Re: [PATCH] Fix longstanding load balancing bug in the scheduler.

[Nick Piggin <npiggin@suse.de>]

On Sat, Sep 09, 2006 at 12:56:16PM -0700, Christoph Lameter wrote:
> On Fri, 8 Sep 2006, Christoph Lameter wrote:
> 
> > > one or more, it is unnecessary for the common case.
> > 
> > The common case is an arch with much less cpus. The maxinum on i386
> > f.e. is 255 meaning 8 bytes. That fits in the cacheline that is already
> > used for the stack frame of the calling function. 
> 
> Ughh. Wrong. 255 cpus require 32 bytes. System rarely have that 
> much. If you configure a kernel with less than 32 cpus then this will be 
> one word on the stack.
> 
> Also note that the patch restricts the search to online cpus. The 
> scheduler will check offline cpus without this patch. That may actually 
> result in speed improvements since the cachelines from offline cpus are
> no longer brought in during the search for the busiest group / cpu.

This should not be the case. The sched-domain structure should always
reflect the online CPUs only (see our hotplug cpu handler), and if you
find otherwise then that would be a bug.