We are trying to setup a cluster with jumbo frames in the hope that it would
speed up the NFS.
well, what kind of speed are you getting with normal frames? it seems pretty
easy to saturate gigabit with 1500B, so the main question is whether jumbo is
worth the effort.
The problem is the booting of the diskless nodes.
As the nodes boot, the kernel level IP configuration uses the default MTU
size (1500). This results in their inability to mount the root partition from
the NFS server as the server is configured to use the jumbo frames.
well, you definitely don't have to use kernel autoconf. it's quite
convenient, but the initrd approach isn't that hard, either.
you'd run dhclient, and have the privilege of configuring your
interface however you like from /sbin/dhclient-script.
here's a cute hack: if you're using pxelinux, it has an "ipappend" feature,
which just adds an argument to the kernel's commandline like this:
ip=10.49.0.66:10.49.0.254:10.49.0.254:255.255.0.0
which avoids the need to re-dhcp. after all, once you get to the stage
of downloading a kernel, you've already dhcp'ed once.
I haven't had time to try this...
regards, mark hahn.
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
