On Fri, 16 Oct 2015 12:24:39 +0100
Adrian Jackson <[email protected]> wrote:

> Hi,
> 
> It looks like INSTANT_SYSTEM_CALL_METHOD=OS_SYSTEM fixes the problem
> we were seeing.  Does this have a performance impact for jobs using
> only one node?

Great! OS_SYSTEM uses more tempfiles rather than pipes so potentially
being more IO demanding. But with the current implementation of JIT
chain this is probably not critical. Moreover there are some annoying
side effects for DOLFIN test suite,
https://bitbucket.org/fenics-project/dolfin/issues/316

Jan

> 
> thanks
> 
> adrianj
> 
> On 16/10/2015 10:46, Jan Blechta wrote:
> > On Thu, 15 Oct 2015 15:36:44 +0100
> > Adrian Jackson <[email protected]> wrote:
> > 
> >> '/work/z01/z01/adrianj/.instant/ffc_form_e1514c0a3fffa4b6c0eda574739266f451a578aa/compile.log'
> > 
> > It could be useful to see this file.
> > 
> > BTW, isn't it InfiniBand (or similar OFED-based) cluster? InfiniBand
> > software stack is prone to segfault on fork calls involved in
> > JITting which can be circumvented by experimenting with
> > INSTANT_SYSTEM_CALL_METHOD environment variable, see Instant README
> > https://bitbucket.org/fenics-project/instant/
> > 
> > Jan
> > 
> 

_______________________________________________
fenics-support mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics-support

Reply via email to