You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@trafficserver.apache.org by zw...@apache.org on 2014/01/02 23:36:03 UTC

[2/4] git commit: TS-1365 Update documentation

TS-1365 Update documentation


Project: http://git-wip-us.apache.org/repos/asf/trafficserver/repo
Commit: http://git-wip-us.apache.org/repos/asf/trafficserver/commit/a74ef5a0
Tree: http://git-wip-us.apache.org/repos/asf/trafficserver/tree/a74ef5a0
Diff: http://git-wip-us.apache.org/repos/asf/trafficserver/diff/a74ef5a0

Branch: refs/heads/master
Commit: a74ef5a049347a864a0597364d3575ec297eeeb5
Parents: eccb33c
Author: Leif Hedstrom <zw...@apache.org>
Authored: Thu Jan 2 15:18:52 2014 -0700
Committer: Leif Hedstrom <zw...@apache.org>
Committed: Thu Jan 2 15:30:53 2014 -0700

----------------------------------------------------------------------
 .../configuration/records.config.en.rst         | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)
----------------------------------------------------------------------


http://git-wip-us.apache.org/repos/asf/trafficserver/blob/a74ef5a0/doc/reference/configuration/records.config.en.rst
----------------------------------------------------------------------
diff --git a/doc/reference/configuration/records.config.en.rst b/doc/reference/configuration/records.config.en.rst
index c92f351..08aaf1a 100644
--- a/doc/reference/configuration/records.config.en.rst
+++ b/doc/reference/configuration/records.config.en.rst
@@ -2225,6 +2225,26 @@ Sockets
 
    Same as the command line option ``--accept_mss`` that sets the MSS for all incoming requests.
 
+.. ts:cv:: CONFIG proxy.config.net.poll_timeout INT 0
+
+   Same as the command line option ``--poll_timeout``, or ``-t``, which
+   specifies the timeout used for the polling mechanism used. This timeout is
+   always in milliseconds (ms). On Linux, this is the timeout to
+   ``epoll_wait()``. The default value is ``10`` on all platforms except
+   Solaris, where it is ``30``.
+
+   Changing this configuration can reduce CPU usage on an idle system, since
+   periodic tasks gets processed at these intervals. On busy servers, this
+   overhead is diminished, since polled events triggers more
+   frequently. However, increasing the setting can also introduce additional
+   latency for certain operations, and timed events. It's recommended not to
+   touch this setting unless your CPU usage is unacceptable at idle
+   workload. Some alternatives to this could be::
+
+        Reduce the number of worker threads (net-threads)
+        Reduce the number of disk (AIO) threads
+
+
 Undocumented
 ============
 


Re: [2/4] git commit: TS-1365 Update documentation

Posted by Theo Schlossnagle <je...@lethargy.org>.
You can change it to 10ms.   However, if we’re reliably using net_signal_hook_function to wake up the event thread, it’s not very relevant as that will immediately cause a return.  Consistency is valuable, I’m +1 on changing it to 10ms.

--  
Theo Schlossnagle
http://lethargy.org/~jesus/


On Friday, January 3, 2014 at 12:13 PM, Leif Hedstrom wrote:

> On Jan 2, 2014, at 5:03 PM, Igor Galić <i.galic@brainsware.org (mailto:i.galic@brainsware.org)> wrote:
>  
> >  
> >  
> > > + ``epoll_wait()``. The default value is ``10`` on all platforms except
> > > + Solaris, where it is ``30``.
> > >  
> >  
> >  
> > *Why* is it different on Solaris?
>  
>  
> Fine question. I don’t know. Theo or anyone else using Solaris: Why would we want the timeout to kevent() to be 30ms instead of 10ms as we do for Linux? Can / should we change it to 10ms across the board?
>  
> — Leif  


Re: [2/4] git commit: TS-1365 Update documentation

Posted by Leif Hedstrom <zw...@apache.org>.
On Jan 2, 2014, at 5:03 PM, Igor Galić <i....@brainsware.org> wrote:

> 
> 
>> +   ``epoll_wait()``. The default value is ``10`` on all platforms except
>> +   Solaris, where it is ``30``.
> 
> *Why* is it different on Solaris?


