On Fri, 4 Mar 2016 08:36:44 -0800
Alexei Starovoitov <alexei.starovoi...@gmail.com> wrote:
> On Fri, Mar 04, 2016 at 02:01:14PM +0100, Jesper Dangaard Brouer wrote:
> > This patchset use the bulk ALLOC side of the kmem_cache bulk APIs, for
> > SKB allocations. The bulk free side got enabled in merge commit
> > 3134b9f019f2 ("net: mitigating kmem_cache free slowpath").
> >
> > The first two patches is a followup on the free-side, which enables
> > bulk-free in the drivers mlx4 and mlx5 (dev_kfree_skb -> napi_consume_skb).
> >
> > Rest of patchset is focused on bulk alloc-side. We start with a
> > conservative bulk alloc of 8 SKB, which all drivers using the
> > napi_alloc_skb() call will benefit from. Then the API is extended to,
> > allow driver hinting on needed SKBs (only some drivers know this
> > size), and mlx5 driver is the first user of hinting.
>
> patches 1-5 look very good to me. Should help all cases afaik.
> As far as 6-7 about hints I have a question. Does this hint
> actually makes the difference? The fixed bulk alloc of 8 probably
> easier for the main slub, but when mlx5 starts doing 'work_done' as
> a hint there will be more 'random' bulking going on.
> Was wondering whether you have the perf numbers to back up 6/7
Yes, it makes a difference. I did some performance numbers with
dropping in the mlx5 driver, plus the RX loop cache-miss avoidance.
With all my optimizations I reached 12Mpps, with this hint optimization
I could reach 13Mpps. It sounds nice also percentage wise (8.3%), but
in nanosec this optimization "only" corresponds to 6.4 ns. For real
workloads, we might see a higher "nanosec" improvement, as this invoke
kmem_cache_alloc_bulk() less times resulting in less icache-misses.
So, yes it makes a difference.
--
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
