APR and Apache Time Format Questions

[Jean-Jacques Clar <JJ...@novell.com>]

I have two Questions/Suggestions to improve the current performance 
of Apache 2.
 
1: 
Apache and Apr are keeping the time stamp in uSEC, it implies multiple

calls to apr_time_sec(), which is just doing a 64 bits division to
transform 
uSEC in sec. 
I think this is a waste of CPU cycles for nothing.
 
What about creating a new define called: 
 
#define APR_HAS_USEC.
 
The else of that option could allow apache to run using ONLY sec for
all 
apr_time_t field removing all 64 bits division to transform uSEC to
sec.
The size of the apr_time_t variables could also be changed to 32 bits
and 
calls to gettimeofday() changed to time() in that configuration.
 
Why is apache using uSEC? 
Time variables are mostly used in seconds at run-time, and http 
requirements use seconds.
 
The performance gain is minor but measurable: between 1 and 2.5% on 
my box (from 1 to 4 CPUs).
 
-----------

 
Questions: 
2-
If one does not fly, it should also be valuable to change the return
format 
of apr_date_parse_http() to be in seconds instead of uSEC.
This is a snippet of the last part of that function:
 
--- boc ---
    /* ap_mplode_time uses tm_usec and tm_gmtoff fields, but they
haven't 
     * been set yet. 
     * It should be safe to just zero out these values.
     * tm_usec is the number of microseconds into the second.  HTTP
only
     * cares about second granularity.
     * tm_gmtoff is the number of seconds off of GMT the time is.  By
     * definition all times going through this function are in GMT, so
this
     * is zero. 
     */
    ds.tm_usec = 0;
--- eoc ---
 
and then the time is exploded to match the apr_time_t format.
 
A call to apr_date_parse_httpd() is usually followed by a call to
apr_time_sec(), 
which do a 64 bits division to get the time in second.
What about having a new function called apr_date_parse_httpd_sec(),
which 
skip the end part of apr_time_exp_get() and return seconds?
It implies a small change to apr_time_exp_get() and moving the core
code in 
apr_time_exp_get() to apr_time_exp_get_sec().:
 
--- boc ---
APR_DECLARE(apr_status_t) apr_time_exp_get(apr_time_t *t,
apr_time_exp_t *xt)
{
    apr_status_t ret;
 
    ret = apr_time_exp_get_sec(t, xt);
    if (ret == APR_SUCCESS)
        *t = *t * APR_USEC_PER_SEC + xt->tm_usec;
 
    return ret;
}
--- eoc ---
 
Thank you,
 
Jean-Jacques 

Re: APR and Apache Time Format Questions

["William A. Rowe, Jr." <wr...@rowe-clan.net>]

You can't accomplish your patch and maintain binary compatibility
between different 'flavor builds' of APR.  Plus the number of bugs that
would occur would be a huge pain.

This has been debated many times on list, please query the archives
<http://marc.theaimsgroup.com/?l=apr-dev>

Different proposals have included using tm structures; going back to
seconds (not apr_time_t) for those functions that clearly only benefit
from 1sec resolution; and even binary microsecond resolution.  There
really hasn't been one definitive agreement, claims that "httpd only
benefits from 1sec resolution" are balanced against "most platforms
support sub-sec resolution for file times and other values, either 1ms,
1us, and sometimes even sub-1us values like NT's 100ns units".

It's a very uneasy consensus and one that is not very stable.  But any
arguments to change the definition of "apr_time_t" are likely to fall on
deaf ears - if you care to propose another type-name to represent 
whatever it is you are inventing - the list would be much more receptive.
Many bugs were cause by changing the time definition, more bugs 
will inevitably be introduced by changing it yet again.  A different type 
would help developers identify their mistakes, a type that doesn't
auto-resolve to an int would help even more :-)

Oh - this discussion belongs on apr@ ... discussing changing the
definition of time within httpd belongs on httpd@.

Bill

At 01:03 PM 3/3/2003, Jean-Jacques Clar wrote:
>I have two Questions/Suggestions to improve the current performance 
>of Apache 2.
> 
>1: 
>Apache and Apr are keeping the time stamp in uSEC, it implies multiple 
>calls to apr_time_sec(), which is just doing a 64 bits division to transform 
>uSEC in sec. 
>I think this is a waste of CPU cycles for nothing.
> 
>What about creating a new define called: 
> 
>#define APR_HAS_USEC.
> 
>The else of that option could allow apache to run using ONLY sec for all 
>apr_time_t field removing all 64 bits division to transform uSEC to sec.
>The size of the apr_time_t variables could also be changed to 32 bits and 
>calls to gettimeofday() changed to time() in that configuration.
> 
>Why is apache using uSEC? 
>Time variables are mostly used in seconds at run-time, and http 
>requirements use seconds.
> 
>The performance gain is minor but measurable: between 1 and 2.5% on 
>my box (from 1 to 4 CPUs).
> 
>-----------
> 
>Questions: 
>2-
>If one does not fly, it should also be valuable to change the return format 
>of apr_date_parse_http() to be in seconds instead of uSEC.
>This is a snippet of the last part of that function:
> 
>--- boc ---
>    /* ap_mplode_time uses tm_usec and tm_gmtoff fields, but they haven't 
>     * been set yet. 
>     * It should be safe to just zero out these values.
>     * tm_usec is the number of microseconds into the second.  HTTP only
>     * cares about second granularity.
>     * tm_gmtoff is the number of seconds off of GMT the time is.  By
>     * definition all times going through this function are in GMT, so this
>     * is zero. 
>     */
>    ds.tm_usec = 0;
>--- eoc ---
> 
>and then the time is exploded to match the apr_time_t format.
> 
>A call to apr_date_parse_httpd() is usually followed by a call to apr_time_sec(), 
>which do a 64 bits division to get the time in second.
>What about having a new function called apr_date_parse_httpd_sec(), which 
>skip the end part of apr_time_exp_get() and return seconds?
>It implies a small change to apr_time_exp_get() and moving the core code in 
>apr_time_exp_get() to apr_time_exp_get_sec().:
> 
>--- boc ---
>APR_DECLARE(apr_status_t) apr_time_exp_get(apr_time_t *t, apr_time_exp_t *xt)
>{
>    apr_status_t ret;
> 
>    ret = apr_time_exp_get_sec(t, xt);
>    if (ret == APR_SUCCESS)
>        *t = *t * APR_USEC_PER_SEC + xt->tm_usec;
> 
>    return ret;
>}
>--- eoc ---
> 
>Thank you,
> 
>Jean-Jacques