[geeks] A Linux Tech Support Question...

[Andrew Jones]

On 02/08/2013 11:17 AM, Jonathan Katz wrote:
>
> Device:         rrqm/s   wrqm/s   r/s   w/s   rsec/s   wsec/s avgrq-sz
> avgqu-sz   await  svctm  %util
> sda6              0.00  1254.33 39.33 73.00   605.33 10618.67    99.92
> 0.15    1.38   1.17  13.17
> sdb             141.33   289.67 243.67 29.33 53261.33  2658.67   204.84
> 4.26   15.73   2.48  67.70
> sdc              81.67   422.00 109.33 53.67 25730.67  3805.33   181.20
> 2.28   14.00   2.24  36.57
> dm-0              0.00     0.00 577.00 793.33 79077.33  6346.67    62.34
> 21.32   15.95   0.73  99.90

If you're IO-bound, it's not a throughput problem.

You were right to question this: the "%util" parameter can sometimes 
have funny values, since it's the kernel's estimate of what's going on.

The "avgqu-sz" and "await" are real, hard numbers, though -- queue depth 
and average latency to service an item in the queue.  They're really 
quite low on your box.

If your I/O devices were pounded for throughput, I would expect queue 
depths and wait times to be in the high 10s or 100s, not the low 10s. 
My completely idle RAID array has await in the 10 to 20 range.

Other possibilities for the IO-blocked processes: waiting for network 
traffic, locking implmented with files, just tons of blocking operations 
that are hurt by the 10-15 ms latency of hitting disk.