/tmp, /var/log, /var/tmp as /dev/md - why?

[Ian Lepore]

On Thu, 2014-07-03 at 12:55 +0200, John Hay wrote:
> On Thu, Jul 03, 2014 at 04:47:24AM -0600, Warren Block wrote:
> > On Wed, 2 Jul 2014, Mattia Rossi wrote:
> > 
> > >
> > > Am 01.07.2014 21:27, schrieb Andreas Schwarz:
> > >> Speed and speed, but I can't understand why using md here, there is already 
> > >> tmpfs,
> > >> which optimzed for such cases (dynamic allocation, etc.).
> > >> 
> > >> root at pizelot:~ # df
> > >> Filesystem     1K-blocks    Used   Avail Capacity  Mounted on
> > >> /dev/mmcsd0s2a    983680   57252  847736     6%    /
> > >> devfs                  1       1       0   100%    /dev
> > >> /dev/mmcsd0s2d   8106716 3068708 4389472    41%    /usr
> > >> /dev/mmcsd0s2e   8106716  155976 7302204     2%    /var
> > >> /dev/mmcsd0s2f   8106716236 7457944  <sip:2367457944>      0%    /home
> > >> tmpfs            1097160       4 1097156     0%    /tmp
> > >> tmpfs            1097160       4 1097156     0%    /var/tmp
> > >> 
> > >
> > > On an embedded systems with little memory I prefer to limit the partitions to 
> > > a certain size, like 32M, so dynamic allocation is no advantage. What other 
> > > differences are there between tmpfs and a simple md device?
> > > I'd be interested in knowing any tricks, that can make the system faster :-)
> > 
> > The white paper on tmpfs (wiki.deimos.fr/images/1/1e/Solaris_tmpfs.pdf) 
> > says:
> > 
> >    "RAM disks use memory inefficiently; file data exists twice in both
> >     RAM disk memory and kernel memory, and RAM disk memory that is not
> >     being used by the file system is wasted.  RAM disk memory is
> >     maintained separately from kernel memory, so that multiple
> >     memory-to-memory copies are needed to update file system data."
> > 
> > So a limited-size tmpfs will be faster and use less memory overall.  A
> > benchmark comparison would be interesting.
> 
> Last time I looked the rc scripts that create /etc, /var and /tmp
> ramdisks only did it using md devices. It would be great if it was
> easily tunable from say rc.conf or if could detect which one is
> available and use that.
> 
> John

I have patches ready to commit that do exactly that, but they weren't
exactly enthusiastically received when I posted them on arch@ for
review.

-- Ian