Hello,

On Mon, Mar 26, 2018 at 12:47:02PM +0300, Slawa Olhovchenkov wrote:
> On Mon, Mar 26, 2018 at 12:38:57PM +0300, Slawa Olhovchenkov wrote:
> 
> > > Could you check with your previous configuration (nbproc 2 and no 
> > > nbthread parameter) ? It could help to know if it is a threads related 
> > > problem or not.
> > > 
> > > Then, try to disable kqueue (nokqueue option in global section or the 
> > > command-line argument "-dk"), just for the test.
> > 
> > 1. disable kqueue. CPU consume less (from currently 100% per core to
> > 80% per core). Shrink user part, system about same.
> > 
> > 2. nbproc 2
> > 
> > global
> >         nokqueue
> >         nbproc 2
> >         cpu-map auto:1/1-2 0-1
> >         stats bind-process 1
> > 
> > Totaly trashed
> > 
> > 3. nbproc 2 + kqueue -- same as 2)
> > 
> > https://postimg.org/image/4ple6ujfr/
> > 
> > 12:21 (1)
> > 12:28 (2)
> > 12:31 (3)
> > 
> > 1.7.10 work good. This is regression in 1.8
> 
> https://postimg.org/image/tk4w189if/
> 
> 12:38 shutdown 1.8 and start 1.7 (start done at 12:40).
> 
> red line is 'sys' cpu time
> green line is 'user' cpu time

This is very useful, thank you. I'm seeing overall that when you're on
1.7.10+kqueue and 1.8.5+poll the overall %user is the same. However
it's the system which makes a huge difference there (to be expected
with poll under load). However things are highly different when you
run with kqueue, this seems to indicate that we may be missing some
events in the kqueue loop or that we need multiple calls for some of
them to be processed, causing some suboptimal behaviour. There is
some important locking in 1.8 to deal with the fd cache which has been
revisited differently in 1.9. And the differences are important enough
not to be realisticly backportable. So it may be possible that what you
observe is a limit of the current 1.8 code, though we'll take a closer
look in case we find something. However since we haven't yet heard of
anything similar affecting epoll, it possibly is an issue affecting
only the kqueue code, so there's hope that we can spot something there.

thanks,
Willy

Reply via email to