10 ms·
> The top shows that single-core performance is the bottleneck, and all CPU is consumed by ksoftirqd. Trying out VPP instead of Linux kernel networking might m
by ivan4th 3y ago
> The top shows that single-core performance is the bottleneck, and all CPU is consumed by ksoftirqd.
Trying out VPP instead of Linux kernel networking might make some sense: https://fd.io https://fd.io
It will likely reach the target performance, although the NIC requirements need to be rechecked.
But... well, this software is not for the faint of heart ;)
- tgtweak 3y agoYou can/should also use hardware offloaded queues wherever possible, or look into kernel-bypass networking.
- civiloid 3y agoWell... I did more than I've described in the post (I'm the article's author). It just became too large, so I've decided to stop and publish a second part once I run more tests and fix some problems. So far, I have yet to figure out all the problems and settings with Netronome (or give up on it). Still, it is Netronome-specific as I can easily get line rate over TCP single stream with Mellanox cards, and the CPU doesn't even break a sweat (however, I have relatively simple routing rules). I want to try out programmable NIC, and as far as I understand, Netronome has some support there; I just wanted to get a simple setup running first and then optimise/improve it. I tend to perfect the setup, and if I let myself do more and more and more - I will always need more time to finish it.