The following code copies data, whereas the copy can be avoided using a memory view:

chunk = self._input[self._input_offset:self._input_offset + _PIPE_BUF]
self._input_offset += os.write(key.fd, chunk)

It should be replaced with:

input_view = memoryview(self._input)
...
chunk = input_view[self._input_offset:self._input_offset + _PIPE_BUF]
self._input_offset += os.write(key.fd, chunk)


This issue is a reminder for one of my comment of issue #18923.

Attached subprocess_memoryview.patch uses a memoryview() to avoid memory copies.

The behaviour is undefined if two threads share the same Popen object and call the communicate() method at the same time. But I'm not sure that Popen is thread safe, it doesn't use any lock.

Hi, why does your patch change listobject.c?

> Hi, why does your patch change listobject.c?

Oops, the change is unrelated (#19578).

What is a performance for small chunks?

> What is a performance for small chunks?

I have no idea. Does it matter?

I don't know. But a function call have a cost.

New patch without the unwanted change in listobject.c.

> Serhiy Storchaka added the comment:
>
> I don't know. But a function call have a cost.

I'm with Serhiy here.
Writing a performance optimization without benchmark is generally a
bad idea (I won't quote Knuth :-).

Could someone with a dual-core machine try the attached simplistic
benchmark with and without Victor's patch?
I can see some user-time difference with 'time' on my single-core
machine, but I'm curious to see how this would affect things were both
the parent and the child subprocess can run concurrently.

On a quad-core machine:

- without Victor's patch:

$ time ./python test_sub.py 
0.39263957000002847

real	0m0.521s
user	0m0.412s
sys	0m0.238s

- with Victor's patch:

$ time ./python test_sub.py 
0.3856174530001226

real	0m0.516s
user	0m0.404s
sys	0m0.247s

Best of 10 runs.

Unpatched:  3.91057508099766
Patched:    3.86466505300632

Impressive speedup :-)

New changeset 44948f5bdc12 by Gregory P. Smith in branch '3.3':
Fixes issue #19506: Use a memoryview to avoid a data copy when piping data
http://hg.python.org/cpython/rev/44948f5bdc12

New changeset 5379bba2fb21 by Gregory P. Smith in branch 'default':
Fixes issue #19506: Use a memoryview to avoid a data copy when piping data
http://hg.python.org/cpython/rev/5379bba2fb21

fwiw, this patch made sense.  using test_sub.py on my dual-core amd64 box with an opt build is saw the times drop from ~0.90-0.97 to ~0.75-0.8.  more speedup than i was really anticipating for this use case.  it probably depends upon what the value of _PIPE_BUF is on your system.