LWN: Comments on "Accelerating networking with AF_XDP"

Accelerating networking with AF_XDP

dankamongmen — Sun, 22 Jan 2023 17:15:43 +0000

irq load will be a function of how often NAPI posts from its loop, which can be controlled with sysctls.

Accelerating networking with AF_XDP

f18m — Tue, 05 May 2020 14:38:26 +0000

Hi Jonathan,
Thanks for this very interesting article.
One question though: is AF_XDP still hindered by interrupts? I mean in high-performance applications the "ksoftirqd" thread will jump up to 100% of CPU usage in scenarios where you're receiving a lot of packets per second... on 100Gbps the theoretical max is 148 MPPS... unless you use DPDK framework, which does use polling instead of interrupts, you will never achieve that PPS rates on Linux. Is AF_XDP using a polling mechanism or relies on interrupts?

Thanks!

related writeup on getting packets to user-space

wingo — Thu, 19 Apr 2018 08:23:59 +0000

Nice article. I wrote up some notes on Magnus Karlsson's FOSDEM talk on AF_XDP here (scroll down to the second talk):

https://wingolog.org/archives/2018/02/05/notes-from-the-f...

I thought the following talk by François-Frédéric Ozog was an interesting counterpoint. In any case, if the kernel can get packets to userspace in a fast, generic way, that would definitely be a step forward.

Accelerating networking with AF_XDP

RamiRosen — Wed, 11 Apr 2018 20:59:26 +0000

Regarding your comment about appealing to DPDK users: actually it seems that you are right, and a patchset for PMD driver for
AF_XDP was sent to the dpdk-dev mailing list about a month ago. See:

http://dpdk.org/ml/archives/dev/2018-February/091502.html

Rami Rosen

Hardware support

sbates — Wed, 11 Apr 2018 12:02:11 +0000

Thanks! I also found this a rather awesome resource.

https://qmonnet.github.io/whirl-offload/2016/09/01/dive-i...

And yes, from what I can see, only Netronome offer kernel support for HW offload of eBPF programs today (but I'd love to be corrected if this is not true).

Accelerating networking with AF_XDP

mjthayer — Wed, 11 Apr 2018 07:14:45 +0000

Seems to be partly answered in the next two comments.

Accelerating networking with AF_XDP

blitzkrieg3 — Tue, 10 Apr 2018 21:25:08 +0000

Looks like it according to page here:

https://www.solarflare.com/ultra-low-latency

Accelerating networking with AF_XDP

blitzkrieg3 — Tue, 10 Apr 2018 17:35:08 +0000

Will stuff like libonload eventually use this? Or are they to remain separate forever?

Hardware support

corbet — Tue, 10 Apr 2018 12:52:59 +0000

Much of the available information seems to be in the form of slidware, but Netronome has a card that does BPF offload. See this netdev talk [PDF] or this FOSDEM talk [PDF].

Accelerating networking with AF_XDP

sbates — Tue, 10 Apr 2018 12:41:03 +0000

Nice overview. I guess this is intended to appeal to users of DPDK and the like.

“and, for some hardware, to offload the BPF program into the interface itself.”

Does anyone have a link to which devices support this and any patches needed to enable this. Offloading arbitrary BPF programs to the NIC sounds very interesting given the path the kernel seems to be taking.

Accelerating networking with AF_XDP

mjthayer — Tue, 10 Apr 2018 06:06:48 +0000

I wonder (as someone who does not know enough about networking) whether this will be usable in combination of in-device processing (e.g. a card's hardware TCP/IP stack) as well in future. I seem to recall that Linux preferred to do as much of that processing on the CPU as possible until recently, but I don't know whether that is changing or likely to change - I would expect that using specialised network processing hardware has at least the potential to be a performance benefit. I am sure that the answers will enlighten me...