Re: [PATCH net-next] tuntap: introduce tx skb ring

[Michael S. Tsirkin]

On Wed, May 18, 2016 at 06:42:10PM +0800, Jason Wang wrote:
> 
> 
> On 2016å05æ18æ 17:55, Michael S. Tsirkin wrote:
> >On Wed, May 18, 2016 at 11:21:29AM +0200, Jesper Dangaard Brouer wrote:
> >>On Wed, 18 May 2016 11:21:59 +0300
> >>"Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:
> >>
> >>>On Wed, May 18, 2016 at 10:16:31AM +0200, Jesper Dangaard Brouer wrote:
> >>>>On Tue, 17 May 2016 09:38:37 +0800 Jason Wang <jasowang@xxxxxxxxxx> wrote:
> >>>>>>>And if tx_queue_length is not power of 2,
> >>>>>>>we probably need modulus to calculate the capacity.
> >>>>>>Is that really that important for speed?
> >>>>>Not sure, I can test.
> >>>>In my experience, yes, adding a modulus does affect performance.
> >>>How about simple
> >>>	if (unlikely(++idx > size))
> >>>		idx = 0;
> >>So, you are exchanging an AND-operation with a mask, for a
> >>branch-operation.  If the branch predictor is good enough in the CPU
> >>and code-"size" use-case, then I could be just as fast.
> >>
> >>I've actually played with a lot of different approaches:
> >>  https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/include/linux/alf_queue_helpers.h
> >>
> >>I cannot remember the exact results. I do remember micro benchmarking
> >>showed good results with the advanced "unroll" approach, but IPv4
> >>forwarding, where I know I-cache is getting evicted, showed best
> >>results with the more simpler implementations.
> >This is all assuming you can somehow batch operations.
> >We can do this for transmit sometimes (when linux
> >is the source of the packets) but not always.
> >
> >>>>>Right, this sounds a good solution.
> >>>>Good idea.
> >>>I'm not that sure - it's clearly wasting memory.
> >>Rounding up to power of two.  In this case I don't think the memory
> >>wast is too high.  As we are talking about max 16 bytes elements.
> >It almost doubles it.
> >E.g. queue size of 10000 (rather common) will become 16K, wasting 6K.
> 
> It depends on the user, e.g default tx_queue_len is around 1000 for real
> cards. If we really care about the wasting, we can add a threshold and fall
> back to normal linked list during resizing.

That looks like a lot of complexity.

> >
> >>I am concerned about memory in another way. We need to keep these
> >>arrays/rings small, due to data cache usage.  A 4096 ring queue is bad
> >>because e.g. 16*4096=65536 bytes, and typical L1 cache is 32K-64K. As
> >>this is a circular buffer, we walk over this memory all the time, thus
> >>evicting the L1 cache.
> >Depends on the usage I guess.
> >Entries pointed to are much bigger, and you are
> >going to access them - is this really an issue?
> >If yes this shouldn't be that hard to fix ...
> >
> >>-- 
> >>Best regards,
> >>   Jesper Dangaard Brouer
> >>   MSc.CS, Principal Kernel Engineer at Red Hat
> >>   Author of http://www.iptv-analyzer.org
> >>   LinkedIn: http://www.linkedin.com/in/brouer