ones and zeros, turn it into information and push results out to interested parties. Paraphrasing a little "Its all about I/O stupid".

With this in mind I'm constantly surprised by the imbalance of I/O put into servers both disk and network. Its not unusual to see a 4 cpu server running with the latest generation Intel and AMD CPUs but with a single HBA and dual ported NIC. Whilst memory is cheap many of these servers still run 32 bit kernels. This typically means only a small proportion of the database is cached in memory be it in the SGA or file cache (don't me started on file cache). I'd make a rash guess that whilst the size of the memory in a typical database server has increased the average size of the SGA hasn't increased in line with this trend. To make matters worse the typical size of a database has got significantly bigger. This has to lead us to the conclusion that less of the database is cached and as a result a bigger proportion of its is located on disk. As I said this is just a guess but its backed up with real customer engagements. What would be of interest is to have performed an analysis over the last 10 years to see if the wait event for scattered and sequential reads had decreased or increased as a proportion of the total wait event in production databases.

What I'm driving at is the need to move I/O way up the agenda when sizing a server for databases. The number of CPUs needs to be married to the number of I/O channels available. It makes no sense to buy database licenses for a machine that will simply sit and wait on I/O, Its simply wasting money. Equally it makes no sense to stuff a 4 cpu machine full of HBAs for a database application that will perform index lookups on a index that fits comfortably in the cache. Adding HBAs later to an existing server isn't necessarily a simple option either especially for a mission critical application or one that has hard coded paths to disk.

The next obvious question is "well thats well and good but how do I size the ratio of HBAs to CPUs." and in a typically vague fashion I reply "well that depends". The type of application and the type of processor should heavily influence the decision. Certainly the CPU has been winning the race in terms of performance over the last few years and it needs a lot more I/O to keep it busy. But the equation also needs to be balanced with the amount of memory available on the box. A large SGA will certainly reduce the need to visit disk. The best advice I can give is to speak to your hardware supplier and find out what the current state of play is. Also check the latest TPC-C and TPC-H figures show. Whilst these are generally edging towards the extremes of performance it does show what a hardware supplier believed was needed to show their hardware in the best light.

blog comments powered by Disqus