100G is boring, so I went 200G instead on the rx side. Implemented zc send in the test app (kperf) as well (io_uring already supports this). Single CPU thread, maxes out 200G link as well. See below for basic details:
https://lore.kernel.org/io-uring/9bbab76f-70db-48ef-9dcc-7fedd75582cb@kernel.dk/
https://fosstodon.org/@axboe/113278217854097854
=> More informations about this toot | More toots from axboe@fosstodon.org
@axboe when you say 200G it's not gbps because in the link says 96gbps. what's 200G and are you trying to just make cpu 100 percent? sorry for my ignorance
=> More informations about this toot | More toots from thejvmbender@techhub.social
@thejvmbender Not sure what link you're referring to, but it's 98.4 and 98.6, so 197Gbps in total. Where's the 96 from?
=> More informations about this toot | More toots from axboe@fosstodon.org
@axboe link in the first post, i read as 98 and 96 but it was actually 98.6 🤓 I don't have my glasses sorry, so 200g means gbps 😇 thank you
=> More informations about this toot | More toots from thejvmbender@techhub.social
@axboe If I read this correctly that's TCP streams, right? Now let's do UDP in UDP (vxlan) ... runs :)
=> More informations about this toot | More toots from theuni@social.tchncs.de
@theuni yep it's TCP, over IPv6 to be exact. And no, I think I'm done testing for now, just wanted to see what the limits were at. Didn't find it, but I guess that's fine as I don't have faster nics (or a switch) than this anyway. "Good Enough" for now.
=> More informations about this toot | More toots from axboe@fosstodon.org
@axboe Yeah, just kidding. UDP in UDP measurements when we introduced evpn/vxlan was nasty and required a bit of tuning even on 10g links ...
=> More informations about this toot | More toots from theuni@social.tchncs.de This content has been proxied by September (ba2dc).Proxy Information
text/gemini