Buffer Cache Hit Ratio – my “guilty” Confession

My Friday Philosophy this week was on Rules of Thumb on buffer gets per row returned.

Piet de Visser responded with a nice posting of his own, confessing to using ratios to help tuning {We seem to be playing some sort of blog-comment tag team game at the moment}.

Well, I have a confession so “guilty” or “dirty” that I feel I cannot inflict it on someone else’s blog as a comment.

I use the Buffer Cache Hit Ratio.

And the Library Cache Hit Ratio and the other Ratios.

As has been blogged and forum’d extensively, using these ratios is bad and stupid and anyone doing so does not know what they are doing as they do not help you solve performance problems. I mean, hell, you can download Connor McDonald’s/Jonathan Lewis’s script  to set it to what you want so it must be rubbish {go to the link and chose “tuning” and pick “Custom Hit Ratio” – it’s a rather neat little script}.

The point I am trying to make is that once the Buffer Cache Hit Ratio (BCHR) was wrongly elevated to the level of being regarded as a vital piece of key information but the reaction against this silly situation has been that it is now viewed by many (I feel) as the worst piece of misleading rubbish. Again a silly situation.

I think of the BCHR as similar to a heart rate. Is a heart rate of 120 good or bad? It’s bad if it is an adult’s resting heart rate, but pretty good if it is a kitten’s resting heart rate. It’s also probably pretty good if it is your heart rate as you walk briskly. Like the BCHR it can be fudged. I can go for a run to get mine higher, I can drain a couple of pints of my blood from my body and it will go up {I reserve the right not to prove that last one}. I can go to sleep and it will drop. Comparing my resting heart rate to yours (so like comparing BCHRs between systems) is pretty pointless, as I am a different size, age and metabolism to you {probably} but looking at mine over a year of dieting and exercising is very useful. If only I could keep up dieting and exercising for a year…

So what do I think the much-maligned Buffer Cache Hit Ratio gives me? It gives me what percentage of sql access, across the whole database activity, is satisfied from memory as opposed to disc. Or, put another way, the percentage of occurences a block has to be got from the I/O subsystem. Not how many blocks are read from storage or memory though, but you can get that information easily enough. As Physical IO is several orders of magnitude slower than memory access {ignoring I/O caches I should add} , it gives me an immediate feel for where I can and can’t look for things to improve.

If I am looking at a system that is overall very slow (eg high process wait queues under l/unix, the client has said the system is generally slow) and I see that the BCHR is low, say below 90%, this tells me I probably can get some performance increase by reducing physical access. I’ll go and look for those statements with the highest physical IO and the hottest tablespaces/objects in the DB.
If the BCHR is already up at the 99% level, I need to look at other things, such as tuning sort, looking at removing activity in the database, to be very mindful of nested loop access where maybe it is not the best access method (very likely due to old stats on tables).

When I have got to know a system and what it’s BCHR generally sits at, a sudden change, especially a drop, means there is some unusual physical IO going on. If the phones start going and someone is complaining “it’s all slow”, the BCHR is one of the first things to look at – especially as it is available from so many places.

Another thing the BCHR gives me is, if I am looking at a given SQL statement or part of an application, it’s specific BCHR can be compared to the system BCHR. this does not help me tune the statement itself, but I know if it’s specific BCHR is low then it has unusually high IO demands compared to the rest of the system. Further, Reducing it might help the whole system, so I might want to keep an eye on overall system throughput. If I reduce the statement’s execution time by 75% and the whole system IO by 1%, the client is likely to be more happy, especially if that 1% equates to other programs running a little faster “for free”.

So, I don’t use the BCHR to tune individual statements but I feel confident using it to track the general health of my database, so long as I am mindful of the impact of new functionality or upgrades. It’s a rule of thumb. It’s a database heart rate. (and so is redo generation and half a dozen other things).