benchmarking in general and using xheap
Sven R. Kunze
srkunze at mail.de
Fri Feb 19 04:52:23 EST 2016
Hi everybody,
I've finally had the time to do the benchmarks and here you go:
http://srkunze.blogspot.com/2016/02/the-xheap-benchmark.html
The benchmark compares heapq, Heap, OrderHeap, RemovalHeap and XHeap
regarding their operation heapify, push and pop.
As expected wrapping results in some overhead. Most of the overhead
consists of wrapper, super and descriptor calls. As with the current
optimizations efforts, I expect this to be reduced even further. But
even using current CPython 2.7 or 3.5, the overhead for simple heaps,
heaps with a custom orders or heaps with removal can be considered
affordable given the maintenance benefits.
@srinivas
The current removal implementation uses a index-tracking approach with
quite some overhead for other operations. I am not sure if that is
remediable with a mark-and-sweep approach but given the time I will
definitely look into it for another benchmark post now that I have build
the infrastructure for it.
@all benchmark friends
Not sure how you do your benchmarks, but I am somewhat dissatisfied with
the current approach. I started out using unittests as they integrated
nicely with my current toolchain and I could write actual code. Then I
threw anything away and used timeit as suggested here
https://mail.python.org/pipermail/python-list/2016-January/702571.html
and grew my own set of tools around it to produce readable results;
writing code as strings. :-/
And from what I know, there is no official infrastructure (tools,
classes, ... such as there is for unittests) around timeit to
encapsulate benchmarks, choosing a baseline, calculate ratios etc (and
write code instead of strings).
Does somebody have an idea here?
Best,
Sven
More information about the Python-list
mailing list