Re: Leopard performance penalty (3x slower), NSPopAutoreleasePool

FROM : Gerd Knops
DATE : Sat Nov 17 20:45:34 2007

On Nov 17, 2007, at 1:12 PM, j o a r wrote:

>
> On Nov 17, 2007, at 10:53 AM, Gerd Knops wrote:
>

>> I have a complex Cocoa application, 10.4 SDK (no GC etc). Loading a 
>> document (and performing a large number of calculations triggered 
>> by that) takes 4 seconds under Tiger.
>>
>> Under Leopard, loading the same document takes 8 seconds until it 
>> displays, after which the application is unresponsive for another 4 
>> seconds, the entire time being spent in NSPopAutoreleasePool.
>>
>> I presume that performance penalty is largely due to Leopards more 
>> complex memory handling. Has anybody else observed this, and knows 
>> of any strategies to minimize that overhead?

>
>
> What "more complex memory handling" are you referring to? Do you 
> compare loading the document between Tiger and Leopard on the same 
> machine (same hardware and amount of RAM)?
>

Yes, at least initially. Now to speed things up I am using a (actually 
slower, but same setup, memory etc) machine to run Tiger, the results 
are the same as when I run on the same machine.

> You should probably investigate trying to create fewer temporary 
> objects, and fewer autoreleased objects, as a way to fix this 
> problem on your end.
>

That would be a less than fun task, given >50.000 lines of code...

I did sprinkle a number of local NSAutoReleasePools, so that 
autoreleased objects do not amass. That made no (measurable) 
difference in performance at all. I no longer have the 4 second period 
of unresponsiveness, but the overall process takes 4 seconds longer, 
so that was a wash. By adding timestamps before and after the 
[autoReleasePool release] I can see that they now roughly share the 
burden (eg take about the same time). So having all the temporary 
objects in one large pool or a number of smaller pools makes no 
difference.

> You may also want to file a performance regression bug report with 
> Apple. I would suggest including: Shark Time Profile, Shark Time 
> Profile (All Thread States), System Profile, top output. Make sure 
> to sample the whole duration of loading the document, and include 
> reports from both Tiger and Leopard.



I doubt that is going to be helpfull, as Shark records the time spent 
in NSPopAutoreleasePool as being in main(), and nothing else really 
jumps out, the slowness is all across the board.

<rant>Also do radars ever actually help? I have spent many hours 
filing detailed radars, and best case (if I get an answer at all) a 
year or so later I hear "we can no longer reproduce this in our 
upcoming release". Somewhat frustrating, I'd rather spend the time 
playing with the dogs... For all the time I invested, I have never 
ever got a single helpful line back (like do x and things will be 
better). Makes me feel like I am spending my time (and money) to help 
Apple debug their stuff, for zero (other than the hope it might be 
better in a year) return. So basically the ROI on radars is extremely 
poor, and I can't make a business case for them. You may call that 
anti-social, but not working for a large company these radars do cost 
me actual dollars. Sorry for the rant...</rant>

Gerd