Redistribute traffic over bond

[ph0x]

Hey there!
I have three nodes in a cluster with Ceph connected to a switch with two NICs each in 802.3ad bond mode.
I noticed that one node hardly uses the second NIC and with most of the traffic going through one NIC the available bandwidth for cluster operations is more limited than with the other two nodes.
The reason is, I guess, that the hash algorithm calculates that NIC to be used for both of the other nodes.
Is there an easy way to tweak this behavior and have the hash algorithm distribute the NIC usage more evenly or do I have to play around with the hash algorithms until I find a suitable one?

Thanks for any suggestions on that topic!

Regards
Marco

 

[spirit]

best way is to use layer3+4 hash algorithm on the hypervisor, for traffic going out the hypervisor,

and the same on the switch, for traffic going out the switch (so coming to your hypervisor). On switch, it can be called something like : src-dest-srcport-desport algorithm.

as ceph use multiple daemon-ports by osd, it'll loadbalance fine .

 

[ph0x]

I will give it a try, then. Thanks!

 

[Magneto]

~

best way is to use layer3+4 hash algorithm on the hypervisor, for traffic going out the hypervisor,

and the same on the switch, for traffic going out the switch (so coming to your hypervisor). On switch, it can be called something like : src-dest-srcport-desport algorithm.

as ceph use multiple daemon-ports by osd, it'll loadbalance fine .

Please explain this?