Date: Tue, 19 Jan 2010 12:46:17 -0600
From: Dan Nelson <dnelson@allantgroup.com>
To: Bernard van Gastel <bvgastel@bitpowder.com>
Cc: freebsd-hackers@freebsd.org
Subject: Re: pthread_{mutex,cond} & fifo/starvation/scheduling policy
Message-ID: <20100119184617.GB50360@dan.emsphone.com>
In-Reply-To: <71A129DC-68A0-46C3-956D-C8AFF1BA29E1@bitpowder.com>
References: <71A129DC-68A0-46C3-956D-C8AFF1BA29E1@bitpowder.com>
next in thread | previous in thread | raw e-mail | index | archive | help
In the last episode (Jan 19), Bernard van Gastel said:
> I'm curious to the exact scheduling policy of POSIX threads in relation to
> mutexes and conditions. If there are two threads (a & b), both with the
> following code:
>
> while (1) {
> pthread_mutex_lock(mutex);
> ...
> pthread_mutex_unlock(mutex);
> }
>
> What is the scheduling policy of the different thread libraries? Are both
> threads getting an equal amount of time? Are there no starvation issues
> (are they executed in alternating turns)? (a test program of mine
> indicates that libpthread and libthr both have starvation issues, in
> contrary to Mac OS X 10.6)
There's no guarantee of fairness when dealing with mutexes afaik. My guess
is that if thread "a" unlocks the mutex and still has time left in its
quantum, it'll be able to grab it again without even going to the kernel.
>From the POSIX docs on mutexes:
http://www.opengroup.org/onlinepubs/9699919799/functions/pthread_mutex_lock.html#tag_16_439_08
"Mutex objects are intended to serve as a low-level primitive from which
other thread synchronization functions can be built. As such, the
implementation of mutexes should be as efficient as possible, and this
has ramifications on the features available at the interface.
The mutex functions and the particular default settings of the mutex
attributes have been motivated by the desire to not preclude fast,
inlined implementations of mutex locking and unlocking."
The idea being that mutexes should be held for as little time as possible.
Is there a way to write your code so that most of the CPU-consuming activity
is done outside of the mutex? Perhaps use a job queue of some sort, and
only lock the mutex when pushing/popping elements. Then worker processes
can run without holding the mutex, and will be fairly scheduled by the
kernel.
--
Dan Nelson
dnelson@allantgroup.com
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20100119184617.GB50360>