Fine question. I don’t know. Theo or anyone else using Solaris: Why would we want the timeout to kevent() to be 30ms instead of 10ms as we do for Linux? Can / should we change it to 10ms across the board?

— Leif

Re: [2/4] git commit: TS-1365 Update documentation

Posted by Igor Galić <i....@brainsware.org>.
> > +   ``epoll_wait()``. The default value is ``10`` on all platforms except
> > +   Solaris, where it is ``30``.
> 
> *Why* is it different on Solaris?

165f3686 -> 
a5345da8 -> https://svn.apache.org/viewvc?view=revision&revision=r915922

is that supposed to not have any merge tracking?

-- 
Igor Galić

Tel: +43 (0) 664 886 22 883
Mail: i.galic@brainsware.org
URL: http://brainsware.org/
GPG: 8716 7A9F 989B ABD5 100F  4008 F266 55D6 2998 1641


Re: [2/4] git commit: TS-1365 Update documentation

Posted by Igor Galić <i....@brainsware.org>.

----- Original Message -----
> TS-1365 Update documentation

This should've been in the same commit as the adding of the feature.

> Project: http://git-wip-us.apache.org/repos/asf/trafficserver/repo
> Commit: http://git-wip-us.apache.org/repos/asf/trafficserver/commit/a74ef5a0
> Tree: http://git-wip-us.apache.org/repos/asf/trafficserver/tree/a74ef5a0
> Diff: http://git-wip-us.apache.org/repos/asf/trafficserver/diff/a74ef5a0
> 
> Branch: refs/heads/master
> Commit: a74ef5a049347a864a0597364d3575ec297eeeb5
> Parents: eccb33c
> Author: Leif Hedstrom <zw...@apache.org>
> Authored: Thu Jan 2 15:18:52 2014 -0700
> Committer: Leif Hedstrom <zw...@apache.org>
> Committed: Thu Jan 2 15:30:53 2014 -0700
> 
> ----------------------------------------------------------------------
>  .../configuration/records.config.en.rst         | 20 ++++++++++++++++++++
>  1 file changed, 20 insertions(+)
> ----------------------------------------------------------------------
> 
> 
> http://git-wip-us.apache.org/repos/asf/trafficserver/blob/a74ef5a0/doc/reference/configuration/records.config.en.rst
> ----------------------------------------------------------------------
> diff --git a/doc/reference/configuration/records.config.en.rst
> b/doc/reference/configuration/records.config.en.rst
> index c92f351..08aaf1a 100644
> --- a/doc/reference/configuration/records.config.en.rst
> +++ b/doc/reference/configuration/records.config.en.rst
> @@ -2225,6 +2225,26 @@ Sockets
>  
>     Same as the command line option ``--accept_mss`` that sets the MSS for
>     all incoming requests.
>  
> +.. ts:cv:: CONFIG proxy.config.net.poll_timeout INT 0
> +
> +   Same as the command line option ``--poll_timeout``, or ``-t``, which
> +   specifies the timeout used for the polling mechanism used. This timeout
> is
> +   always in milliseconds (ms). On Linux, this is the timeout to
> +   ``epoll_wait()``. The default value is ``10`` on all platforms except
> +   Solaris, where it is ``30``.

*Why* is it different on Solaris?

> +   Changing this configuration can reduce CPU usage on an idle system, since
> +   periodic tasks gets processed at these intervals. On busy servers, this
> +   overhead is diminished, since polled events triggers more
> +   frequently. However, increasing the setting can also introduce additional
> +   latency for certain operations, and timed events. It's recommended not to
> +   touch this setting unless your CPU usage is unacceptable at idle
> +   workload. Some alternatives to this could be::
> +
> +        Reduce the number of worker threads (net-threads)
> +        Reduce the number of disk (AIO) threads
> +

We should link here to the appropriate settings that would require tuning,
or to the (still rather poor) performance tuning document.

> +
>  Undocumented
>  ============
>  
> 
> 

-- 
Igor Galić

Tel: +43 (0) 664 886 22 883
Mail: i.galic@brainsware.org
URL: http://brainsware.org/
GPG: 8716 7A9F 989B ABD5 100F  4008 F266 55D6 2998 1641