Re: [PATCH] continous statistics

Hi Krzysztof !

On Sat, Nov 24, 2007 at 10:16:49PM +0100, Krzysztof Oledzki wrote:
> >From baff40f9c50af1f62ff02d8c8269bd6a23157717 Mon Sep 17 00:00:00 2001
> From: Krzysztof Piotr Oledzki <ole#ans.pl>
> Date: Sat, 24 Nov 2007 22:12:47 +0100
> Subject: [MEDIUM] continous statistics

Thanks, I've merged your patch. BTW, thanks for having updated the doc! I tried again, and it's working fine.

Also, as I never trust gcc's code on 64-bit data, I took a look at what it produced for session_process_counters, and (as expected), it clearly was suboptimal. The problem is always the same: gcc knows very few things when optimizing 64-bit for a 32-bit platform. For instance, it does not know that a == b if (a - b) == 0.

So after merging your patch, I slightly tweaked the code to help gcc, and by a single small trick, the function above went down from 318 bytes to 220 bytes, the difference relying in data passing from register to reg/memory, and useless comparisons and value reloading.

I measured a net gain of almost 2% using this change, which I cannot attribute to this code alone, so I think that it simply changed some other function's alignment resulting in this difference.

I've attached the patch for your convenience.

For your info, I finally managed to merge the dynamic weights code. It was kind of hard, the initial code looked like a mine field! But now it works (I added a quick and dirty patch to change a server's weight on the fly to verify this). I'm currently merging the branches back to master. The changes are not yet well documented, but they only affect developpers. I still have two features to implement this week (slowstart and server maintenance mode) and I'll be able to release 1.3.14 with all of this.

I'll keep you updated about this.

Regards,
Willy