Subject: Re: IPMI experiences anyone ?
To: None <port-i386@netbsd.org>
From: Herb Peyerl <hpeyerl@beer.org>
List: port-i386
Date: 02/16/2006 19:29:52
On Feb 16, 2006, at 1:03 PM, Manuel Bouyer wrote:
> I'd like to buy some servers with real remote-mamangement support  
> (up to now,
> my x86 servers are built on high-end workstation motherboards, without
> real remote management). My ideal system would be managable from LAN
> with standard tools like telnet or ssh, and would give access to  
> the system's
> serial console, as well as power and reset switches. I'm not sure this
> exists.
>
> However most vendors now seems to support the IPMI standard. Is use  
> a custom
> UDP protocol, but maybe some boards also support telnet or ssh  
> connectivity ?
> Also, I'm not sure the console is seen as a serial port on the  
> system side
> (I've seen mails suggesting some supermicro IPMI boards would be  
> snooping
> the video memory for characters, slowing down the whole system).
>
> Has anyone experience with IPMI boards (from any vendors) ?

I don't have experience with anything you're likely to want to buy.   
I work with ATCA systems all day long and some of them have dual-xeon  
blades.  ATCA is all IPMI.  There's an I2C bus running across the  
midplane and there are shelf-management controllers that control the  
blades.  We've tried a few "PC" based blades and without fail, they  
are all 'flakey' as far as IPMI goes.

One specific example comes to mind.  You know how with a normal PC,  
if the PCI configuration screws up and the board wedges, you can't  
shut it down with the soft-power switch; so you have to hold the  
switch for 5 seconds which causes it to turn off?  Well, the PC based  
ATCA blades are really just PC's in an ATCA form-factor with an IPMC  
controller glommed onto the board, wired up to the power/reset lines  
that you're used to on a regular PC.  If the PC blade in your chassis  
goes into the aforementioned "mode", then there is no IPMI support  
for "holding the power switch for 5 seconds" and you have to  
physically yank the board.  Pulling the handle which would normally  
take the board OOS doesn't have any effect.

So, in short; I'm relatively un-impressed with IPMI as implemented on  
PC's.