[POLLING] strange interrupt/system load

[Barney Cordoba]

--- On Tue, 9/15/09, peterjeremy at acm.org <peterjeremy at acm.org> wrote:

> From: peterjeremy at acm.org <peterjeremy at acm.org>
> Subject: Re: [POLLING] strange interrupt/system load
> To: "Barney Cordoba" <barney_cordoba at yahoo.com>
> Cc: freebsd-net at freebsd.org
> Date: Tuesday, September 15, 2009, 3:38 AM
> On 2009-Sep-13 07:19:24 -0700, Barney
> Cordoba <barney_cordoba at yahoo.com>
> wrote:
> >64bits must be faster than 32bits is patently
> misguided. My rule of 
> >thumb is that if I don't need 64bits for something, I
> avoid it.
> 
> It's not quite that cut-and-dry.  The 64-bit ISA is
> significantly
> different to the 32-bit ISA and has different subroutine
> calling
> conventions.  Yes, you do need to lug 64-bit pointers
> around but the
> overall codesize is comparable (looking at /usr/bin and
> /lib suggests
> about a 5% increase in size going from i386 to amd64) - a
> lot of this
> is probably because amd64 has a 16-bit offset mode so
> there's much
> less need for 32-bit offsets.  Having twice as many
> registers is a
> win in some areas (less spilling to memory) and a loss in
> others (more
> state to save/restore on a context switch).
> 
> If performance is critical, it's probably worthwhile
> benchmarking
> both i386 and amd64 variants and seeing which works best
> for you.
> 
"Rules of Thumb" are generally for those times when you don't have
a pressing preference and you don't want to spend your life endlessly
benchmarking.

I don't think its the code, necessarity, but rather the significant
increase in the size of data structures, and the memory that has to
be moved around.

I haven't tried with the latest compiler but I can't see why it would
have any benefit for systems used for high capacity networking other than
incrementing counters. 

Barney