processutils.execute is 10 times slower than clean subprocess.run

[Denis Kolvakh]

There is a bug discovered by me in eventlet with incorrect PIPE_BUF size assignment: https://github.com/eventlet/eventlet/issues/813. In short, fallback to default 512 bytes due to absence of PIPE_BUF constant in patched 'select' module decrease I/O operations speed up to 10 times in comparison with Linux-style 4096 bytes.

Due to wide usage of eventlet around Openstack, this issue affects many Openstack components and flows, especially when significant amount of data should be sent/received via Popen.

While issue in eventlet is still orphan, I suggest the following w/a to speed up I/O processes up to 10 times (also attached).

Index: oslo_concurrency/processutils.py
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/oslo_concurrency/processutils.py b/oslo_concurrency/processutils.py
--- a/oslo_concurrency/processutils.py (revision 774f604c16b47ad4ce47e6390ec30f9fc8f30c67)
+++ b/oslo_concurrency/processutils.py (date 1697534335526)
@@ -53,6 +53,16 @@
         subprocess.threading = eventlet.patcher.original('threading')
     else:
         from eventlet.green import subprocess
+ # NOTE(dvkolvakh): during monkey_patch 'eventlet' doesn't care about
+ # constants in built-in 'select' module, that's why 'subprocess'
+ # falls back to default 512 bytes for IO buffer trying to lookup
+ # for non-existing constant in 'eventlet.green.select' module,
+ # see https://github.com/eventlet/eventlet/issues/813.
+ # By default, linux systems has PIPE_BUF = 4096 in 'select' module,
+ # so rely on constant definition in original 'subprocess' module
+ # to speed-up IO operations.
+ subprocess_orig = eventlet.patcher.original('subprocess')
+ subprocess._PIPE_BUF = subprocess_orig._PIPE_BUF

     from eventlet import tpool