From: Harald Anlauf <anlauf@gmx.de>
To: fortran@gcc.gnu.org
Subject: Re: Problem setting buffer size for gfortran ( v 11.2)
Date: Wed, 23 Feb 2022 21:57:11 +0100 [thread overview]
Message-ID: <9cd697e4-eb65-bf4f-be28-8c67b7acdcd6@gmx.de> (raw)
Message-ID: <20220223205711.FNpAhkrYSBS1DCMV8FsPBp294HsYErHHMp8OggxTqpY@z> (raw)
In-Reply-To: <cc6f4e7f1f184bdcb430025c71e92a02@gsi.de>
Hi Denis,
Am 23.02.22 um 21:28 schrieb Bertini, Denis Dr. via Fortran:
> Hi Harald,
>
> In attachment you will find a typical strace output when running with
>
> gfortran8 on lustre filsystem.
unfortunately I do not see any attachment to your mail.
Note: GFORTRAN_FORMATTED_BUFFER_SIZE & friends was added in 10.1,
and possibly backported to some 9.y, but *not* to 8.x.
> As i said for gfortran version 8.x, 10.x, 11.2 setting the runtime
>
> env. variable to change buffer size did not work for me as you could
>
> see in the strace output.
>
> You could also see that "mmap" is used by the lustre client.
>
> The optimal blocksize given back by the STAT() routine indeed
>
> favor a 4 MiB buffer size.
It might make sense to show your issue to the vendor of your system.
Parallel file systems - not only Lustre - require special handling
not only to get optimum performance, but also correct behavior.
(E.g. the HDF5 library is known to hit bugs in our Lustre.)
Cheers,
Harald
>
> Best
>
> Denis
>
>
> ________________________________
> From: Harald Anlauf <anlauf@gmx.de>
> Sent: Wednesday, February 23, 2022 7:26 PM
> To: Bertini, Denis Dr.; fortran@gcc.gnu.org
> Subject: Re: Problem setting buffer size for gfortran ( v 11.2)
>
> Hi Denis,
>
> Am 23.02.22 um 17:10 schrieb Bertini, Denis Dr. via Fortran:
>> Hi
>> Investigating some sporadical performance drop when using Fortran IO on our Lustre
>> shared file system, we realized that the buffer size used as default by the gnu fortran
>> compiler is 8 kiB only.
>> We wanted to change this default and according to the documentation, one has just to set
>> the environment variables:
>>
>> GFORTRAN_UNFORMATTED_BUFFER_SIZE
>> GFORTRAN_FORMATTED_BUFFER_SIZE=
>> GFORTRAN_UNBUFFERED_ALL=n
>>
>> to increase the buffer size during runtime.
>
> this works for me on gcc-12 development for the following
> test program on Linux/x86:
>
> open(10,file="bigfile")
> write(10,*) "data"
> close(10)
> end
>
> Compile, link, and run under strace with
>
> GFORTRAN_FORMATTED_BUFFER_SIZE=4194304 strace ./a.out
>
> [...]
> openat(AT_FDCWD, "bigfile", O_RDWR|O_CREAT|O_CLOEXEC, 0666) = 3
> fstat(3, {st_mode=S_IFREG|0644, st_size=6, ...}) = 0
> mmap(NULL, 4198400, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1,
> 0) = 0x146b69b1c000
> write(3, " data\n", 6) = 6
> ftruncate(3, 6) = 0
> munmap(0x146b69b1c000, 4198400) = 0
> close(3) = 0
>
>> Unfortunately isetting these variables does not change the default buffer size.
>> I tried with gfortran 8, 10, and the new 11.2.
>> Looking a the code, i found in
>> io/unix.c
>> that the default is set to the following value,
>>
>> static const int FORMATTED_BUFFER_SIZE_DEFAULT = 8192;
>> static const int UNFORMATTED_BUFFER_SIZE_DEFAULT = 128*1024;
>>
>> Of course, changing the default value and recompiling works and change accordingly the buffer size
>> It should nevertheless be possible to change these values without recompiling.
>> Is there something one should be aware of when setting these variables?
>
> You said you're using Lustre. I guess you use the default
> blocksize (4MB). Does your Lustre client support mmap?
> What does strace report?
>
> I remember I once saw a system where we had to disable the use
> of mmap in a library.
>
> Cheers,
> Harald
>
>> Best regards,
>> Denis
>>
>> ---------
>> Denis Bertini
>> Abteilung: CIT
>> Ort: SB3 2.265a
>>
>> Tel: +49 6159 71 2240
>> Fax: +49 6159 71 2986
>> E-Mail: d.bertini@gsi.de
>>
>> GSI Helmholtzzentrum für Schwerionenforschung GmbH
>> Planckstraße 1, 64291 Darmstadt, Germany, www.gsi.de
>>
>> Commercial Register / Handelsregister: Amtsgericht Darmstadt, HRB 1528
>> Managing Directors / Geschäftsführung:
>> Professor Dr. Paolo Giubellino, Dr. Ulrich Breuer, Jörg Blaurock
>> Chairman of the GSI Supervisory Board / Vorsitzender des GSI-Aufsichtsrats:
>> Ministerialdirigent Dr. Volkmar Dietz
>>
>
>
next prev parent reply other threads:[~2022-02-23 20:57 UTC|newest]
Thread overview: 20+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-02-23 16:10 Bertini, Denis Dr.
2022-02-23 18:26 ` Harald Anlauf
2022-02-23 18:26 ` Harald Anlauf
2022-02-23 20:28 ` Bertini, Denis Dr.
2022-02-23 20:57 ` Harald Anlauf [this message]
2022-02-23 20:57 ` Harald Anlauf
2022-02-23 20:59 ` Bertini, Denis Dr.
[not found] ` <trinity-9bf94afe-e773-4366-b0b2-ae398de80664-1645651097363@3c-app-gmx-bs49>
2022-02-23 21:23 ` Bertini, Denis Dr.
2022-02-23 21:38 ` Harald Anlauf
2022-02-24 7:57 ` Bertini, Denis Dr.
2022-02-24 8:13 ` Tobias Burnus
2022-02-24 8:15 ` Bertini, Denis Dr.
2022-02-24 8:20 ` Tobias Burnus
2022-02-25 17:28 ` Bertini, Denis Dr.
[not found] ` <trinity-59001f50-077d-424e-a648-24645944dd0c-1645813711633@3c-app-gmx-bap28>
2022-02-25 20:01 ` Bertini, Denis Dr.
2022-02-25 20:44 ` Harald Anlauf
2022-02-25 20:44 ` Harald Anlauf
2022-02-25 20:53 ` Bertini, Denis Dr.
2022-02-26 9:20 ` Thomas Koenig
2022-02-26 15:37 ` Bertini, Denis Dr.
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=9cd697e4-eb65-bf4f-be28-8c67b7acdcd6@gmx.de \
--to=anlauf@gmx.de \
--cc=fortran@gcc.gnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).