[20920] in SIPB IPv6
Re: limekiller pkgsrc update
daemon@ATHENA.MIT.EDU (Ken Raeburn)
Thu Dec 31 17:49:31 2009
From: Ken Raeburn <raeburn@MIT.EDU>
To: Quentin Smith <quentin@mit.edu>
In-Reply-To: <alpine.DEB.1.10.0912311149430.26286@dr-wily.mit.edu>
Date: Thu, 31 Dec 2009 17:49:24 -0500
Cc: Greg Troxel <gdt@work.lexort.com>, sipbv6@mit.edu
On Dec 31, 2009, at 11:50, Quentin Smith wrote:
> Didn't we express concern over jitter as a result of the hypervisor
> scheduling? I think I'd be a little wary of using a domU on a
> heavily loaded host as a router.
I raised the question, but mainly because I don't know enough about VM
responsiveness in such configurations. I do know from experience that
VoIP service on a virtual machine can be a bit of a problem; I presume
running VoIP (or worse, video conferencing) *through* a VM router
would be similarly affected, but it would have the advantage of not
needing to push the packet data into and out of user mode, and
shouldn't need many CPU cycles to process, so maybe it wouldn't be as
bad. I doubt anyone's doing VoIP over our IPv6 setup, so maybe it
doesn't matter. (If anyone is, please say so!) Unless the VM
responsiveness is really bad, I wouldn't expect ssh to be disrupted
much.
(Maybe the work on supporting real-time apps under Xen would help too,
someday.)
I think it'd be worth doing the experiment, at least, especially if we
gather some numbers with the real hardware and with the VM to compare
performance, variations in latency, etc. Aside from splitting off the
w20net support (and it apparently wasn't a big deal that that was
offline for a while), it *should* be pretty easy to take a NetBSD VM
on the right network, and make it our new IPv6 router. It'll need to
use the same address, so that does mean coordinating with keeping the
current limekiller offline (or at a different address, maybe
lemonkiller's if that's still offline, and not sending router
advertisements, etc). I wouldn't reconfigure limekiller as the local
w20net IPv6 router right away, as if we do find problems with the VM
router, we'll want to be able to re-enable limekiller as the central
IPv6 router easily...
Hrm... actually, a *different* experiment would be easy enough to set
up. Create a virtual machine on xvm with two IPv6 tunnels, or one
tunnel plus 6to4, and compare its performance with limekiller's. At
least one test endpoint should be a machine on campus, with IPv6
service through limekiller also (preferably directly, meaning
something on SMR-net or w20-net, or one of the on-campus routers), and
the other, if not on campus, at least not getting too much variability
in IPv4 RTTs to campus. I'm heading out tonight, but maybe I could
set something up in the next few days, if no one else gets to it.
Ken
