Re: Impact of hardware offloads on network stack performance

[craigt <[e-mail]>]

The BBC folk involved in that work aren't on this list. If there are
questions about the implementation you would like answered, get in touch
with me and I'll try to get some time.

c

On 10 May 2018 at 16:02, Mikkel Fahnøe Jørgensen <[e-mail]> wrote:

> Some of you probably have inside knowledge of this:
>
> https://www.bbc.co.uk/rd/blog/2018-04-high-speed-networking-
> open-source-kernel-bypass
>
>
> On 10 May 2018 at 16.57.35, Mikkel Fahnøe Jørgensen ([e-mail])
> wrote:
>
> Netmap is has a fairly low line count and complexity for what it achieves,
> it is not tied to any vendor specific hardware, it patches do seem to
> arrive for various cards.
>
> I recently did a user space emulation of netmap where separate user
> threads drains and fills the tx and rx rings respectively. I’m not sure
> this is worth it in terms of performance since it adds more synchronization
> (netmap makes certain assumptions because some updates are in the kernel).
> However, it does provide a uniform buffering layer and it makes it easier
> to plug in proper netmap when available. It does move some per message
> latency out of the way.
>
>
> On 10 May 2018 at 16.03.28, Eggert, Lars ([e-mail]) wrote:
>
> On 2018-5-10, at 14:57, Boris Pismenny <[e-mail]> wrote:
> > Also, is there a reason to prefer netmap over other userspace network
> stacks like DPDK/VPP/etc?
>
> For me personally, the netmap API is a lot more minimal and therefore
> easier to integrate. DPDK really wants to take over scheduling, memory
> management and all kinds of other stuff besides packet I/O. YMMV.
>
> Lars
>
>