Repository navigation
MPTCP and hardware offloading #4325
Unanswered
AndreaLeidi
asked this question in
Q&A
Replies: 1 comment
|
GSO and GRO shouldn't give issues in most case. Others should be disabled. This will be in next release. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I'd like to ask your opinion about using the hardware offload functions of ethernet NICs that support them:
Generic Receive Offload (GRO),
Generic Segmentation Offload (GSO),
TCP Segmentation Offload (TSO),
Large Receive Offload (LRO),
UDP Fragmentation Offload (UFO)
in an MPTCP environment.
This is because the intrinsic packet aggregation logic for maximizing single-path throughput (the context for which these functions were designed, and which covers 99% of customer configurations) would conflict with MPTCP's path manager/scheduler needs to fully manage the traffic and split it into subflows across the various interfaces.
I have three WANs, each on its own ethernet hardware and I disabled these features some time ago (Luci: Network > Interfaces > Devices > [DEVICE] - Configure > Advanced device options), resulting in a significant improvement in system stability and predictability with SQM.
It should also be considered that on embedded systems with strictly sufficient resources, disabling these hardware functions can lead to a computational burden on the CPU that could undermine the advantages obtained.
All reactions