Re: [PATCH RFC net-next] netif_receive_skb performance

[Alexei Starovoitov <ast@plumgrid.com>]

On 4/28/15 10:23 PM, Eric Dumazet wrote:
> On Tue, 2015-04-28 at 19:11 -0700, Alexei Starovoitov wrote:
>> Hi,
>>
>> there were many requests for performance numbers in the past, but not
>> everyone has access to 10/40G nics and we need a common way to talk
>> about RX path performance without overhead of driver RX. That's
>> especially important when making changes to netif_receive_skb.
>
> Well, in real life, having to fetch RX descriptor and packet headers are
> the main cost, and skb->users == 1.

yes. you're describing the main cost of overall RX including drivers.
This pktgen rx mode is aiming to benchmark RX _after_ drivers.
I'm assuming driver vendors equally care a lot about performance of
their bits.

> So its nice trying to optimize netif_receive_skb(), but make sure you
> have something that can really exercise same code flows/stalls,
> otherwise you'll be tempted by wrong optimizations.
>
> I would for example use a ring buffer, so that each skb you provide to
> netif_receive_skb() has cold cache lines (at least skb->head if you want
> to mimic build_skb() or napi_get_frags()/napi_reuse_skb() behavior)

agree as well, but cache cold benchmarking is not a substitute for
cache hot. Both are valuable and numbers from both shouldn't be blindly
used to make decisions.
This pktgen rx mode simulates copybreak and/or small packets when
skb->data/head/... pointers and packet data itself is cache hot,
since driver's copybreak logic just touched it.
The ring-buffer approach with cold skbs is useful as well, but it will
benchmark different codepath through netif_receive_skb.
I think at the end we need both. This patch tackles simple case first.

> Also, this model of flooding one cpu (no irqs, no context switch) mask
> latencies caused by code size, since icache is fully populated, with a
> very specialized working set.
>
> If we want to pursue this model (like user space (DPDK and alike
> frameworks)), we might have to design a very different model than the
> IRQ driven one, by dedicating one or multiple cpu threads to run
> networking code with no state transition.

well, that's very different discussion. I would like to see this type
of model implemented in kernel, where we can dedicate a core for
network only processing. Though I think irq+napi are good enough for
doing batch processing of a lot of packets. My numbers show
that netif_receive_skb+ingress_qdisc+cls/act can do tens of millions
packet per second. imo that's a great base. We need skb alloc/free
and driver RX path to catch up. TX is already in good shape. Then
overall we'll have very capable packet processing machine from
one physical interface into another.