Lustre discuss - May 2006 - Using lctl for "bonding", recovery questions

Hello,



I''d like to ask for some insight on lustre networking. We have the=20
folowing setup:

MDS: Ethernet only
OSTs: Myrinet (TCP) and Ethernet
A-Clients: Ethernet only
B-Clients: Myrinet (TCP) and Ethernet


I configured two addresses for the OSTs and B-clients. lconf however=20
configures lustre in such a way that the "Myrinet"-connection is
not=20
used and everything goes over ethernet. What we do is to call lctl=20
after lustre start and issue a

connect <myrinet-address-of-server> 988

for every OST - then both connections are used in parallel


Now my question is whether and how lustre recovers the network when eg.=20
an OST is restarted - my impression is that something works differently=20
(that means connections are not restored) by as there are many timeout=20
values in the recovery processes I would like to see a short high-level=20
overview when lustre will reconnect to a particular server (=20
"--failover" is configured for the OSTs as I assume this is
necessary=20
to enable transparent recovery after an OST reboot, right?)


Thanks!

  - Christoph



--=20
Leibniz Rechenzentrum M=FCnchen (LRZ)
http://www.lrz.de
Abteilung Hochleistungssysteme
Barer Str. 21 - 80333 M=FCnchen - Germany
Tel. ++49-(0)89 / 289-28853, Raum S1527