Re: Optimizing Objective-C code

[Richard Frith-Macdonald]

On Friday, March 8, 2002, at 09:27 PM, Chris B. Vetter wrote:

> Hi,
>
> I just stumbled across Mulle Kybernetik, who run a seemingly interesting
> series on ways to make ObjC code run faster.
>
> Might be of interest ...
>
>         http://www.mulle-kybernetik.com/artikel/Optimization/


It was ... I had a look at the chapter on 'optimising foundation 
classes' and
was amazed at the performance increase they managed to get replacing 
NSArray,
and didn't see how that could possibly have been managed.

So I decided to try it under GNUstep.  As I was running on an 850 MHz 
i686
rather than a 450MHz cube, I doubled the number of iterations the test 
was
doing.

The unoptimised code took 2.7 seconds.
The naive optimisation (avoiding method lookup overheads) reduced that 
to 1.9 seconds.
I couldn't see how the rewritten classes they used to replace 
NSArray/NSString would
help much, so I decided to try on my 500MHz MacOS-X system -
the code which took 1.9 seconds on my GNU/Linux machine took around 65 
seconds there
(averaged over a few runs).

Next, I modified the code to detach a thread before the test ...
ensuring that lots of locking was turned on (all reference count
work has to be lock protected in a multithreaded environment).
This pushed the time under GNUstep up to 6.3 seconds
Under MacOS-X the timing was unchanged ... perhaps MacOS-X performs lno 
optimisations
for running single threaded?

Finally, I tried varying the number of iterations down to 5 millionm, 
one million,
and a hundred thousand.
With GNUstep, the timing change was linear ... 6.5 -> 3.2 -> 0.64 -> 0.06
With MacOS-X it was not ... 65 -> 20 -> 3.14 -> 0.24

What does all that tell you?

1. You can't really assume anything about the efficiency of particular 
methods/classes -
you have to try them out.  And they can vary radically from 
implementation to
implementation, or release to release, or even depending on the size of 
the data
being handled!

2. Running multithreaded really slows you down under GNUstep.
An optimisation for really good performance is to avoid using threads or
using any libraries which use threads.  Remember, starting a thread at 
any point will
damage the performance of your program and the performance will not be 
regained when
the thread exits.

3. Enumerating and adding elements to an array is currently very much 
faster in GNUstep than
MacOS-X, even when running threaded... You probably shouldn't bother 
trying to optimise it.