Re: Semantics of copying mutex/rwlock

[Sad Clouds <cryintothebluesky%googlemail.com@localhost>]

On Wed, 16 Jun 2010 13:21:49 -0400 (EDT)
der Mouse <mouse%Rodents-Montreal.ORG@localhost> wrote:

> > OK I wrote a simple benchmark.  It turns out malloc() is quite
> > expensive, it's about 6 times slower to call malloc() for each
> > mutex, compared to init/destroy pair.
> 
> (a) But you need to malloc only the new mutexes, whereas you need to
> destroy/init all, including the old ones.  Or do you mean you profiled
> actual use and found this difference between the two paradigms?
> 

I timed malloc() for 1000 mutex objects.

Then I timed 2x init() + 2x destroy() mutex calls for 1000 mutex
objects.

The time for malloc run was 6 times longer. So eventually the time
difference will probably level off, but it'll start happening after I
increase the number of mutex objects quite a few times.

Actually I've decided on the optimal solution:

1) malloc() not individual mutex objects, but arrays of objects. Join
allocations into a linked list.

2) Have a single array of pointers to mutex objects. Every time you
need to grow it, just call realloc() and set mutex pointers.


[12345678]->[12345678]->NULL /* linked list of arrays with 8 objects */

[^^^^^^^    ^^^^^^^^] /* main array of pointers to mutex objects */