Bonded head nodes
becker at scyld.com
Fri Nov 8 12:10:04 PST 2002
On Fri, 8 Nov 2002 ivan at sixfold.com wrote:
> The Multicast extension to TFTP described in the PXE standard has no
I recently wrote both a TFTP client and server. I initially did some
research on what already existed, and was surpised at high ratio of talk
vs. implementation of multicast TFTP. "Interoperating with" isn't the
same as "currently implements".
> mention of the concept of bonded ethernet channels. I am not familiar
> enough if ethernet bonding to say with certainty, but my initial guess
> would be that the ROM-driven stage of the PXE boot process would most
> definitely not support this due to the fact that a single card on a
> given client node is driven by the PXE code in the BIOS.
That's right -- most built-in PXE code isn't even able to enough to
attempt to contact a server on all interfaces simultaneously. That
means your PXE server should only talk on one channel, not the bonded
> If, however, bonded ethernet were supported in the second stage boot,
> which is driven by PXELinux or a similar bit of client code, there's
PXELinux doesn't do this.
> Mixing the concepts of multicast and ethernet channel binding seem odd
Actually, it works fine. This is a case where you really want two (or
N) switches, rather than a single partitioned switch.
Another approach is to multicast on only one of the physical interfaces,
but that can lead to unbalanced traffic
> If someone out there is working on improvements like
> this, I would like to hear about it. We are very interested in
> participating in the development of higher performance boot client
> code for Linux clusters.
We've had high performance boot code in our system since mid-2000, using
TCP instead of TFTP for performance and reliability and a light-weight
phase 2/3 image.
Our new series boot code has some interesting changes in phase2, but
"phase 1" booting is not the critical issue that it used to be.
Every current motherboard you might use in a cluster has native PXE
boot. It's not the protocol you would choose, but it's a ubiquitous
standard. It's done. It's workable. That means it won't change for
about a decade.
[[ Why do I complain about TFTP in PXE? It made sense in the days of
8KB boot ROMs, but it's a pointlessly discards reliability with 8MB BIOS
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
410 Severn Ave. Suite 210 Scyld Beowulf cluster system
Annapolis MD 21403 410-990-9993
More information about the Beowulf