Re: [virtio-dev] packed ring layout proposal v2

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Mar 07, 2017 at 04:53:53PM +0100, Cornelia Huck wrote:
> On Wed, 22 Feb 2017 18:43:05 +0200
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > On Thu, Feb 09, 2017 at 05:11:05PM +0100, Cornelia Huck wrote:
> > > > >>> * Non power-of-2 ring sizes
> > > > >>>
> > > > >>> As the ring simply wraps around, there's no reason to
> > > > >>> require ring size to be power of two.
> > > > >>> It can be made a separate feature though.
> > > > >>
> > > > >> Power of 2 ring sizes are required in order to ignore the high bits of
> > > > >> the indices.  With non-power-of-2 sizes you are forced to keep the
> > > > >> indices less than the ring size.
> > > > > 
> > > > > Right. So
> > > > > 
> > > > > 	if (unlikely(idx++ > size))
> > > > > 		idx = 0;
> > > > > 
> > > > > OTOH ring size that's twice larger than necessary
> > > > > because of power of two requirements wastes cache.
> > > > 
> > > > I don't know.  Power of 2 ring size is pretty standard, I'd rather avoid
> > > > the complication and the gratuitous difference with 1.0.
> > > 
> > > I agree. I don't think dropping the power of 2 requirement buys us so
> > > much that it makes up for the added complexity.
> > 
> > I recalled why I came up with this. The issue is cache associativity.
> > Recall that besides the ring we have event suppression
> > structures - if we are lucky and things run at the same speed
> > everything can work by polling keeping events disabled, then
> > event suppression structures are never written to, they are read-only.
> > 
> > However if ring and event suppression share a cache line ring accesses
> > have a chance to push the event suppression out of cache, causing
> > misses on read.
> > 
> > This can happen if they are at the same offset in the set.
> > E.g. with L1 cache 4Kbyte sets are common, so same offset
> > within a 4K page.
> > 
> > We can fix this by making event suppression adjacent in memory, e.g.:
> > 
> > 
> > [interrupt suppress]
> > [descriptor ring]
> > [kick suppress]
> > 
> > If this whole structure fits in a single set, ring accesses will
> > not push kick or interrupt suppress out of cache.
> > Specific layout can be left for drivers, but as set size is
> > a power of two this might require a non-power of two ring size.
> > 
> > I conclude that this is an optimization that needs to be
> > benchmarked.
> 
> This makes sense. But wouldn't the optimum layout not depend on the
> platform?

There's generally a tradeoff between performance and portability.
Whether it's worth it would need to be tested.
Further, it might be better to have platform-specific optimization
tied to a given platform rather than a feature bit.

> > 
> > I also note that the generic description does not have to force
> > powers of two *even if devices actually require it*.
> > I would be inclined to word the text in a way that makes
> > relaxing the restriction easier.
> > 
> > For example, we can say "free running 16 bit index" and this forces a
> > power of two, but we can also say "free running index wrapping to 0
> > after (N*queue-size - 1) with N chosen such that the value fits in 16
> > bit" and this is exactly the same if queue size is a power of 2.
> > 
> > So we can add text saying "ring size MUST be a power of two"
> > and later it will be easy to relax just by adding a feature bit.
> 
> A later feature bit sounds good.

No need to delay benchmarking if someone has the time though :)

-- 
MST