|
Here
are a few belated comments ... sorry I didnt get them in
earlier.
- Foster et al in their "Anatomy" paper wisely suggested focusing on
interactions rather than internals in their discussion of GRID architecture.
I'm hoping that this paper will be applicable to a range of optical networking
architectures/technologies.
- This
raises the issue of defining what sorts of networks are in scope.
I think something like the following will cover the networks we're likely to
see in the long-haul world for some time: Connection-oriented, synchronous,
constant bit-rate, very low BER, bit rates 10**8-10**11/sec, connection set-up
time sub-second but not sub-millisecond, rapidly restorable (~100 ms) if
desired. Somebody wisely pointed out that "optical" is a much-misused
term ... this definition covers most of what's currently called "optical
networking" and includes both optical and electrical switch
fabrics. If we want to cover LAN's, this will need some
extension. Including optical packet switching seems less important in
the short term, altho definitely of interest longer term. Seems like quite a
challenge to say something meaningful without bounding the network
characteristics somehow.
- The
VO's described in the "Anatomy" paper block out a very broad set of
applications/users. If we want to have this breadth of use in mind I think
it's crucial that from the very start it's assumed that providing connectivity
for a VO will require cobbling together heterogenous networks. I'd
expect to see a single VO using a mixture of customer-owned and carrier
facilities, for example, and multiple administrative domains. Also a
variety of switch technologies. This has a number of implications:
- Inter-domain control has to be considered. There's quite a bit
going on here that can be drawn on ... Canarie's OBGP. Also the OIF
demo-ed an optical NNI at OFC in June, with ~10 vendors demonstrating the
ability to rapidly set up connections with other vendors (required control
plane interworking).
- Achieving complete user control of routing, etc. is going to be a
major challenge for technical reasons (routing constraints may depend on
technology and so vary between domains) and also for
scalability/controllability (there's a strong similarity to the situation in
the internet, where the information advertised by BGP between domains is
quite limited). In carrier space, I'd suggest focusing on something
like optical VPN's. This is a very active research/standards area right
now.
- Consequently I think the GRID world needs to think carefully about
just what they do want to control and what state information they really
want/need. If you specify complete information about network internals
you're going to get the same push-back that Internet ISP's provide. But (my
opinion) you definitely need much more information than is advertised
by the Internet's BGP.
- If
you accept that you need to interwork with the control planes of commercial
ON's and commercial ON equipment, detailed attention needs to be paid to the
work in IETF (GMPLS), OIF (UNI, ONNI), and ( :-( ) ITU (ASON).
None of these are solutions to the GRID problem, but the GRID middleware will
need to interwork with them to get needed state information and for
reconfiguration. I'd also suggest that it'd be good to use terms
like "peer-to-peer" and UNI in a way that's consistent with the terminology
being used in these groups ... it'll get rid of a lot of downstream
communications problems.
- As a
GRID outsider, I would find it useful if explicit connections were made with
the GRID layered model as Foster et al laid it out (or some similar
model).
- The
motivation given for interest in ON's is bandwidth and cost. Are there
more subtle opportunities provided? For example, might scheduling take
advantage of the transmission time predictability inherent in synchronous CBR
connections?
I
could provide more detail/background/references on most of these items if
desired.
John
John Strand
AT&T Transport Network Evolution
Dept.
200 Laurel Ave., Room A5-1D33
Middletown, NJ
07748
+1 732 420-9036
jls@research.att.com
Here is the draft memo on optical networking for GRID
services.
Please send your contribution
&comments.
I will be happy to compile the final
document.
Dimitra
P.S. I will be away until the 20th of July
but my colleague Reza Nejabati (rnejab@essex.ac.uk) will be around to
deal with any urgent queries.
------------------------------------------------------------------------------------
Dimitra
Simeonidou
Professor, High
Capacity Networks & Distributed Systems
Photonic Networks
Laboratory
ESE
Department
University of
Essex
Colchester CO4
3SQ
UK
tel: +44 (0)1206
873496
fax: +44 (0)1206
872900
http://esewww/~dsimeo/
|