Syslinux - Jun 2003 - gigabit networks and kernel loads

I have been using pxelinux with Intel 850 1.7 Ghz P4 motherboards using
10/100 ethernet switches for about a year in a multichannel visual
generation system.  We are shifting to SuperMicro P4SAA 2.8 Ghz P4
motherboards with onboard gigabit ethernet and gigabit ethernet
switches. Put are hung up because the diskless slave nodes will not load
the (Linux) kernel from the master 2.8 Ghz computer.

Everything seems to be fine up to the kernel load.  The diskless node
does a PXE boot and requests DHCP data from master. Pxelinux.0 loads and
finds the configuration file. The menu comes up and the diskless node
waits for time out.  After time out the diskless nodes start displaying
messages about being unable to find bzImage.2.4.19 (the kernel file).
My current belief at the moment is that this has something to do with
packet handling being too slow/fast in pxelinux.

Other observations:
- Connecting a single diskless node to a master via one straight cable
works fine.  The subnode boots every time.
- Connecting a single diskless node to a master via a gigabit ethernet
switch fails almost every time.
- Connecting a single diskless node to a master via a 10/100 ethernet
switch works.
- The first boot try via gigabit switch after power off of all nodes
usually works.
- Running tcpdump on master while booting subnode via gigabit switch has
worked.
- 1.7 Ghz (10/100 onboard ethernet) via gigabit switch to gigabit master
works.
- I have had sessions where everything gigabit works fine for X>5
reboots and then quits working at all.

Mitchell James
mejames at link.com


I looked through pxelinux.asm for a wait loop that could be reduced.
But nothing popped out at me. I am running the latest syslinux version.

Mitchell E James wrote:
> I have been using pxelinux with Intel 850 1.7 Ghz P4 motherboards using
> 10/100 ethernet switches for about a year in a multichannel visual
> generation system.  We are shifting to SuperMicro P4SAA 2.8 Ghz P4
> motherboards with onboard gigabit ethernet and gigabit ethernet
> switches. Put are hung up because the diskless slave nodes will not load
> the (Linux) kernel from the master 2.8 Ghz computer.
> 
> Everything seems to be fine up to the kernel load.  The diskless node
> does a PXE boot and requests DHCP data from master. Pxelinux.0 loads and
> finds the configuration file. The menu comes up and the diskless node
> waits for time out.  After time out the diskless nodes start displaying
> messages about being unable to find bzImage.2.4.19 (the kernel file).
> My current belief at the moment is that this has something to do with
> packet handling being too slow/fast in pxelinux.
Sorry, that's most likely baloney.  TFTP is a lockstep protocol in part 
just to avoid having that kind of dependencies.
> Other observations:
> - Connecting a single diskless node to a master via one straight cable
> works fine.  The subnode boots every time.
> - Connecting a single diskless node to a master via a gigabit ethernet
> switch fails almost every time.
Sounds like there is a problem with the gigabit switch.  Perhaps your 
server has jumbo frames enabled and your switch doesn't handle them or 
doesn't have it enabled?

	-hpa