> > Check out http://cesdis.gsfc.nasa.gov/linux/beowulf/beowulf.html it's the
> > Beowulf linux cluster project they have a patch that will allow multible nic
> > act as one through bonding (Just like two b-channels in ISDN
This is missing the point. Simple load balancing is not a solution
because there needs to be some inteligence. I think that the
ifenslave is the best place to start since I would, in a sense, just
want to modify the logic it uses in sending, but it would involve the
kernel knwoing which interface a given MAC address lay on. Although
the structures seem to be there for this, the arp layer doesn't seem
to handle this properly.
Alan was closer in his last message. It is in a sense routing on MAC
address instead of IP. Perhaps an ASCII diagram is in order:
+----+ pc1 pc2
/----------\ | | | |
| | | + + + + +
| Server | | [switch one]---+
| | | |
| eth0-----+ |
| | | Switch Interconnect
| eth1----------+ pc3 pc4 |
| | | | | |
| eth2-----+ + + + + + |
| | | [switch two]---+
\---------/ | |
| |
+----+ pc5 pc6 |
| | | |
+ + + + + |
[switch three]--+
The switches themselves to the briding and in a sense "routing" between
themselves. That's what a switch does, after all...
If these were separate IP networks, it would be done with subnetting and
you could setup simple IP rules. I want a solution that doesn't involve
subnetting or special routes, etc etc. I want the server to have only one
IP address, but appear to be a localhost on each of the switches..
Another less costly solution would be to replace each switch by a hub, and
use a single switch for the "switch interconnect". Now the benefit really
should show itself.
By putting the heavily used server on all three switches/hubs/etc you
significantly reduce the traffic both along the backplane and traffic from
any client to the server doesn't have to travel through anything but the
one switch.
I hope this clarifies what I'm looking to do. I thank you all for your
continued careful comment.
-Mark
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu