INTRODUCTION TO SQL PERFORMANCE TUNING
A few days ago, with a little spare time on my hands I was sitting at a clients site skipping through some performance views on a development server when a statement jumped out, grabbed me by the throat and started to strangle me, I was fighting for breath, dripping with sweat and beginning to feel the life draining from my body.
I could smell burning, was it nausea from the lack of oxygen... and what was that cackling laugh I kept hearing? Determined to find out I managed to extricate myself from the murderous statement and looked on from a distance.
Clearly this statement had military training - it was fighting on two fronts! As well as suffocating me it was trying to kill the server, merrily ripping ram chips from it and cackling away as they were destroyed, and was that glue I could see dripping from the drive arrays causing them to run at a crawl? Surely not, in between ram chips, this evil little statement was pumping glue into the drives.
At this rate, the server would soon be on it's knees begging. No wonder it was so slow, what with it's diminished ram capacity, and constant struggle against the rapidly hardening glue, this server was in a fight to the death with this statement.
This intrigued me, Oracle servers have highly trained, highly motivated and adaptable armies of men who can jump to it's defence against statements like this. They can analyse it from a distance before it even gets close to their defences and prepare to outmanouvre it, adapt to it's offensive capabilities and spit out the remains in a blink of an eye.
So what on earth was this statement, I had to find out. It must be a really complex statement, multiple sorts, table scans, complex inner and outer joins, I could feel this was going to be a biggie. I assembled some trusted allies and went deep into enemy territory, prepared to meet my maker in this behemoth of a statement.
SELECT MPRN FROM TLM_CONSUMERS WHERE ZGC = 21991022Hardly a battle hardened scary thing is it?
The need for tuning, a simple example
Taking the above statement and checking the table, there was a unique index on the ZGC column. So what was the problem?
I obtained an explain plan for the statement and sure enough Oracle was doing a full table scan for this statement, each and every time it was run. The statement had been executed 331 times over the past few days and caused disk reads of 180 Gigs. In fact this statement and its multiple executions were responsible for approximately 90% of the workload on the Server. Here was the glue in the hard drives.
Quite simply, the ZGC column (and therefore the index) is a VARCHAR2 column. The value being looked for was a numeric and so Oracle didn't use the index. It was performing a full table scan on 1.2 million rows and had done so 331 times over the past few days.
To put that in context, Oracle was having to physically read an average of 200Mb per minute, every minute just to satisfy this statement. I imagine the execution time of this statement was somewhere in the 10's of minutes.
A simple change to the statement of adding single quotes around the value like this
SELECT MPRN FROM TLM_CONSUMERS WHERE ZGC = '21991022'resulted in a lightning fast execution time as would be expected, and almost zero workload on the server. Not only would the user or process executing the statement notice the difference, everyone using that server would notice the difference.
Poor SQL or a slow server... who can we shout at?
I once worked with a developer who had a written a piece of PL/SQL to perform a task, it was taking three hours to complete. There was absolutely no way to account for this as the volume of data on the box was so small you simply could not imagine any possible way for anything to take so long.
Having a quick look at the code, I pointed out to him where the code could be tuned and where the code was badly inefficient. His view was "I'm paid to make it work, not to make it fast" and left it. His entire view was quite simply that performance was not his problem. This was the same developer who constantly complained about how slow Oracle is.
Had this code made it to production, there would have been 500 users also wondering why Oracle was so slow. The reality was the 'couldn't care less developer' moved on and one of the other developers spent 15 minutes tuning the code. It resulted in a job which ran in 10 minutes and left the users with a pleasant, softly focused, cosy feeling about Oracle.
So who should we shout at? The developers? No, not necessarily. The murderous statement I mentioned in the first section above was actually generated by a report designer tool. The developer of the report had little or no control of the SQL it generated. In fact it was a deficiency of the tool in that it was not correctly querying the Oracle dictionaries before constructing it's SQL.
Statements like that appear for many reasons, maybe someone has dropped an index from a table and what was once a highly efficient statement turns into a dog.
Maybe that couldn't care less developer has just started on your site?
Maybe automatically generated code which people have no control over is grinding the server to it's knees.
In my experience most developers are conscientious people who take pride in their work and would readily spend a few minutes tuning their SQL for performance. The problem often lies in a lack of knowledge of exactly where to start, or from a foundation in some other desktop database where SQL optimisation was not possible. Hopefully this document will aid those in their quest for better performance.
The purpose of this document is not to offer a complete, detailed look at SQL tuning and what Oracle offers - the Oracle documentation already does that. The problem with the Oracle documentation is that it is so detailed you would have to read it for six months to get a grasp on how to tune your statements.
Instead, this document is aimed at those getting started in Oracle/SQL or those developers who have not yet ventured into tuning their statements. Hopefully it will provide a fairly easy route into the fascinating area of Oracle performance.
Back to Contents