[Cython] Acquisition counted cdef classes

[mark florisson]

On 26 October 2011 11:23, Dag Sverre Seljebotn
<d.s.seljebotn at astro.uio.no> wrote:
> On 10/26/2011 11:45 AM, mark florisson wrote:
>>
>> On 26 October 2011 08:56, Stefan Behnel<stefan_ml at behnel.de>  wrote:
>>>
>>> Greg Ewing, 26.10.2011 00:27:
>>>>
>>>> Dag Sverre Seljebotn wrote:
>>>>
>>>>> I'd gladly take a factor two (or even four) slowdown of CPython code
>>>>> any
>>>>> day to get rid of the GIL :-). The thing is, sometimes one has 48 cores
>>>>> and consider a 10x speedup better than nothing...
>>>>
>>>> Another thing to consider is that locking around refcount
>>>> changes may not be as expensive in typical Cython code as
>>>> it is in Python.
>>>>
>>>> The trouble with Python is that you can't so much as scratch
>>>> your nose without touching a big pile of ref counts. But
>>>> if the Cython code is only dealing with a few Python objects
>>>> and doing most of its work at the C level, the relative
>>>> overhead of locking around refcount changes may not be
>>>> significant.
>>>>
>>>> So it may be worth trying the strategy of just acquiring
>>>> the GIL whenever a refcount needs to be changed in a nogil
>>>> section, and damn the consequences.
>>>
>>> Hmm, interesting. That would give new semantics to "nogil" sections,
>>> basically:
>>>
>>> """
>>> You can do Python interaction in nogil code, however, this will slow down
>>> your code. Cython will generate C code to acquire and release the GIL
>>> around
>>> any Python interaction that your code performs, thus serialising any
>>> calls
>>> into the CPython runtime. If you want to avoid this serialisation, use
>>> "cython -a" to find out where Python interaction happens and use static
>>> typing to let Cython generate C code instead.
>>> """
>>>
>>> In other words: "with gil" sections hold the GIL by default and give it
>>> away
>>> on explicit request, whereas "nogil" sections have the GIL released by
>>> default and acquire it on implicit need.
>>>
>>> The advantage over object level locking is that this does not increase
>>> the
>>> in-memory size of the object structs, and that it works with *any* Python
>>> object, not just extension types with a compile time known type.
>>>
>>> I kind of like that.
>>
>> My problem with that is that if there if any other python thread,
>> you're likely just going to sleep for thousands of CPU cycles as that
>> thread will keep the GIL. Doing this implicitly for operations with
>> such overhead would be unacceptable. I think writing 'with gil:' is
>> fine, it's the performance that's the problem in the first place which
>> prevents you from doing that, not the 9 characters you need to type.
>
> You are sure about the complete impossibility of having a seperate thread
> doing all INCREFs and DECREFs posted to it asynchronously (in the order they
> are posted), without race conditions?

No I think it is possible, but I don't believe it will solve the
DECREF C compiler optimization prevention problem (unlikely() should
help there though) as it will still have to submit an asynchronous
DECREF without races which means it has to call some kind of
(synchronized or atomically operating) function (which prevented the
optimization). It would be nice to have as it would mean you can pass
stuff around in nogil mode without acquisition counting, and it would
mean you can implement these types that can be used in nogil mode and
can synchronize using their own lock (if needed).

I wonder if deferring INCREFs are safe though. What if you have one
reference, you INCREF (deferred, because you don't have the GIL), you
call some function that steals your reference (after you obtained the
GIL), you somehow cause the program to lose the stolen reference which
causes it to be collected, and then the reference counter thread
decides to do the INCREF (too late). You also cannot atomically
INCREF, and Python doesn't do that, so there could be a race there as
well. So I think you really need the GIL to INCREF, and you need to do
it synchronously (I'm not completely sure, please feel free to poke
holes in my logic any time :).

I think it would be nicer to just fix this in CPython in any case,
though. Reference counting is terrible to work with in general
(regardless of whether you do them immediately or defer them), and
it's part of the reason why we have a GIL (although really not the
only one). As long as CPython does reference counting, removing the
GIL is an absolute no-go (although I wonder how many architectures
don't support atomic reference counting).

Refcounting has upsides too, though. One is more deterministic
collection of objects and destructor calling. Of course this argument
becomes moot if you have a reference cycle somewhere.

Has anyone ever attempted to implement a garbage collector for
CPython? Or did everyone who wanted this feature move to PyPy?

>>
>> What I would like is having Cython infer whether the GIL is needed for
>> a function, and mark it "implicitly nogil", so it can be called from
>> nogil contexts without actually having to declare it nogil. This would
>> only work for non-extern things, and you would still need to declare
>> it nogil in your pxd if you want to export it. Apparently many users
>> (even those that have used Cython quite a bit) are confused with what
>> nogil on functions actually does (or they are not even aware it
>> exists).
>
> There's a long thread by me and Robert (and some of Stefan) on this from a
> couple of months back, don't know if you read it. You could support exports
> across pxds as well. Basically for *every* cdef function, export two
> function pointers:
>
>  1) To a wrapper to be called if you hold the GIL (outside nogil sections)
>
>  2) To a wrapper to be called if you don't hold the GIL, or don't know
> whether you hold the GIL (the wrapper can acquire the GIL if needed)
>
> Taking the address of a function (for passing to C, e.g.) would give you the
> one that can be called without holding the GIL.
>
> The implications should hopefully be getting rid of "with gil" and "nogil"
> on function declarations entirely.

Oh, this was about functions. I agree that for functions that would be
neat. For inlined code in functions I don't like it very much,
although (unconditional) warnings help a lot in that regard.

However, this would mean that e.g. adding a print statement to your
function makes it acquire the GIL for nogil contexts, and since it
doesn't automatically release it again it may just call another
function that was really supposed to operate without the GIL (because
it's going to/may take a long time). Overall making all this
transparent to the user would be great, people care about their code,
not about how CPython is implemented.

> Dag Sverre
> _______________________________________________
> cython-devel mailing list
> cython-devel at python.org
> http://mail.python.org/mailman/listinfo/cython-devel
>