Re: POSIX rwlocks vs. glib2 rwlocks

[Sebastian Wilhelmi <wilhelmi ira uka de>]

Hi Erik,

> While investigating the g_type_instance_is_a() performance problem, I
> looked into how rwlocks are done in glib2.  I was surprised to learn
> that there is no implementation of rwlocks that uses the pthread
> functions that implement this primitive.  glibc-2.2 has rwlocks fully
> implemented from what I can see in the library and headers (though I
> haven't confirmed that they are fully functional, I would assume so),
> and I have little doubt that they're faster than the routines currently
> in glib.

I made a test program, and here are the results for 1000000 locks and unlocks.

GLib Reader locks: 2.6676460000s
GLib Writer locks: 2.6764440000s
PThread Reader locks: 2.8518880000s
PThread Writer locks: 2.5612170000s

So I think we can continue using out home brewed version. The point in not
using the pthread version is, that our thread system is fully replaceable by a
user provided implementation. That would mean to either 

1) let rwlocks not be replaceable or

2) add those 8 functions to the already fat GThreadFunc vector

Both options are not that nice for a dubious win.

Bye,
Sebastian
-- 
Sebastian Wilhelmi
mailto:wilhelmi ira uka de
http://goethe.ira.uka.de/~wilhelmi