[pypy-dev] Idea for speed.pypy.org

[Maciej Fijalkowski]

On Mon, Dec 13, 2010 at 10:51 PM, Miquel Torres <tobami at googlemail.com> wrote:
> @Maciej: it doesn't make a lot of sense. Looking at this graph:
> http://speed.pypy.org/comparison/?exe=2%2B35,4%2B35,1%2B172,3%2B172&ben=11,14,15&env=1&hor=false&bas=none&chart=normal+bars
>
> slowspitfire is much faster than the other two. Is that because it
> performs more iterations?

I think it's apples to oranges (they have different table sizes and
different number of iterations)

>
> Also, how come pypy-c-jit is faster than cpython or psyco precisely in
> cstringio, where performance should be dependent on cstringIO and thus
> be more similar across interpreters?

because having a list of small strings means you have a large (old)
object referencing a lot of young objects, hence GC cost. It's not the
case with cstringio where you have a single chunk of memory which does
not contain GC pointers.

>
>
> 2010/12/13 Leonardo Santagada <santagada at gmail.com>:
>> why not have only 2 versions, both with the same size table and name
>> one spitfire_cstringio and the other spitfire_strjoin? I think it
>> would make things clearer.
>>
>>
>> On Mon, Dec 13, 2010 at 2:43 PM, Maciej Fijalkowski <fijall at gmail.com> wrote:
>>> Hi.
>>>
>>> spitfires are confusing.
>>>
>>> slowspitfire and spitfire use ''.join(list-of-strings) where
>>> spitfire_cstringio uses cStringIO instead.
>>>
>>> spitfire and spitfire_cstringio use smaller table to render (100x100 I
>>> think) which was the default on original benchmarks
>>>
>>> slowspitfire uses 1000x1000 (which is why it used to be slower than
>>> spitfire) and was chosen by US guys to let the JIT warm up. We should
>>> remove _slow these days.
>>>
>>> On Mon, Dec 13, 2010 at 5:08 PM, Miquel Torres <tobami at googlemail.com> wrote:
>>>> sorry, I meant the opposite. To recap, according to
>>>> http://code.google.com/p/unladen-swallow/wiki/Benchmarks,
>>>> spitfire: psyco
>>>> slowspitfire: pure python
>>>>
>>>> in addition we have spitfire_cstringio, which uses a c module (so it
>>>> is even faster).
>>>>
>>>> what is vanilla spitfire in our case?
>>>>
>>>>
>>>> 2010/12/13 Miquel Torres <tobami at googlemail.com>:
>>>>> @Carl Friedrich & exarkun: thanks, I've added those.
>>>>>
>>>>> only spectral-norm, slowspitfire and ai to go.
>>>>>
>>>>> slowspitfire is described at the Unladen page as using psyco, but it
>>>>> doesn't make sense in our case?
>>>>>
>>>>>
>>>>>
>>>>> 2010/12/13  <exarkun at twistedmatrix.com>:
>>>>>> On 08:20 am, tobami at googlemail.com wrote:
>>>>>>>
>>>>>>> Thanks all for the input.
>>>>>>> I've compiled a list based on your mails, the Unladen benchmarks page
>>>>>>> (http://code.google.com/p/unladen-swallow/wiki/Benchmarks), and the
>>>>>>> alioth descriptions. Here is an extract of the current speed.pypy.org
>>>>>>> admin:
>>>>>>>
>>>>>>> [snip]
>>>>>>> twisted_iteration
>>>>>>
>>>>>> Iterates a Twisted reactor as quickly as possible without doing any work.
>>>>>>>
>>>>>>> twisted_names
>>>>>>
>>>>>> Runs a DNS server with Twisted Names and then issues requests to it over
>>>>>> loopback UDP.
>>>>>>>
>>>>>>> twisted_pb
>>>>>>
>>>>>> Runs a Perspective Broker server with a no-op method and invokes that method
>>>>>> over loopback TCP with some strings, dictionaries, and tuples as arguments.
>>>>>>
>>>>>
>>>> _______________________________________________
>>>> pypy-dev at codespeak.net
>>>> http://codespeak.net/mailman/listinfo/pypy-dev
>>>>
>>> _______________________________________________
>>> pypy-dev at codespeak.net
>>> http://codespeak.net/mailman/listinfo/pypy-dev
>>
>>
>>
>> --
>> Leonardo Santagada
>>
>