This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [Patch, fortran] PR 56981 Improve unbuffered unformatted performance
- From: Jerry DeLisle <jvdelisle at charter dot net>
- To: Janne Blomqvist <blomqvist dot janne at gmail dot com>
- Cc: Fortran List <fortran at gcc dot gnu dot org>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Sun, 28 Apr 2013 15:46:02 -0700
- Subject: Re: [Patch, fortran] PR 56981 Improve unbuffered unformatted performance
- References: <CAO9iq9HLQzJud_ew_y17ZPvsaNJWy-uwX-q8XGKcQ2CEcA_ROQ at mail dot gmail dot com> <CAO9iq9HwA3cbtzWtjKOaRpt_sYBbXU7JdfmDdBSeZmbp7xpMRA at mail dot gmail dot com>
On 04/28/2013 11:31 AM, Janne Blomqvist wrote:
> PING
>
> On Fri, Apr 19, 2013 at 1:30 PM, Janne Blomqvist
> <blomqvist.janne@gmail.com> wrote:
>> Hi,
>>
>> the attached patch improves the performance for unformatted and
>> unbuffered files. Currently unbuffered unformatted really means that
>> we don't buffer anything and use the POSIX I/O syscalls directly. With
>> the patch, we use the buffer but flush it at the end of each I/O
>> statement.
>>
>> (For formatted I/O we essentially already do this, as the format
>> buffer (fbuf) buffers each record).
>>
>> For the ever important benchmark of writing small (containing a single
>> scalar 4 byte value) unformatted sequential records to /dev/null, the
>> patch reduces the number of syscalls by a factor of 6, and performance
>> improves by more than double.
>>
>> For trunk, time for the benchmark in the PR:
>>
>> real 0m0.727s
>> user 0m0.272s
>> sys 0m0.452s
>>
>> With the patch:
>>
>> real 0m0.313s
>> user 0m0.220s
>> sys 0m0.092s
>>
>> For comparison, writing to a file where we use buffered I/O:
>>
>> real 0m0.202s
>> user 0m0.180s
>> sys 0m0.020s
>>
>>
>> As a semi-unrelated minor improvement, the patch also changes the
>> ordering when writing out unformatted sequential record markers.
>> Currently we do
>>
>> write bogus marker
>> write record data
>> write tail marker
>> seek back to before the bogus marker
>> write the correct head marker
>> seek to the end of the record, behind the tail marker
>>
>> With the patch we instead do
>>
>> write bogus marker
>> write record data
>> seek back to before the bogus marker
>> write the correct head marker
>> seek to the end of the record data
>> write tail marker
>>
>> With the patch, the slightly shorter seek distances ever-so-slightly
>> increase the chance that the seeks will be contained within the buffer
>> so we don't have to flush.
>>
>> Regtested on x86_64-unknown-linux-gnu, Ok for trunk?
>>
OK Janne and thanks for the patch.
What are your thoughts about special casing nul devices/
Jerry