Detlef Michael wrote:
>
> Hallo,
> does anybody know, if Linux-SMP uses lock-free message buffers??
> The background of my question:
> I've got the ch_shmem device of MPICH to work on my system, [...]
Hey, how did you manage THAT?
I think you had to patch the mpich sources? At least with version 1.1.1
the stock sources weren't able to use ch_shmem?
> [...] and I'd like to
> know if it makes sense to try the ch_lfshmem device. The MPICH documentation
> says only, on LINUX systems shared memory devices sometimes will work and
> sometimes not.
As far as I understand the whole thing, mpich tries to use /dev/zero for
the shmem-communication, but the linux kernel won't allow this (at least
it works for SysV, AFAIK. In fact I do have a Solaris ix86 on my SMP
machine and really want to compare the results, but I'm not that
experienced with Solaris and really think, that there's no compiler
installed...).
I don't know, if there are plans to work around this or allow user
processes to share memory via special kernel functions, but I don't
know, how... and cause there's a code freeze on the kernel (right?), I
don't think such a feature will get into 2.2.x
A workaround may be a ram drive mounted with sticky bit (e.g. /ram with
chmod 1777) and patch the mpich sources to use a file /ram/shmem instead
of /dev/zero).
BUT: This is UNTESTED... Use it at your own risk. You're welcome to
report, if it worked for you...
BTW: Another good starting point for MPICH discussions is
comp.parallel.mpi and, there should be an mpich mailing list!
cul8r,
Frank
