Commit c6addb05 authored by Kirill Smelkov's avatar Kirill Smelkov

X Switch Zope to use poll() instead of select()

Recently we are starting to get following errors on a production
instance:

      File ".../bin/runzope", line 211, in <module>
        sys.exit(Zope2.Startup.run.run())
      File ".../eggs/Zope2-2.13.24-py2.7.egg/Zope2/Startup/run.py", line 26, in run
        starter.run()
      File ".../eggs/Zope2-2.13.24-py2.7.egg/Zope2/Startup/__init__.py", line 111, in run
        Lifetime.loop()
      File ".../eggs/Zope2-2.13.24-py2.7.egg/Lifetime/__init__.py", line 43, in loop
        lifetime_loop()
      File ".../eggs/Zope2-2.13.24-py2.7.egg/Lifetime/__init__.py", line 53, in lifetime_loop
        asyncore.poll(timeout, map)
      File ".../parts/python2.7/lib/python2.7/asyncore.py", line 145, in poll
        r, w, e = select.select(r, w, e, timeout)
    ValueError: filedescriptor out of range in select()

Initially we thought the reason here is that number of file descriptors this
process uses goes beyond allowed limit (65K open files on that instance)
because wendelin.core used 1 fd per an array view (opening
/dev/shm/ramh.XXXXX) but that turned out to be not strictly that case:

The reason here is that select() has limit for how many #fd it can
monitor at all. The limit is system-specific but on Linux it is usually
1024 - http://man7.org/linux/man-pages/man2/select.2.html#BUGS :

    $ cat s.c
    #include <sys/select.h>
    #include <stdio.h>

    int main() {
            printf("%d\n", FD_SETSIZE);
            return 0;
    }

    $ tcc -run s.c
    1024

Also select() can monitor only file descriptors which are by itself are
"< FD_SETSIZE", e.g. it cannot monitor fd=1025 even if there are only 10
opened file descriptors because 1025 is not < 1024.

As was said above in wendelin.core every array view uses 1 file
descriptor, so if we are using not so small amount of arrays, even
though #fd does not go beyond process-specific ulimit, because of
select() usage the total number of allowed opened file descriptors
becomes essentially 1024.

So let's switch from select() to poll() which does not have this 1024
 #fd limit. Asyncore already support using poll() out of the box -
either via passing use_pull=True to asyncore.loop()

https://docs.python.org/2/library/asyncore.html#asyncore.loop

or by using asyncore.poll2() instead of asyncore.poll()

https://github.com/python/cpython/blob/2.7/Lib/asyncore.py#L170
https://github.com/python/cpython/blob/2.7/Lib/asyncore.py#L125

--------

This patch switches asyncore.poll() -> asyncore.poll2() for only 1 place
in Zope. There are however many such places in Zope and other software.

For this reason, for me, what makes sense to do is not to patch all such
places, but instead change via runtime-patching asyncore.poll to be
asyncore.poll2 - this way all software will automatically benefit from
poll() usage instead of select.

P.S.

What might also make sense to do in the future is to actually let
asyncore.poll to use epoll(), as both select() and poll() are doing all
fd registration on _every_ call, so when #fd grows this preparation
time grows too. For epoll() file descriptors are registered only once.
For this to work asyncore.socket_map has to be patched also, since there
are places in code which modify this fd registry directly (e.g. remove
fd from there etc)
parent ecfd7b1b
...@@ -50,7 +50,7 @@ def lifetime_loop(): ...@@ -50,7 +50,7 @@ def lifetime_loop():
map = asyncore.socket_map map = asyncore.socket_map
timeout = 30.0 timeout = 30.0
while map and _shutdown_phase == 0: while map and _shutdown_phase == 0:
asyncore.poll(timeout, map) asyncore.poll2(timeout, map)
def graceful_shutdown_loop(): def graceful_shutdown_loop():
......
  • Switch Zope to use poll() instead of select()

    /cc @klaus, @Tyagov, @jm, @arnau, @yusei, @nexedi

  • Typical select gotcha :) : data structure proportional to the value of the monitored descriptor and not the number of monitored descriptor.

    +1 on the principle, and I trust you for the practical part.

  • @vpelletier thanks for feedback.

  • For the reference: on a production project @klaus manages we patched cpython's stdlib this way:

    --- lib/python2.7/asyncore.py.kirr      2017-02-03 13:18:01.000000000 +0100
    +++ lib/python2.7/asyncore.py   2017-07-06 21:15:31.858839043 +0200
    @@ -657,3 +657,13 @@
                 self.socket = file_wrapper(fd)
                 self._fileno = self.socket.fileno()
                 self.add_channel()
    +
    +
    +# monkeypatch to switch everyone's default to use poll instead of select
    +# (select has 1024 fd limit)
    +# TODO it is better to use epoll
    +_poll_select = poll
    +_poll_poll   = poll2
    +poll = _poll_poll
    +
    +print >>sys.stderr, 'USING POLL INSTEAD OF SELECT'

    and it is running stable for 3 days already.

  • It is running stable for many months.

    (the patch above was extracted by gendiff-nxd0 srv/runner/software/... .kirr - see misc@2ac16a8f)

  • The patch to switch cpython's stdlib asyncore default from select to poll was merged to slapos master: nexedi/slapos@e5c107c4 .

Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment