Re: [PATCH 5/5] sched_ext: Make scx_has_op a bitmap

[Andrea Righi <arighi@nvidia.com>]

On Wed, Apr 09, 2025 at 08:51:24AM -1000, Tejun Heo wrote:
> Hello,
> 
> On Wed, Apr 09, 2025 at 09:57:59AM +0200, Andrea Righi wrote:
> > On Tue, Apr 08, 2025 at 01:06:05PM -1000, Tejun Heo wrote:
> > > scx_has_op is used to encode which ops are implemented by the BPF scheduler
> > > into an array of static_keys. While this saves a bit of branching overhead,
> > > that is unlikely to be noticeable compared to the overall cost. As the
> > > global static_keys can't work with the planned hierarchical multiple
> > > scheduler support, replace the static_key array with a bitmap.
> > > 
> > > In repeated hackbench runs before and after static_keys removal on an AMD
> > > Ryzen 3900X, I couldn't tell any measurable performance difference.
> > 
> > At this point I'm wondering if we should just do something like:
> > 
> >  #define SCX_HAS_OP(op) (scx_ops.op != NULL)
> > 
> > Do you think the bitmap can provide some measurable benefits? For the most
> > frequently used hot paths (enqueue, dispatch, select_cpu, running,
> > stopping) we likely have to fetch scx_ops.op anyway, so cache-wise the
> > difference should be minimal.
> 
> Performance-wise, unlikely. However, we need to be able to enable ops in
> stages during init - CPU hotplug ops need to be enabled before other ops. We
> can do that by setting the op pointers in stages but the code was already
> structured in a way which is easy to convert to bitmap, so that was the path
> of least resistance. We surely can change it.

Ah good point about the init stages. In that case, I agree that it's safer
to introduce the bitmap for now and, later, we can refactor the code to use
pointers if we want. Thanks for the clarification.

Acked-by: Andrea Righi <arighi@nvidia.com>

-Andrea