VMD-L Mailing List

From: Bogdan Costescu (Bogdan.Costescu_at_IWR.Uni-Heidelberg.De)
Date: Mon Jan 10 2005 - 09:56:14 CST

On Thu, 6 Jan 2005, John Stone wrote:

> If not, then there'd be no reason to add that to the new startup
> script, but if so, then I agree that may be a smart move, to prevent
> this problem from affecting others.

The threading problem comes from differences between the
compile/link-time environment and the runtime environment. It's mostly
a difference between your choice of Linux distribution for
compiling/linking and the choice of Linux distribution by the user at
runtime. Different distributions chose different ways and different
times of dealing with the new threading model, that's why these
problems appear; however, so far, the incidence of this problem with
VMD is low enough that a general solution might not be worth doing.
Providing a VMD binary distribution means making a choice about the
compiling/linking behaviour; you are obviously not able to control the
runtime environment, so I think that there are several ways of dealing
with it:
- just document what change is needed to the startup script for those
few that encounter the problem
- make the change to the startup script and hope that this covers all
cases
- create 2 different binary distributions that are compiled/linked in
the 2 different threading enviroments

> Honestly though, this is really a problem with the Linux threads
> runtime and not VMD itself. I'd rather that the Linux distros fix
> their threads library to avoid this kind of problem in the first
> place.

The threading "problem" comes from switching from the older pthreads
to the new NPTL. The switching involves both kernel and glibc changes
and therefore some distributions avoided them until recently when they
started being enabled by default (unlike Red Hat Linux 9, which is
quite old, but has NPTL support). NPTL is here to stay, so it's quite
safe to say that future builds of VMD (hopefully on some up-to-date
distribution) will actually be built in a NPTL environment and no
longer in a pthreads one, therefore setting LD_ASSUME_KERNEL will not
be needed anymore (and might actually be harmful).

For reference, the LD_ASSUME_KERNEL values are explained at:

http://people.redhat.com/drepper/assumekernel.html

> No other OS has this type of problem, AFAIK.

When was the last time an OS changed their threading implementation ? ;-) 

-- 
Bogdan Costescu
IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu_at_IWR.Uni-Heidelberg.De