Yevgeny Kliteynik a écrit :
Hi Nicolas,
Nicolas Morey Chaisemartin wrote:
Hello,
I am conducting some tests on routing non-pure fat-tree network using
the fat tree algorithm of OpenSM.
The network I am experimenting on is a 3 level fat tree, with a
pruned 3rd layer.
By providing the root_guid_file, the algorithm works great !
The problem is, we would like to add some service nodes directly on
the 3rd level switches.
I have added the cn_guid_file so the network is still recognize as a
fat tree.
OpenSM once more manage to create the routing for the network. It
provides full connectivity,
except there are no routes between non computes nodes.
I understand that the point of setting these node as not compute node
should intend they won't talk to each other, but we still need a bit
of connectivity between them to exchange few datas (pings and such).
A simple min-hop or such should be enough to generate those routes.
It will probably desequilibrate the number of routes going through
the top links, but those additional link makes virtually no traffic
at all, so in practical it shouldn't be a problem.
Fat-tree should create full connectivity as long as there is an up/down
route between ports. Do you get connectivity between these nodes with
up/down routing algorithm?
Try running it with the same root_guid_file.
-- Yevgeny
Well the route would be more down/up compared to the rest of the transfer.
(Im not sure I was clear, but when i talk of 3rd level, I mean top
level. 1st level begin the switches just above the compute nodes)
I'll try this tomorrow
Thanks
Nicolas
Is there any reasons such a behaviour wasn't implemented yet? Should
there be one?
Regards
Nicolas Morey-Chaisemartin
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit
http://openib.org/mailman/listinfo/openib-general
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
