On Thu, Sep 15, 2016 at 04:09:26PM -0700, Eric Dumazet wrote:
> On Thu, 2016-09-15 at 19:11 -0300, Thadeu Lima de Souza Cascardo wrote:
> > Instead of using flow stats per NUMA node, use it per CPU. When using
> > megaflows, the stats lock can be a bottleneck in scalability.
> >
> > On a E5-2690 12-core system, usual throughput went from ~4Mpps to
> > ~15Mpps when forwarding between two 40GbE ports with a single flow
> > configured on the datapath.
> >
> > This has been tested on a system with possible CPUs 0-7,16-23. After
> > module removal, there were no corruption on the slab cache.
> >
> > Signed-off-by: Thadeu Lima de Souza Cascardo <casca...@redhat.com>
> > Cc: pravin shelar <pshe...@ovn.org>
> > ---
>
> > + /* We open code this to make sure cpu 0 is always considered */
> > + for (cpu = 0; cpu < nr_cpu_ids; cpu = cpumask_next(cpu,
> > cpu_possible_mask))
> > + if (flow->stats[cpu])
> > kmem_cache_free(flow_stats_cache,
> > - (struct flow_stats __force
> > *)flow->stats[node]);
> > + (struct flow_stats __force
> > *)flow->stats[cpu]);
> > kmem_cache_free(flow_cache, flow);
> > }
> >
> > @@ -757,7 +749,7 @@ int ovs_flow_init(void)
> > BUILD_BUG_ON(sizeof(struct sw_flow_key) % sizeof(long));
> >
> > flow_cache = kmem_cache_create("sw_flow", sizeof(struct sw_flow)
> > - + (nr_node_ids
> > + + (nr_cpu_ids
> > * sizeof(struct flow_stats *)),
> > 0, 0, NULL);
> > if (flow_cache == NULL)
>
> Well, if you switch to percpu stats, better use normal
> alloc_percpu(struct flow_stats)
>
> The code was dealing with per node allocation so could not use existing
> helper.
>
> No need to keep this forever.
The problem is that the alloc_percpu uses a global spinlock and that affects
some workloads on OVS that creates lots of flows, as described in commit
9ac56358dec1a5aa7f4275a42971f55fad1f7f35 ("datapath: Per NUMA node flow
stats.").
This problem would not happen on this version as the flow allocation does not
suffer from the same scalability problem as when using alloc_percpu.
Cascardo.
