Python: Make algebra faster by using pool of vectors

[Joachim Haga]

In Python, the natural way to write matrix algebra involves a lot of temporary vector creation. By implementing a pool of vectors, I see a 15-20% performance improvement in a matrix-algebra heavy workload. The basic idea is to add a method to the GenericMatrix interface (Python only, probably). Then Matrix.__mul__ (and everywhere else a vector is needed) can use this method whenever they need to create a vector, and they are transparently re-used whenever they go out of scope.

    @vec_pool
    def create_vec(self, dim=1):
        vec = dolfin.Vector()
        self.resize(vec, dim)
        return vec

The implementation of the vec_pool decorator is found here: http://bazaar.launchpad.net/~jobh/cbc.block/trunk/download/head:/block/object__pool.py/object_pool.py
(code copyright me, LGPL or whatever you prefer.)

[Johan Hake (johan-hake)]

Looks like a nice thing to have but I am not sure it is appropriate to expand DOLFIN with this functionality. It would be more natural to extend libraries using this, such as cbc.block, for this functionality.

Also, using:
  D = A.copy()
  A += B

instead of:

  D = A+B

In this particular case it does not matter but in general it will.

Johan

[Garth Wells (garth-wells)]

This isn't a bug. It should be turned into a Blueprint (for discussion).

[Joachim Haga (jobh)]

Sure. It's a wishlist "bug". Feel free to reject if these are unwanted. I
have only very limited internet connection here, but can do the blueprint
thing later in the summer.

-j.
Den 5. apr. 2011 23.50 skrev "Garth Wells" <email address hidden>
følgende:
> This isn't a bug. It should be turned into a Blueprint (for discussion).
>
> --
> You received this bug notification because you are a direct subscriber
> of the bug.
> https://bugs.launchpad.net/bugs/747318
>
> Title:
> Python: Make algebra faster by using pool of vectors
>
> Status in DOLFIN:
> New
>
> Bug description:
> In Python, the natural way to write matrix algebra involves a lot of
> temporary vector creation. By implementing a pool of vectors, I see a
> 15-20% performance improvement in a matrix-algebra heavy workload. The
> basic idea is to add a method to the GenericMatrix interface (Python
> only, probably). Then Matrix.__mul__ (and everywhere else a vector is
> needed) can use this method whenever they need to create a vector, and
> they are transparently re-used whenever they go out of scope.
>
> @vec_pool
> def create_vec(self, dim=1):
> vec = dolfin.Vector()
> self.resize(vec, dim)
> return vec
>
> The implementation of the vec_pool decorator is found here:
http://bazaar.launchpad.net/~jobh/cbc.block/trunk/download/head:/block/object__pool.py/object_pool.py
> (code copyright me, LGPL or whatever you prefer.)
>
> To unsubscribe from this bug, go to:
> https://bugs.launchpad.net/dolfin/+bug/747318/+subscribe

[Joachim Haga (jobh)]

This code is already in cbc.block via code injection (aka monkey patching).
Just trying to move the good stuff upstream. I don't see how your examples
differ (except for cache behaviour) - either way a new vector must be
created, no?
Den 5. apr. 2011 23.41 skrev "Johan Hake" <email address hidden>
følgende:
> Looks like a nice thing to have but I am not sure it is appropriate to
> expand DOLFIN with this functionality. It would be more natural to
> extend libraries using this, such as cbc.block, for this functionality.
>
> Also, using:
> D = A.copy()
> A += B
>
> instead of:
>
> D = A+B
>
> In this particular case it does not matter but in general it will.
>
> Johan
>
> --
> You received this bug notification because you are a direct subscriber
> of the bug.
> https://bugs.launchpad.net/bugs/747318
>
> Title:
> Python: Make algebra faster by using pool of vectors
>
> Status in DOLFIN:
> New
>
> Bug description:
> In Python, the natural way to write matrix algebra involves a lot of
> temporary vector creation. By implementing a pool of vectors, I see a
> 15-20% performance improvement in a matrix-algebra heavy workload. The
> basic idea is to add a method to the GenericMatrix interface (Python
> only, probably). Then Matrix.__mul__ (and everywhere else a vector is
> needed) can use this method whenever they need to create a vector, and
> they are transparently re-used whenever they go out of scope.
>
> @vec_pool
> def create_vec(self, dim=1):
> vec = dolfin.Vector()
> self.resize(vec, dim)
> return vec
>
> The implementation of the vec_pool decorator is found here:
http://bazaar.launchpad.net/~jobh/cbc.block/trunk/download/head:/block/object__pool.py/object_pool.py
> (code copyright me, LGPL or whatever you prefer.)
>
> To unsubscribe from this bug, go to:
> https://bugs.launchpad.net/dolfin/+bug/747318/+subscribe

[Johan Hake (johan-hake)]

As I said, this example won't make a difference as it start of with a copy. But for a longer algorithm it might be worth while instantiating some tmp vectors and use imul. Imul does not generate additional vectors.

Johan

On Apr 7, 2011, at 12:07, Joachim Haga <email address hidden> wrote:

> This code is already in cbc.block via code injection (aka monkey patching)