Performance optimization in av::fixup - use buffered IO, not mapped file

[Jin-woo Ye]

On 2012-12-28 AM 7:53, Ryan Johnson wrote:
> On 14/12/2012 4:18 AM, Corinna Vinschen wrote:
>> On Dec 13 19:53, Corinna Vinschen wrote:
>>> On Dec 13 10:32, Eric Blake wrote:
>>>> On 12/13/2012 07:30 AM, Corinna Vinschen wrote:
>>>>> Below is my cut on the issue.  It introduces a "sparse" mount
>>>>> option and
>>>>> sparsifies a file only if this mount option is set.  I also slightly
>>>>> improved the code in fhandler_base::write and ftruncate so that it
>>>>> will
>>>>> not try to set the sparse flag on already sparse files.
>>>>>
>>>>> The code compiles, so it's basically correct.  It's just not
>>>>> tested. ;)
>>>> Also untested (so far) on my end, but I can agree to this - anyone that
>>>> _wants_ sparse files (such as for virtual disk image) will have to
>>>> enable that mount point option for the directory of those files in
>>>> question, but no one else is forced to use them.  As sparse files are
>>>> _supposed_ to be an optimization, not being sparse is not a loss in
>>>> functionality, just in potential optimizations; but if Windows is not
>>>> taking advantage of optimizations even when a file IS sparse, then that
>>>> argues that we aren't losing much.
>>>>
>>>> It should not be too hard for me to rig the coreutils testsuite to
>>>> ensure that its tests of sparse operations are done on a mount
>>>> explicitly set up to allow sparse files.
>>> Yup, just add a temporary user mount point.
>> I applied my patch with a single fix.  The logic for adding the "sparse"\
>> option string to mnt_opts was upside down.
> I just finished building gcc-4.7 using the latest cygwin1 snapshot, and
> performance is vastly, drastically improved. Stages 2 and 3 went as fast
> as Stage 1, and the disk light only flickered occasionally instead of
> staying pegged like it used to.
>
> Thanks for finding and fixing this!
>
> Ryan
>
>
Building lib{java|fortran|stdc++} are now much faster than before the fix.
I don't want to remind how slow building gcc was !
Before that fix, I had to defragment a hard disk during building gcc or 
it would take more than 3 hours.
Thanks for Corinna for fixing this.

-- 
Regards.