[Python-Dev] "Fixing" the new GIL

[Nir Aides]

Hello all,

I would like to kick this discussion back to life with a simplified
implementation of the BFS scheduler, designed by the Linux kernel hacker Con
Kolivas: http://ck.kolivas.org/patches/bfs/sched-BFS.txt

I submitted bfs.patch at http://bugs.python.org/issue7946. It is work in
progress but is ready for some opinion.

On my machine BFS gives comparable performance to gilinter, and seems to
schedule threads more fairly, predictably, and with lower rate of context
switching. Its basic design is very simple but nevertheless it was designed
by an expert in this field, two characteristics which combine to make it
attractive to this case.

The problem addressed by the GIL has always been *scheduling* threads to the
interpreter, not just controlling access to it, and therefore the GIL, a
lock implemented as a simple semaphore was the wrong solution.

The patches by Antoine and David essentially evolve the GIL into a
scheduler, however both cause thread starvation or high rate of context
switching in some scenarios:

With Floren't write test (http://bugs.python.org/issue7946#msg101120):
2 bg threads, 2 cores set to performance, karmic, PyCon patch, context
switching shoots up to 200K/s.
2 bg threads, 1 core, set to on-demand, karmic, idle machine, gilinter patch
starves one of the bg threads.
4 bg threads, 4x1 core xeon, centos 5.3, gilinter patch, all bg threads
starved, context switching shoots up to 250K/s.

With UDP test (http://bugs.python.org/file16316/udp-iotest.py), add
zlib.compress(b'GIL') to the workload:
both gilinter and PyCon patches starve the IO thread.

The BFS patch currently involves more overhead by reading the time stamp on
each yield and schedule operations. In addition it still remains to address
some issues related to timestamps such as getting different time stamp
readings on different cores on some (older) multi-core systems.

Any thoughts?

Nir



On Sun, Mar 14, 2010 at 12:46 AM, Antoine Pitrou <solipsis at pitrou.net>wrote:

>
> Hello,
>
> As some of you may know, Dave Beazley recently exhibited a situation
> where the new GIL shows quite a poor behaviour (the old GIL isn't very
> good either, but still a little better). This issue is followed in
> http://bugs.python.org/issue7946
>
> This situation is when an IO-bound thread wants to process a lot of
> incoming packets, while one (or several) CPU-bound thread is also
> running. Each time the IO-bound thread releases the GIL, the CPU-bound
> thread gets it and keeps holding it for at least 5 milliseconds
> (default setting), which limits the number of individual packets which
> can be recv()'ed and processed per second.
>
> I have proposed two mechanisms, based on the same idea: IO-bound
> threads should be able to steal the GIL very quickly, rather than
> having to wait for the whole "thread switching interval" (again, 5 ms
> by default). They differ in how they detect an "IO-bound threads":
>
> - the first mechanism is actually the same mechanism which was
>  embodied in the original new GIL patch before being removed. In this
>  approach, IO methods (such as socket.read() in socketmodule.c)
>  releasing the GIL must use a separate C macro when trying to get the
>  GIL back again.
>
> - the second mechanism dynamically computes the "interactiveness" of a
>  thread and allows interactive threads to steal the GIL quickly. In
>  this approach, IO methods don't have to be modified at all.
>
> Both approaches show similar benchmark results (for the benchmarks
> that I know of) and basically fix the issue put forward by Dave Beazley.
>
> Any thoughts?
>
> Regards
>
> Antoine.
>
>
> _______________________________________________
> Python-Dev mailing list
> Python-Dev at python.org
> http://mail.python.org/mailman/listinfo/python-dev
> Unsubscribe:
> http://mail.python.org/mailman/options/python-dev/nir%40winpdb.org
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-dev/attachments/20100412/999d297c/attachment-0001.html>