Dear experts,

we are operating a compute cluster with a number of storage systems 
serving data via NFS via LACP-bonded 10 Gbit/s links (usually 2 or 4).

 From time to time, users may overload a server and use up all available 
bandwidth for hours and sometimes days at a time and I am currently 
exploring if we could use tools like nftlb[1] to achieve the following:

(a) For (NFS|any) traffic have at least two tiers (farms?), one are 
standard compute nodes and one are interactive log-in nodes/web servers. 
The former should have a much lower bandwidth priority as the latter. 
These targets can easily be distinguished by IP ranges/netmasks.

(b) Sometimes we need to perform bandwidth intensive/sensitive 
operations, e.g. just today I would like to move a user's file system 
from one too busy box to another but sending the ZFS snapshot over via 
mbuffer takes really long and I would like to prioritize this connection 
as well, i.e. I know the source and destination IP as well as target TCP 
port.

I am still learning to migrate from iptables to nft while I stumbled 
over nftlb which looks to support what I want, but I am not sure yet - 
and I have not found many documents describing potential set-ups and 
thus I wanted to ask the experts here first.

Base of all our systems is currently Debian 10 (buster) and thus kernel 
4.19.

Cheers and thanks a lot in advance for any insights/pointers/...!

Carsten

[1] https://github.com/zevenet/nftlb
-- 
Dr. Carsten Aulbert, Max Planck Institute for Gravitational Physics,
Callinstraße 38, 30167 Hannover, Germany, Phone +49 511 762 17185