dolfin team mailing list archive
-
dolfin team
-
Mailing list archive
-
Message #15036
Re: [HG DOLFIN] Add a 'mpi_jit_decorator' function which generalizes the paralell jit
On Sat, Aug 29, 2009 at 03:14:27PM +0200, Anders Logg wrote:
> On Sat, Aug 29, 2009 at 12:31:24AM +0200, Johan Hake wrote:
> > On Friday 28 August 2009 22:35:08 Anders Logg wrote:
> > > On Wed, Aug 19, 2009 at 01:41:50PM +0200, DOLFIN wrote:
> > > > One or more new changesets pushed to the primary dolfin repository.
> > > > A short summary of the last three changesets is included below.
> > > >
> > > > changeset: 6801:d3897056b4c92df1a2decddb8481f01011468683
> > > > tag: tip
> > > > user: "Johan Hake <hake@xxxxxxxxx>"
> > > > date: Wed Aug 19 13:41:45 2009 +0200
> > > > files: site-packages/dolfin/compile_extension_module.py
> > > > site-packages/dolfin/jit.py description:
> > > > Add a "mpi_jit_decorator" function which generalizes the paralell jit
> > > > compilation algorithm.
> > > >
> > > > It utilize a nice feature in Python: decorators
> > > > To apply it to a jit_compiler just import the decorator and decorate
> > > > your jit function:
> > > >
> > > > @mpi_jit_decorator
> > > > def some_jit_compiler(some):
> > > > ....
> > >
> > > Pretty cool, but why the threads? I don't understand why that should
> > > be needed.
> >
> > If we want a message that tells the user that a form will be jit compiled on
> > the first process, only when it actually is compiling, I couldn't find another
> > non-intrusive way to implement it.
> >
> > The same process that starts the compiling needs to find out if an actual
> > compilation has started. This is done using a thread. It starts it, wait a
> > while 0.2 s I think, using the join function), check if the thread is alive.
> > If the thread is alive, assume it is compiling and issue a message. If the
> > thread is not alive we assume the module is picked up from cache, and we
> > proceed without issuing a message. Also note that if the cache is used, the
> > present thread will finish the join statement before the 0.2 s.
>
> Maybe it's enough that we get a message from the actual JIT function
> that it is compiling? That message will only appear (at least from
> FFC) when code is not cached.
I removed this so that the threads are not needed. We can set a prefix
"Process 0:" for the FFC (UFL) log system which will accomplish the
same thing.
--
Anders
Attachment:
signature.asc
Description: Digital signature
Follow ups
References