From: Paul Jakma (paulj at domain itg.ie)
Date: Sat 19 May 2001 - 03:22:53 IST
On Fri, 18 May 2001, Smelly Pooh wrote:
> Plan9 is not a distributed OS, if you had a network of plan9
> machines, they can all be seen as distinct instances of Plan9
> (each machine running its own services with its own address and so
> on). Plan9 doesn't do any more networking/distribution than Linux
> does, but it does make it more seamless. Everything, including
> remote resources appear to Plan9 as files.
hhmm.. yes. Al Viro seems very taken with plan9, as does Linus..
From discussions on l-k I've a feeling 2.6 is going to be rather
different. Eg, Al wants to write an fs for drivers to plug into - in
effect giving each driver it's own fs for info/control purposes.
linus wants to have things like /dev/tts/0,baud=9600,foo=bar.
ie, in effect try to banish ioctl's.
> That's not just NFS type files but everything from acessing a
> remote Plan9 Windows server as a device file to messing around
i'd love to see plan9 alright..
> I don't see how that will make your kernel/kernels any more efficient, you're
> basically applying a distributed computing model to parallel computing which
> involves more work
well, i've probably taken lm's idea, got it completely arseways and
paraphrased it even more backwards.. but the guy does a have clue, as
he was involved with scaling work at both SGI and Sun.²
i think the idea in essence is to /contain/ resource mgt to something
like an fs and running multiple less scaleable but less completed
kernels, rather than making a kernel more and more fine-grained. The
latter approach lm feels is the wrong approach.
ie contain the complexity.
but best search for "McVoy smpfs" on google. threads were maybe a year
or two ago on l-k, and he links to 2 papers of his on bitmover.com
where he details it a bit more.¹
As kenn (and lmc does too in his post) points out DEC were light years
ahead in this. Vax clusters essentially worked in this way, and they
were absolutely amazing - 1980s DEC tech that noone else has come
close to in the 15/20 odd years since.
> I think there's more to it than that,
surely must be.. i was just daydreaming. :)
> for example every piece of hardware access (not just memory) will
> have to be dealt with.
[snip]
the idea as proposed is just in germination stage at best i think. i
wouldn't have a clue what'd be involved in implementation though..
obviously those things'd have to be worked out.
the memory access stuff wouldn't be massively difficult (from the
little i know). presumably also takes care of memmory mapped IO. IO
space access can (surely?) be trapped and handled too.
anyway... i'm way out of my depth. but if a guy like larry mcvoy
thinks it's the way forward, and if DEC already essentially have done
it, then maybe there's something in it.
as far as i'm concerned: if one day i can get a bunch of PCs install
linux on them, mount 'smpfs' on all of them and have them /truly/ act
as one from the point of view of userspace, i'll be dead impressed.
and if you could then yank the power cord on one of the PCs and have
the thing continue to work (eg a couple of unlucky processes
segfault), then i'd be in heaven.
:)
--paulj
¹ and from that search i note that i've waffled about this before on
ilug, and that kenn did his usual "well the VAX..." posting too. :)
² and by god is IRIX 6.5 slow. Solaris is an absolute sow too.
This archive was generated by hypermail 2.1.6 : Thu 06 Feb 2003 - 13:10:22 GMT