[Python-Dev] io.BytesIO slower than monkey-patching io.RawIOBase

[Serhiy Storchaka]

On 17.07.12 06:34, Eli Bendersky wrote:
> The second approach is consistently 10-20% faster than the first one
> (depending on input) for trunk Python 3.3
>
> Is there any reason for this to be so? What does BytesIO give us that
> the second approach does not (I tried adding more methods to the patched
> RawIOBase to make it more functional, like seekable() and tell(), and it
> doesn't affect performance)?

BytesIO resizes underlying buffer if it overflowed (overallocating 1/8 
of size and copying old content to new buffer). Total it makes 
log[9/8](N) allocations and copy 8*N bytes (for large N). List uses the 
same strategy, but number of chunks usually significantly less than 
number of bytes. At the end all this chunks concatenated by join, which 
calculates sum of chunk lengths and allocate the resulting array with 
the desired size. That is why append/join is faster than BytesIO in this 
case.

There are other note, about ElementTree.tostringlist(). Creating 
DataStream class in every function call is too expensive, and that is 
why "monkeypatched" version several times is faster than DataStream 
version for small data. But for long data it is faster too, because 
data.append() is on one lookup slower than "monkeypatched" 
write=data.append.

> This also raises a "moral" question - should I be using the second
> approach deep inside the stdlib (ET.tostring) just because it's faster?

Please note that the previous version of Python used a "monkeypatching".