Niagara Support and Ubuntu

[Colm MacCarthaigh]

Hey there, 

This is an ineffective mail, in that the important stuff is at the end,
but it was hard to structure differently, still it may well be worth
skipping right to the last two paragraphs if you don't want to read it
all.

I've been running Ubuntu on SUN's (sparc) Niagara chip for a few weeks
now, ever since it was first made work (we installed a kernel from git -
with help from fabbione!), and it's been going great. We actually have
the machine in-production, as murphy.redbrick.dcu.ie - an active shell
server, serving a few thousand users, running Ubuntu. Fabbione and Dave
Miller have both done a great job, and it's working great.

I've blogged about it a little here;

	http://www.stdlib.net/~colmmacc/category/niagara/

and I've got some libmicro comparisons here;

	http://www.stdlib.net/~colmmacc/libmicro-benchmarks/
	http://www.stdlib.net/~colmmacc/mv-res.html

which shows the relevant return speeds of various system calls on Linux
Vs Nevada (Solaris express betas). It's important to note that Linux has
epoll() which is more performant than Solaris' poll() - but is missing
from the comparison as libmicro does not benchmark it.

Anyhow, I'm also a very (super) lazy Debian package maintainer, wrote my
own in-house Debian installer, and am a committer to the Apache httpd
and apr projects. I'm a big-time Ubuntu user too, and just wanted to
give some background. I was at JavaOne this week, just as an attendee
and behind the scenes a little bit on some other stuff.

Firstly, thanks to my blog, I'm getting a lot of mails from people who
are installing and trying to install Ubuntu on Niagara boxes. A lot of
these are relatively benign questions from people who've never used a
serial connection yet and so on. But I am getting a fair amount of mail
from people who can't use Ubuntu on the T1000. The underlying problem is
relatively clear - the tigon driver isn't working for their broadcom's
(I've suggested trying the other bcm driver, but heard no reports yet).
I dunno if its a problem generic to sparc (ie an dump word-size or
endianess issue, a chipset issue on the particular box or whatever).

I havn't fired up my own T1000 just yet, but once I do, I'll try and so
some more useful and intensive debugging to provide more useful feedback
about that to the kernel channels.

Anyhow, as June rapidly approaches, and as could be judged from the
keynote and the questions I was being asked (including from many Sun
people), Sun are keen that the Niagara platform be supported by the next
Ubuntu release.

Hell, I'm keen too, I think Linux runs great an Niagara, despite the
freshness of the port, I'm not sure 3 months is really long enough of a
burn-in for the kind of messy race-conditions these architectures can
throw up, but I'm more than comfortable running Ubuntu on Niagara in
production.

But I think that as the T1000 is the most common Niagara box right now,
and as it isn't even installable their just-yet (no cd-drom drive), it
might be worth taking that into consideration as to whether the platform
is "officially" supported (to the extent that that even means anything).

Either way; for the sake of my inbox, if there are release notes - I beg
you to include this in the known issues :-)

As a Debian house (it runs on approx 80% of our boxes) which was already
in the process of migrating to Ubuntu, I have to say the bi-directional
Sun/Canonical|Ubuntu support thing will be amazingly great for me. Even
adding Sparc as an official architecture means a great deal to us.  

Some other feedback I wanted to give is this; the installer is
effectively unusable over a 9600 baud serial console, or certainly the
incredibly laggy-to-redraw one that the Niagara boxes ALOM comes with.
It's just about doable if everything goes o.k., but if you need to
change anything - or god help you, manually lay out the partitions or
install with a software raid, it basically wastes a day. I spent 8 hours
installing Ubuntu the second time, and I *really* know what I'm doing.

There *needs* to be some easier way of escaping to a shell for non-vty
users, and non-curses terminal-dialog like way of answering the
questions would be an incredible assistance. A reverse-telnet install
target is another possibility, but that might confuse people. What I've
been doing to get around it is to use my own installer, but that's not a
very scalable option (the installer is a 60-line shell script which just
partitions, mkfs's, mounts, and then does wget | tar -x , nice).

There are some other issues around how SILO works (SILO is generally
terrible, but work on porting grub is underway). But I think Fabbione
has done another good job of fixing those up too.

Anyway, I wanted to provide this feedback for the real point of my mail;
based on what I was hearing at JavaOne, SUN are keen to make a big deal
of Ubuntu running no Niagara. I think that's great, I think Ubuntu runs
brilliantly on Niagara, I'd recommend it to anyone.

However I also think that that kind of PR might create expectations in
the user community, and if users are expected Ubuntu's excellence in
terms of ease-of-use on Niagara, it's not quite going to be there.
There will be frustrated users, and it will generally suck. On a PR
level, it may well rebound on Ubuntu, and that would definitely suck,
because I want to see it do really well there. It makes me look less
insane when I tell the boss that's what we're going to be running on
ours ;)

I think about 90% of the problem is the installer Vs the ALOM's serial
console and the T1000 broadcom issue, but I'd still be a bit wary about
the relative newness of some of the other code. But I just wanted to
chip in and make sure that I'd communicated some of our experiences.

--
Colm MacCárthaigh                        Public Key: colm+pgp at stdlib.net