How to make bonding on two independent switches more correctly?

Question

linux

Good afternoon. There is a proxmox 4.4 cluster installation on three identical servers. On their own base, ceph was raised to return rbd. Each server has 4 gigabit cards. Two links from the server are combined into lacp and on the side of the managed simple tp-link l2 switches are also assembled into lacp, a bridge is raised on top of each for communication with the local network and virtual machines. Ceph is served by two links in balance-rr, each of which is connected to a separate unmanaged switch (there are two of them, only ceph serves, physically separated from the world) for fault tolerance and bandwidth, tuning is minimal (net.ipv4.tcp_reordering = 127). The performance is tolerable, there is no money to switch to 10 gbit. Actually the question is, did you do the right thing with the network for ceph? Or to arrive as well as with a client network? There is a second l2 managed switch,

Answer 1

Mikhail Of course, it's better to find a switch that can LAG which is 802.3ad.
https://en.wikipedia.org/wiki/%D0%90%D0%B3%D1%80%D...
I would really look towards balance-xor .
You will have a transfer rate to each client at the level of the PCI bus 133 approximately :)
Do these things only within the framework of one switch, unless the switches are ironically stacked among themselves.

Answer 2

what to do as you want switches must be able to MLAG

Answer 3

Correctly. Within the same switch, balance-rr can be run through different VLANs

Answer 4

Thanks to the unsubscribers, they do not know how to manage mLAG. If we leave the scheme with two stupid switches, does it make sense to replace balance-rr with balance-alb or with balance-xor? Packet loss on rr is a bit of a concern