SQM makes bufferbloat worse

According to @gwlim you get better performance with fast path enabled, but I am not sure if it works with options like:
option iqdisc_opts 'nat dual-dsthost ingress'
option eqdisc_opts 'nat dual-srchost'

Couldn't really find an answer regarding flow offloading. Some people say yes and others no.

I'm still waiting for my fiber connection to do some real world test.

Software flow offload is supposed to work with qdisc, but hardware doesn't.

Although "working" in this case means, no real speed-up to be expected (as all SQM traffic can't be offloaded by definition, given that SQM needs to micro-manage each individual packet - but it should still work transparently, just not any (at least not a lot) faster).

1 Like

My understanding was that software flow offload mostly avoids a lot of overhead in iptables processing (like filter,postrouting,etc tables won't be run for the packet), so that you can get some speedups. But I don't really know.