Mailing-List: contact dev-help@httpd.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@httpd.apache.org
Message-ID: <3E14B4D3.9030802@cnet.com>
Date: Thu, 02 Jan 2003 13:53:23 -0800
From: Brian Pane <brian.pane@cnet.com>
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.0; en-US;
 rv:1.1b) Gecko/20020722
MIME-Version: 1.0
To: dev@httpd.apache.org
Subject: Re: [PATCH] remove some mutex locks in the worker MPM
References: <4E88DDB5-1DD6-11D7-B5B4-000393B3C494@clove.org>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Aaron Bannert wrote:

> The patch looks good at first glance. Have you done any testing
> to see how much it improves performance (on UP and MP machines)
> and if it has any effect when APR is build with generic atomics?


Here are the performance numbers that I have.  I ran
httpd-2.1.0-dev on an 8x167MHZ CPU Sun with Solaris 8
(32-bit mode).  The client driver sent a fixed number
of concurrent requests for a 1-byte file (to keep the
time spent in network writes from overshadowing the
results).

I tested with both the SPARC V8+ native atomic ops
and APR's mutex-based default atomics:

50 clients
                              load  %CPU  req/s
standard 2.1.0-dev worker     4.59  0.53  1090
patched w/native atomic ops   4.57  0.53  1097
        w/mutex atomic ops    4.70  0.52  1093

100 clients
                              load  %CPU  req/s
standard 2.1.0-dev worker     4.74  0.54  1070
patched w/native atomic ops   4.69  0.54  1083
        w/mutex atomic ops    4.61  0.54  1067

Basically, the patch results in a slightly higher
throughput with lower CPU load.  That matches what
I'd expect from a reduction of mutex contention.
With the mutex-based fallback implementation of the
apr_atomic API, performance was slightly worse than
the original worker code in the 100-client case, but
faster in the 50-client case.  (The effect of using
my worker patch with the mutex-based atomics is to
increase the number of lock calls while reducing
the amount of time spent in each critical region.
These two effects seem to counteract each other.)

Brian