Re: [RPRWG] RPR Perf: Non-uniform distribution
All,
I've been watching the conversations on this topic and agree with the
thoughts being exchanged.
In performance modeling, it has been my experience that having very uniform
patterns would sometimes hide other performance characteristics.
How about if we take a tough scenario that will push most implementations
and help us understand the merits of various proposals.
Something like a ring with 2 points of connection (emulating a ring
connected to two rings or a ring with two POPs), and take one of them to be
much more loaded than the other. So, we can talk about a dest address
distribution of 20% and 80% or even 10% and 90%. I think that this will give
us much more useful information than the uniformly distributed dest address
that we originally had.
Any comments or suggestions?
__________________________________________
*** Please notice new e-mail address ***
Khaled Amer
President, AmerNet Inc.
Architecture Analysis and Performance Modeling Specialists
Address: 13711 Solitaire Way, Irvine, CA 92620
Phone: (949)552-1114 Fax: (949)552-1116
e-mail: amer@xxxxxxxxxxx
Web: www.performancemodeling.com
----- Original Message -----
From: Lewis Eatherton <leathert@xxxxxxxxxxxxxx>
To: Komal Rathi <krathi@xxxxxxxxxxxxxx>
Cc: 'Timbs, Jeffrey L (Jeff)' <timbs@xxxxxxxxx>; <stds-802-17@xxxxxxxx>;
<seti@xxxxxxxxxxxxxx>
Sent: Thursday, April 26, 2001 9:05 PM
Subject: Re: [RPRWG] RPR Perf: Non-uniform distribution
> Actually, in several of our larger rings, we currently have 2 to 4 points
with
> connections to our core. With most LEC infrastructure being OC48, I can
only
> see this trend continuing. On the other hand, in most major markets,
there is
> only one POP (depending on who your long haul connection is). (just the
@Home 2
> cents)
>
> Komal Rathi wrote:
>
> > Hi Jeff,
> >
> > I agree with you that there will probably not be high multi-homing
within a
> > MAN ring in the near term. A more realistic scenario will most likely be
2
> > or more hubs (POPs) on the ring and unequal traffic going "off-ring" at
each
> > hub. Do you have any suggestions for a set of realistic non-uniform
traffic
> > patterns?
> >
> > Thanks,
> > Komal
> >
> > -----Original Message-----
> > From: Timbs, Jeffrey L (Jeff) [mailto:timbs@xxxxxxxxx]
> > Sent: Thursday, April 26, 2001 6:04 AM
> > To: 'stds-802-17@xxxxxxxx'
> > Subject: RE: [RPRWG] RPR Perf: Non-uniform distribution
> >
> > Khaled et al:
> > I agree with Komal's suggestion below. I would like to see the behavior
when
> > a large % of traffic goes "off-ring" through few (1 or 2) nodes. It has
been
> > reported that pass-thru : add/drop traffic is higher than 80:20 due to
the
> > long-distance nature of IP traffic. [Rodriguez-Moral et al, "Optical
Data
> > Networking," Journal of Lightwave Technology, Vol. 18, No. 12, December
> > 2000] A further implication is that most IP traffic is not terminated
within
> > a MAN, leading to hot-nodes (or gateways) which are destinations of the
> > majority of traffic. Assuming that MAN rings are connected to backbone
> > networks via few (one or two) nodes, a high percentage of traffic will
be
> > destined for these few nodes.
> >
> > Some interesting points are that, of 100 "leading" websites, WorldCom
could
> > carry traffic to up to 45 of them entirely on its own backbone while MCI
> > could reach 28 sites directly (Sprint, 18 sites). Also, 35 of the sites
> > were multi-homed on the WorldCom backbone. [cf. OECD pub
> > DSTI/ICCP/TISP(98)7/FINAL via
http://www.oecd.org/dsti/sti/it/cm/index.htm]
> >
> > It's not clear that there would be a high degree of multi-homing within
a
> > MAN ring (an economical factor driven by localized demand). For a high
> > degree of destination stripping distributed across many nodes, several
> > services would need to be offered at many sites (POPs). To me, there is
a
> > current assumption that new service providers want to populate as few
POPs
> > as possible, so that traffic must be carried to a centrallized site
within a
> > region which also decreases the likelihood of uniform distribution.
> >
> > For near-term introduction of RPRs, we should consider that this traffic
> > pattern may be more representative of a MAN RPR and should consider the
> > impact on performance (due to lower overall spatial reuse).
> >
> > Jeff Timbs
> > Agere Systems, Inc.
> >
> > -----Original Message-----
> > From: Komal Rathi [mailto:krathi@xxxxxxxxxxxxxx]
> > Sent: Tuesday, April 24, 2001 1:50 PM
> > To: 'stds-802-17@xxxxxxxx'
> > Subject: [RPRWG] RPR Perf: Non-uniform distribution
> >
> > Khaled and members of the Perf. Adhoc committee,
> >
> > Given that all implementations seem to have destination stripping that
> > results in spatial reuse, it's not clear in my mind what we gain out of
> > running simulations with randomly and uniformly distributed
> > source/destination pairs. I believe that the uniform distribution will
not
> > provide enough information that would help demonstrate the differences
in
> > performance characteristics of various proposed architectures. All it
does
> > is show that we get some spatial reuse, which we know all
implementations
> > do.
> >
> > I would like to suggest that we make the destination addresses
distributed
> > non-uniformly which can show some more interesting performance
> > characteristics of various architectures that are being proposed. For
> > example, how about something like 80% to dest-A and some other
distribution
> > to other destination nodes. Another idea would be to have 90% to
neighboring
> > node A, and 10% to the other neighboring node B.
> >
> > Komal Rathi
> > Lantern Communications
>