On Thu, 2018-09-06 at 08:55 -0600, David Boreham wrote:
On 9/6/2018 8:50 AM, isabella.ghiurea(a)nrc-cnrc.gc.ca wrote:
> This does not justify this since running 1 tread
> takes 0.1564msec/op and running 10 threads takes 0.0590ms/op and
> the last one will require the access.log to be flush more
> frequently I think for 10 threads and I do not see the
> spike in exec time showed for 1 thread. Maybe something else ?
> _______________________________________________
>
I think, looking at the data you posted, the question you're asking
is
"why, when I subject my server to a continuous search operation load,
do
some operations have much longer latency than others?".
If they are doing the same operation repeatedly, then it *is* an issue
because it shows that within the server architecture there is a flaw
that causes non deterministic behaviour.
I have observed this behaviour myself, and traced it to logging (I did
a server build with log_* as a stub function).
But there are plenty of other places in the code which could be issues
- I know plenty of them, but time is the enemy of all things :)
--
Sincerely,
William