talloc pools -- if someone feels like it, please comment

[Jeremy Allison]

On Wed, Jan 09, 2008 at 11:22:22PM +0100, Volker Lendecke wrote:
> On Wed, Jan 09, 2008 at 02:30:21PM -0500, Jim McDonough wrote:
> > So what kinds of usage numbers have you seen with this?   In
> > particular, since it's not reused until the whole pool is freed..
> 
> Well, if I only knew. It seems that my benchmarking machine
> (Dual-Core AMD Opteron(tm) Processor 2210, 6GB) is quite
> random in the results it produces. I get results ranging
> from 118MB/sec up to 130MB/sec, all with the same binary. I
> run the standard 4-client netbench run with 600 seconds. I
> thought I would get significant speedup of about 5% with the
> talloc pools, but I can't reproduce that anymore now.

It's very hard to get reproducible numbers with 4 client
runs. There is a lot of varience.

> What it does do is if you run cachegrind, the number of
> Irefs in malloc goes down from about 12 billion to a bit
> over 9 with talloc pools, where 3.0.28 takes 8.6 billion
> Irefs according to cachegrind. But I'm not sure this relates
> in any way to the performance you will see in real benchmark
> runs. 

I'm running cachegrind runs with one client and 100,000
ops to try and get a reproducible set up to compare between
3.0.28, 3.2 and 3.2+pool-patch.

> It's late here, I won't solve this today anymore.

Don't go nuts over this :-). It's still very impressive
(beautiful!) code :-).

Jeremy.