Re: [xcat-user] [External] IBM IPMI USB interface

[Stuart B. <st...@4g...>]

some ranting below...I doubt there is much support for many of the
things below...

On Thu, 31 May 2018 at 08:51 -0000, Jarrod Johnson wrote:

> On preferring KCS not to be there, is this because it is extraneous or
> because it leaves root-level user access to reconfigure networking and
> credentials for the BMC?  Would that sort of thing be less worrisome
> when configured to only allow read access once in an OS (enough for
> 'sensors', 'fru', and 'sel' sorts of commands to work, but not for 'lan
> set' and such.

I just prefer using the network interface when running lots of nodes.

I don't really object to the KCS interface that much and am not
worried about root access to the device.  It seems a little awkward
and possibly a standard I2C interface would be less awkward (speaking
as someone who hasn't really studied KCS or I2C).

I would also like to get rid of other extraneous hardware (and
software) on the compute components of a large system: VGA/graphics,
audio devices and related connectors seem wasteful.  Disk and RAID
controllers (along with the related sheet metal) seem wasteful for
diskless compute nodes.  CPU, RAM, network, power and cooling are the
primary needs.

Video framebuffer support over the IMM just continues to support the
expectation that graphical output is suitable for system interaction.
I was hoping UEFI would have encouraged people to return to loggable
interaction with the BIOS.  I have this same problem with people who
emulate a display device in most virtual machines.

> To sum up my current take on what the DMTF guys are pushing for and
> the compromise I would propose:

The DMTF has done some good work and IPMI is very useful (if somewhat
awkward).  xCAT has good interfaces to most of what I find interesting
about IPMI (although I do use ipmitool itself on occasion).

The SOL interface with emulated and/or back-to-back serial ports seems
awkward and a better character oriented interface might clean up some
usability issues.  The SOL network protocol also seems pretty awkward
and could probably use some transition to a tcp based protocol.

The "shared" mode 1G interface also is awkward for various reasons
(and I get the impression Intel is making it even worse with their
management engine).  For newer equipment we are going with dedicated
IPMI network interfaces and 10G (maybe 25G, IB and/or OPA) for the
node networking.

Some standardization of firmware loading, BMC/BIOS configuration
variable setting and reading and LED sensors (rvitals leds) might be
useful.

Oh yeah, another complaint for the DMTF folks: the system clock should
be in UTC, not "local time".

Stuart Barkley
-- 
I've never been lost; I was once bewildered for three days, but never lost!
                                        --  Daniel Boone