The FNN (Flat Neighborhood Network) paradox
Georgia Southern Beowulf Cluster Project
gscluster at hotmail.com
Thu Feb 22 15:35:44 PST 2001
This is probably a pipe dream of pipe dreams, but here is a "possible" hack
to get advanced routing to work in an FNN machine. The docs on KLAT2 say
that AFN hardware is to be added soon. Give the speed of the AFN hardware,
it may be possible to impliment the following scenerio. I only make this
suggestion a non-programmer and with a little consideration that AFN is
orders of magnitude faster than an ethernet message.
Many machines are connected via Flat Neighborhood Network (FNN) and
Aggregate Function Network (AFN) topologies. The FNN is used for large data
and the AFN can be used for small data.
1. Machine 1 (M1) realizes it wants to send a message to Machine 2 (M2) and
that it has more than one route to M2 using the FNN.
2. M1 sends a short message over the AFN to M2 telling it that it wishes to
communicate over these several channels at once.
3. M2 loads the channel-bonding module (all NICS are on different switches)
and sends a reply over the AFN to M1 that it is ready for transmission.
4. M1 loads the channel-bonding module (possibly after it sends its first
AFN message and while waiting for M2).
5. M1 and M2 are connected across multiple channels and can communicate
using all available bandwith. FNN communication with other machines is now
not possible since M1 and M2 are channel-bonded.
6. M1 and M2 unload the channel-bonding module and can now communicate with
other machines through AFN and FNN.
Okay, okay, I can name any number of drawbacks to the above scenerio, but it
may be useful for finding other means of advanced routing. Also, I have no
experience with FNN, AFN, and channel-bonding, so please don't flame me.
This is mostly a thought, and I hope someone can try it to let me know if
Get your FREE download of MSN Explorer at http://explorer.msn.com
More information about the Beowulf