[Beowulf] KVM to a compute node - ssh

Tim Small tim at buttersideup.com
Wed Jun 30 08:15:18 PDT 2004


Sorry to jump into an old thread, but I turned this post up with google,
and thought my experiences could be useful.

>Do also consider using IPMI for out-of-band management. Here is a small
>HOWTO:
>
> 2. Configure IP-networking for the BMC (Baseboard Management Controller)
>with the bootable resource CD.
>  
>
You can do this on the box under Linux using the ipmitool program, in
conjunction with the OpenIPMI driver (which is included the Linux
kernel, but I found I needed the newer version for stable operation with
2.4.26 on the SE7501WV2 boards, and the lmsensors "bmcsensors" module).

http://ipmitool.sourceforge.net/
http://openipmi.sourceforge.net/

I haven't tried freeipmi.

> 3. Download the CLI-RPM from Intel and install it on your Linux
>workstation. (Or use the Windows software.)
> 4. Run "dpcproxy" (from the CLI-RPM) in the background (and configure
>some kind of tool to restart it since it is quite unstable).
> 5. Connect to your server with "dpccli" (from the CLI-RPM). Try these
>commands:
>    * power-on
>    * power-off
>    * reset
>    * sensors
>    * console
>  
>
ipmitool also does this, and has (IMHO) a nicer interface, but you still
have to use the Intel (freely downloadable) dpcproxy / dpcli programs to
get interactive SOL support.

> 6. With the console command you can enter BIOS, view Grub and the console
>of the operating system.
>
>No need for extra serial cables! (I know that they aren't needed for
>WOL/network boot either.) Depending on your need for security you might
>want the console traffic on a separate network.
>  
>
Yup, IPMI with SOL expects to have hardware flow control tho', so you
can get some dropped characters without this, but it's certainly
useable, and good enough for me at the moment.  Enabling RTS/CTS with
current kernels, is not advisable still as it handles disconnected
terminals badly - i.e. it blocks the kernel for an excessive amount of
time -
http://www.tldp.org/HOWTO/Remote-Serial-Console-HOWTO/bugs-kernelp.html

There is a patch floating around however, but I haven't tried it yet (if
anyone has a 2.4.26 version, that would be helpful!).

>I have found some problems with this, for example I cannot get ENTER-key
>to work inside BIOS, which quite limits the use for entering BIOS. Only
>  
>
This is odd, the ENTER key seems to work if you telnet into the dpcproxy
program (port 623), but not using the dpccli program..

>functions keys from F1 to F10 seem to work and not F11 and F12, and F12 is
>the one used to change boot-device... (It works changing boot-device from
>the Windows-client with help of the service partition, reports a collegue
>of mine.)
>  
>
This seems to be an odd feature of some serial BIOSes (e.g. AMI) - they
don't support the VT100 codes for F11, F12, but instead you can use "Esc
1" to "Esc 0" for F1-F10, "Esc !" for F11, and "Esc @" for F12 (totally
unintuitive with my non-US keyboard with the @ symbol not on the "2"
key, but there you go...).

>I am currently using console=/dev/ttyS0 to redirect the Linux kernel
>console to the virtual serial port over ethernet which works ok, you can
>login as root and all. But it makes the normal KVM-output depressingly
>black. But as far as I know, it should be possible to also use the normal
>VGA for console (it works with Grub) as long as the kernel does not mess
>upp it's video mode.
>
>  
>
If you pass something like " console=tty0 console=ttyS1,19200n8" to the
kernel, you will get kernel messages on the VGA console as well
(although init script output goes to the serial console only).

>I hope that updated software from Intel or maybe one of the open source
>projects will fix some of these problems - I havn't had the time to look
>into them yet. It currently works good enough for me.
>
>  
>
Mmm, Intel's SOL implementation is proprietory for these boards (SOL is
not part of IPMI v1.5) it may be however that it is close enough to the
spec published in IPMI v2.0 that SOL support in free utilities will work
with it, with, or without a bit of reverse engineering..  The most
annoying ommision IMHO is the inability to send a break - so no remote
sysrq functionality yet.

Tim.




More information about the Beowulf mailing list