Message-ID: <19971031104637.52990@deejai.mch.sni.de>
Date: Fri, 31 Oct 1997 10:46:37 +0100
From: Martin Kraemer <Martin.Kraemer@mch.sni.de>
To: new-httpd@apache.org
Subject: Re: strange performance delays
References: <3.0.3.32.19971030195222.008ac800@hyperreal.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
In-Reply-To: <3.0.3.32.19971030195222.008ac800@hyperreal.org>;
 from Brian Behlendorf <brian@hyperreal.org> on Thu, Oct 30,
 1997 at 07:52:22PM -0800
Sender: new-httpd-owner@apache.org
Precedence: bulk
Reply-To: new-httpd@apache.org

On Thu, Oct 30, 1997 at 07:52:22PM -0800, Brian Behlendorf wrote:
> 
>... Once every 1000 requests, I
> see an even larger hit; the largest I've see was 46 seconds (!!!).

I just see that hyperreal has the mod_proxy cache enabled. The proxy cache
is the only place where delays of this magnitude can occur IMHO. I've
observed it taking MINUTES to complete on a large cache when the GC time
has arrived and/or the cache is full. Especially disappointing is the fact
that the request that caused the garbage collection (GC) is served but not
flushed, so in Netscape you see the partially served document and wait and
wait for the final bytes.

Ideas for improvement (without having looked at the code):
 *  Make GC take place in a separate process, not in one of the serving
    threads / preforked servers.
 *  If done in the serving process, then flush the document and finalize
    the request before GC'ing, and return to accepting state after GC.
 *  Don't keep a persistent connection when finding out that GC has to be
    done
 *  Avoid more than one of the serving processes to try to do GC

Is it possible that the cache caused your delays?

    Martin
-- 
| S I E M E N S |  <Martin.Kraemer@mch.sni.de>  |      Siemens Nixdorf
| ------------- |   Voice: +49-89-636-46021     |  Informationssysteme AG
| N I X D O R F |   FAX:   +49-89-636-44994     |   81730 Munich, Germany
~~~~~~~~~~~~~~~~My opinions only, of course; pgp key available on request