As someone who has used IPMP extensively in Sun's internal datacenters and seen 
some interesting failure modes, I would like to throw out one big suggestion:

Please give sysadmins the ability to specify probe targets explicitly in
/etc/default/mpathd (or some other appropriate config file).


We have run into cases where allowing lots of servers' in.mpathd to bang away 
at 
the default router constantly is a bad idea.  Some routers/firewalls 
de-prioritize responding to ICMP when their CPU load is above some threshold. 
This results in every server on the subnet suddenly thinking that both IPMP 
interfaces have failed simultaneously.

If you have other interfaces (like some sort of back-end network for NAS or 
management purposes) that you want to run IPMP on (other than the one on which 
the default route resides), IPMP sends multicast pings to find other hosts to 
probe.  This is also problematic, particularly in hardened environments where 
ip_respond_to_echo_multicast has been disabled (JASS does this).


In both cases, we would like to be able explicitly control who in.mpathd pings. 
  The method of inserting static host routes in the route table seems like a 
kludge at best, and if you list less than 5 per subnet in.mpathd still falls 
back to its other methods anyway.  Therefore, ability to explicitly list probe 
targets in /etc/default/mpathd would be really nice.

-- 

Chuck Cox
Sun Microsystems
Global Internet Services
(303)223-6099 or x69526
Chuck.Cox at Sun.COM

Reply via email to