On 2020-08-02, Ivo Chutkin <[email protected]> wrote: > On 1.8.2020 г. 14:43 ч., Stuart Henderson wrote: >> On 2020-07-31, Ivo Chutkin <[email protected]> wrote: >>> Hello guys, >>> >>> Thanks for suggestions. >>> >>> Tweacking sysctl >>> >>> net.inet.udp.recvspace=131072 >>> net.inet.udp.sendspace=131072 >>> >>> solved the problem. >>> >>> Test between routers that started to drop packets over 10Mbit, now run >>> test at 100Mbit with less than 1% drops (over 50% before). >> >> net.inet.udp.recvspace and net.inet.udp.sendspace only affect traffic >> generated on a machine itself, they do not affect forwarded traffic. >> >> Generating packets on the router itself isn't a good test for how well >> they can forward packets. >> >> > > Hi Stuart, > > I have read it many times that increasing the values do not affect > forwarding performance. That is is why I was skeptical about calomel.org > solution.
The suggestions on that website are a combination of things that may possibly have been slightly relevant some years ago (and which may or may not still be relevant), and things which are totally useless. Combined with the lack of information about what anything does (if the author even knows in the first place) it's a poor information source. > Can you point some directions for increasing forwarding performance? hmm. If you use PF, make sure your state limits are high enough. There isn't really much else I'd consider tweaking "just in case it helps", there are some tools that might help figure out why things aren't working well enough in specific circumstances, but most need some prior knowledge to figure out what to do with what they're showing. (things like netstat -m, systat vmstat, systat mbuf, even just dmesg and top sometimes give clues). > I upgraded all routers to 6.7 with patches and I think problem > disappears, using the same hardware. I don't remember what changed when, but I think there may have been some changes to how "livelock" was detected in that timeframe which may have helped (the kernel tries to backoff on network processing if it thinks it can't cope with the traffic).. > > Thanks for your help, > Ivo > >
