[Cython] cython.parallel tasks, single, master, critical, barriers

[Robert Bradshaw]

On Sun, Oct 9, 2011 at 5:57 AM, Dag Sverre Seljebotn
<d.s.seljebotn at astro.uio.no> wrote:
> On 10/09/2011 02:18 PM, Dag Sverre Seljebotn wrote:
>>
>> On 10/09/2011 02:11 PM, mark florisson wrote:
>>>
>>> Hey,
>>>
>>> So far people have been enthusiastic about the cython.parallel features,
>>> I think we should introduce some new features.

Excellent. I think this is going to become a killer feature like
buffer support.

>>> I propose the following,
>>
>> Great!!
>>
>> I only have time for a very short feedback now, perhaps more will follow.
>>
>>> assume parallel has been imported from cython:
>>>
>>> with parallel.master():
>>> this is executed in the master thread in a parallel (non-prange)
>>> section
>>>
>>> with parallel.single():
>>> same as master, except any thread may do the execution
>>>
>>> An optional keyword argument 'nowait' specifies whether there will be a
>>> barrier at the end. The default is to wait.
>
> I like
>
> if parallel.is_master():
>    ...
> explicit_barrier_somehow() # see below
>
> better as a Pythonization. One could easily support is_master to be used in
> other contexts as well, simply by assigning a status flag in the master
> block.

+1, the if statement feels a lot more natural.

> Using an if-test flows much better with Python I feel, but that naturally
> lead to making the barrier explicit. But I like the barrier always being
> explicit, rather than having it as a predicate on all the different
> constructs like in OpenMP....
>
> I'm less sure about single, since making it a function indicates one could
> use it in other contexts and the whole thing becomes too magic (since it's
> tied to the position of invocation). I'm tempted to suggest
>
> for _ in prange(1):
>    ...
>
> as our syntax for single.

The idea here is that you want a block of code executed once,
presumably by the first thread that gets here? I think this could also
be handled by a if statement, perhaps "if parallel.first()" or
something like that. Is there anything special about this construct
that couldn't simply be done by flushing/checking a variable?

>>> with parallel.task():
>>> create a task to be executed by some thread in the team
>>> once a thread takes up the task it shall only be executed by that
>>> thread and no other thread (so the task will be tied to the thread)
>>>
>>> C variables will be firstprivate
>>> Python objects will be shared
>>>
>>> parallel.taskwait() # wait on any direct descendent tasks to finish
>>
>> Regarding tasks, I think this is mapping OpenMP too close to Python.
>> Closures are excellent for the notion of a task, so I think something
>> based on the futures API would work better. I realize that makes the
>> mapping to OpenMP and implementation a bit more difficult, but I think
>> it is worth it in the long run.

It's almost as if you're reading my thoughts. There are much more
natural task APIs, e.g. futures or the way the Python
threading/multiprocessing does things.

>>> with parallel.critical():
>>> this section of code is mutually exclusive with other critical sections
>>> optional keyword argument 'name' specifies a name for the critical
>>> section,
>>> which means all sections with that name will exclude each other,
>>> but not
>>> critical sections with different names
>>>
>>> Note: all threads that encounter the section will execute it, just
>>> not at the same time
>
> Yes, this works well as a with-statement...
>
> ..except that it is slightly magic in that it binds to call position (unlike
> anything in Python). I.e. this would be more "correct", or at least
> Pythonic:
>
> with parallel.critical(__file__, __line__):
>    ...

This feels a lot like a lock, which of course fits well with the with
statement.

>>> with parallel.barrier():
>>> all threads wait until everyone has reached the barrier
>>> either no one or everyone should encounter the barrier
>>> shared variables are flushed
>
> I have problems with requiring a noop with block...
>
> I'd much rather write
>
> parallel.barrier()
>
> However, that ties a function call to the place of invocation, and suggests
> that one could do
>
> if rand() > .5:
>    barrier()
> else:
>    i += 3
>    barrier()
>
> and have the same barrier in each case. Again,
>
> barrier(__file__, __line__)
>
> gets us purity at the cost of practicality. Another way is the pthreads
> approach (although one may have to use pthread rather then OpenMP to get it,
> unless there are named barriers?):
>
> barrier_a = parallel.barrier()
> barrier_b = parallel.barrier()
> with parallel:
>    barrier_a.wait()
>    if rand() > .5:
>        barrier_b.wait()
>    else:
>        i += 3
>        barrier_b.wait()
>
>
> I'm really not sure here.

I agree, the barrier doesn't seem like it belongs in a context. For
example, it's ambiguous whether the block is supposed to proceed or
succeed the barrier. I like the named barrier idea, but if that's not
feasible we could perhaps use control flow to disallow conditionally
calling barriers (or that every path calls the barrier (an equal
number of times?)).

>>> Unfortunately, gcc again manages to horribly break master and single
>>> constructs in loops (versions 4.2 throughout 4.6), so I suppose I'll
>>> first file a bug report. Other (better) compilers like Portland (and I'm
>>> sure Intel) work fine. I suppose a warning in the documentation will
>>> suffice there.

One can emit conditional #error pragmas in this case, though of course
it's better to produce code that works correctly on all compilers.

>>> If we at some point implement vector/SIMD operations we could also try
>>> out the Fortran openmp workshare construct.
>>
>> I'm starting to learn myself OpenCL as part of a course. It's very neat
>> for some kinds of parallelism. What I'm saying is that at least of the
>> case of SIMD, we should not lock ourselves to Fortran+OpenMP thinking
>> too early, but also look forward to coming architectures (e.g., AMD's
>> GPU-and-CPU on same die design).

+1. I like the idea of providing more parallelism constructs, but
rather than risk fixating on OpenMP's model, perhaps we should look at
the problem we're trying to solve (e.g., what can't one do well now)
and create (or more likely borrow) the right Pythonic API to do it.

- Robert