[cfarm-users] request for tmpfs on gcc220 (was: gcc220 replaced with new hardware)

[Ævar Arnfjörð Bjarmason]

On Fri, Nov 05 2021, Baptiste Jonglez via cfarm-users wrote:

> [[PGP Signed Part:Undecided]]
> Hi,
>
> On 05-11-21, Juan Francisco Cantero Hurtado via cfarm-users wrote:
>> On Fri, 05 Nov 2021 11:14:47 +0100
>> Ævar Arnfjörð Bjarmason via cfarm-users
>> <cfarm-users at lists.tetaneutral.net> wrote:
>> > That's great, yes thanks for providing an OpenBSD machine that's this
>> > capible, it's not long ago that I'd often not test at all on OpenBSD
>> > because there was just some very slow VM'd 1-core box.
>> > 
>> > Request: Would it be possible to provide a tmpfs on that machine
>> > ("man 8 mount_tmpfs" I believe) similar to the /run/user on some
>> > linux boxes? Something in the small number of GBs would be more than
>> > sufficient.
>> 
>> tmpfs was disabled long time ago on OpenBSD and is unsupported. The
>> closest thing to tmpfs is mfs:
>
> Indeed, the mount_tmpfs man page and command are still there, but it
> doesn't work:
>
>   # mount_tmpfs tmpfs /mnt/test
>   mount_tmpfs: tmpfs on /mnt/test: Operation not supported

Sorry, I missed that.

>> https://dataswamp.org/~solene/2018-05-08-mfs-tmp.html
>> 
>> However, the performance is not really great. Also, if you reserve
>> space for mfs, it can't be used as normal RAM.
>
> This system has so much memory that it's not a problem to dedicate a
> portion of memory to this.
>
> After trying different configuration, it seems that MFS is limited to
> around 32 GB of filesytem size:
>
>   # echo "swap /scratch mfs rw,nodev,nosuid,-s=32750m,-m=0 0 0" >> /etc/fstab
>   # mount /scratch
>   # df -h /stratch
>   Filesystem     Size    Used   Avail Capacity  Mounted on
>   mfs:82330     31.0G    1.0K   31.0G     0%    /scratch
>
> Specifying a larger size fails with:
>
>   # mount -t mfs -o -s33G,-m0 swap /mnt/test   
>   mount_mfs: mmap: Cannot allocate memory
>
> 32 GB should be enough already: Ævar, can you give it a try in /scratch/
> and see how it goes?  If it works well, we could enable it for /tmp.

Thanks, that seems to be quite a bit faster, I tested it with:

    hyperfine --show-output -L d /scratch/git,/tmp/git 'ssh gcc220.fsffrance.org "cd git/t && ./t3000-ls-files-others.sh --root={d}"'

And after a while got:

Summary
  'ssh gcc220.fsffrance.org "cd git/t && ./t3000-ls-files-others.sh --root=/scratch/git"' ran
    1.28 ± 0.18 times faster than 'ssh gcc220.fsffrance.org "cd git/t && ./t3000-ls-files-others.sh --root=/tmp/git"'