Am I understanding the output of `top` correctly?

zQUEz · 01-20-2012, 09:14 AM

I would like to get opinions on whether my interpretation below is accurate;

Assume the following scenario:

Intel server, RHEL5, 4 core cpu with hyperthreading turned on, thus, /proc/cpuinfo represents 8 CPU's.
Aside from standard Linux background processes, assume a single user process is running - say MySQL or other.
With 8 CPU's, the default header in `top` will represent a single threaded user space process using 100% CPU utlization as 12.5%us. (100% / 8 cores = 12.5)
But, in the process list, that same process will show 100% CPU because it is using it's full CPU thread.

If that same process is completely waiting on disk I/O, then the header of `top` will show a 12.5%wa and basically nothing on the CPU usage for us or sy. If this true, does this represent an application issue that it doesn't use multiple CPU's to satisfy disk I/O? Or is this behaviour normal?

My understanding is that, disk i/o is handled by kswapd, and if the requested blocks aren't already in file cache, then it is ultimatley a single threaded process to retrieve the block from disk, and if that is true, doesn't this mean in this example you would never seen wait's higher than 12.5%?

It seems to me difficult to see if you are disk bound from this aspect, because if your server usage basically comes does to a single thread that is I/O hungry, you won't ever seen high %wa in any sar reports or other.

Have I completely missed the mark on understanding this?

Nomad-71 · 01-22-2012, 01:05 AM

Hi, as i think:

Code:

wa  --  iowait  Amount of time the CPU has been waiting for I/O to complete.

Quote:

If this true, does this represent an application issue that it doesn't use multiple CPU's to satisfy disk I/O?

It represents that your disk is not fast enough to satisfy your demands.

zQUEz · 01-23-2012, 10:04 AM

Quote:

It represents that your disk is not fast enough to satisfy your demands.

While I agree with that, I would have expected wa to to show closer to 100% because 100% of that thread is waiting for disk. If I had say 4 threads running, but 3 were not io bound and only 1 was, I would still show 12.5% wa, but would really have no idea whether it was 1 thread causing all that io or that was generated by the 4 threads.

However, thanks for your answer because it tells me I must at least be interpreting the output of `top` correctly. I guess that is why you can't use any one tool in isolation.