CPU run queue question

[Court Campbell]

For a few months I have been searching for a good answer as to what is a bad number of processes in the run queue. If anything the answer is it depends. But I would like to get a general clue of what is potentially bad. If I have a box with eight processors and my run queue is say 12 for a marked length of time, am I to assume this is bad?

I have read that the general rule is that your run queue should never be higher than the total number of cpu's in your machine. Any enlightening info would be great. I have read the CPU section of the HP-UX tuning and perf. book, but it does not give you anything specific number wise. If anything you only really learn that you need to baseline to make any future decisions about performance tuning.

[Bill Hassell · Accepted solution]

The reason that there is no number recommended is because "it depends". The run queue gives the number of processes that are currently runnable. For 8 CPUs, a run queue of 12 is just fine -- or it may be delaying all 12 processes. It depends.

The reason it is impossible to give an answer is that the run queue metric is not meaningful by itself. If you are running 500 processes that are all polling once every second (and the average run time for the poll is 2 milliseconds), the run queue might be 20 to 200 and your 8 CPU system will have no delays and no problems. top and Glance will show 95% idle time.

On the other hand, if you have 9 compute-bound processes that run for hours with no I/O at all, then your run queue of 9 will be delaying 8 of the processes as the 9th process is brought in and another is made idle since there aren't enough processors. Virtually all performance issues are much, much more complicated than this and for databases, disk I/O is by far the most important metric for performance.

So a run queue of 12 is fine. You have to determine if those 12 processes must run faster, and if true, add 4 more processors.

[Court Campbell]

Good answer. I was looking for clarification and I got it.