Benchmarking

[lemire]

It would be interesting to benchmark against https://github.com/Ezibenroc/PyRoaringBitMap

[sunzhaoping]

I think PyRoaringBitMap is faster than my cffi binding because it using Cython. I recently working for a project of game analysis which using pypy and dask.distributed for distributed computing. I want to make croaring to work with them properly. I think cffi is the best way and wrote a simple wrapper for it , I will try to use PyRoaringBitMap . thank you for your attention.

[sunzhaoping]

It will be great If there have a distributed service of croaring just like memcached

[lemire]

It would be useful to have hard numbers so that users can make a choice based on real data.

[lemire]

Please see https://github.com/Ezibenroc/PyRoaringBitMap/issues/22#issuecomment-326619594

[Ezibenroc]

@sunzhaoping I think some performance issues can easily be fixed. For instance:

• call roaring_bitmap_from_range for the constructor when the iterable is a range
• call roaring_bitmap_is_subset to implement the method __le__ and roaring_bitmap_is_strict_subset to implement the method __lt__.

Ping me if you fix this, I would be glad to re-run my benchmark!

[sunzhaoping]

Good! thank you for your suggestion.

[sunzhaoping]

@Ezibenroc I have optimized the code according to your suggestion.

[Ezibenroc]

@sunzhaoping the new benchmark results are much better:

operation	pyroaring	python-croaring
range constructor	1.59e-04	1.81e-04
ordered list constructor	3.24e-02	6.03e-02
list constructor	1.23e-01	1.28e-01
ordered array constructor	3.55e-03	4.57e-03
array constructor	1.14e-01	1.15e-01
element addition	8.95e-07	3.86e-05
element removal	2.27e-06	1.62e-06
membership test	9.69e-08	1.86e-06
union	1.88e-04	2.14e-04
intersection	8.85e-04	9.33e-04
difference	1.81e-04	2.14e-04
symmetric diference	1.85e-04	2.13e-04
equality test	8.75e-05	8.05e-05
subset test	1.02e-04	8.34e-05
conversion to list	4.08e-02	2.77e-01
pickle dump & load	4.99e-04	6.90e-04
"naive" conversion to array	4.39e-02	2.91e-01
"optimized" conversion to array	1.32e-03	3.38e-02
selection	9.12e-07	4.42e-05
slice	4.81e-02	3.15e-01

There is no more absurdly high times for python-croaring. What is interesting however is that it is always a bit slower than pyroaring, it seems that using cffi causes a small overhead.

[sunzhaoping]

I think many Type Casting in cffi is using pure python code , All pyroaring code is compiled to c code. So pyroaring can do better. If using pypy, the effect will be better.

[lemire]

Yes, casting is likely the issue. Going from an integer to an uint32_t seems to involve a lot of work.