powerpc kernel flavour

[Tim Gardner]

On 12/06/2010 02:28 AM, Jeremy Kerr wrote:
> Hi all,
>
> Last UDS, we discussed the possibility of dropping the -powerpc kernel
> flavour, as the powerpc-smp kernel will support the same set of machines, and
> means one less flavour to build.
>
> The potential downside of this is that we lose some optimisations for
> uniprocessor machines; for example, spinlocks will not be compiled-out, as can
> be done with the UP kernel.
>
> I've done some brief benchmarking using sysbench's 'thread' test:
>
> 	sysbench --max-requests=100000 --num-threads=64 \
> 		 --test=threads --thread-yields=100 --thread-locks=2 run
>
>
> On my g4 machine, I see a 3.5% performance degradation when using the powerpc-
> smp flavour (which is statistically significant). However, this is a kernel-
> heavy benchmark; we may not see such a change on real workloads. I'm not using
> kernbench as this would be very much disk-bound on this machine.
>
> So, it boils down to: do we care about this 3.5% enough to keep the powerpc
> flavour around?
>
> Cheers,
>
>
> Jeremy
>

I'm all for dropping flavours, but I have no idea how many G4s there are 
still running, and if 3.5% has a noticeable impact.

Doesn't x86 rewrite locks at early boot time depending on UP or SMP? Is 
that something we could do for powerpc ? Though its likely the 
dependency on CONFIG_PPC32=y goes deeper then just spinlocks.

rtg
-- 
Tim Gardner tim.gardner at canonical.com