Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

RE: Traffic characteristics



I think RPR should support other traffic as well, but it should at least address TCP/IP. I agree with BJ that self similar models and trace driven simulations are not sufficient to evaluate PRP protocol performance. We should use real TCP/IP sources in simulations. Traces or models don't capture the interactive flow control of TCP. RPR protocol may very sensitive to this interactivity.
 
We can look at results from sites like www.caida.org to determine the packet size distribution, flow length, number of tcp flows, number of UDP flows, type of applications etc on the backbone.  These parameters should be used to simulate real TCP sources.
 
-Sanjay K. Agrawal
-----Original Message-----
From: Byoung-Joon (BJ) Lee [mailto:bjlee@xxxxxxxxx]
Sent: Thursday, August 24, 2000 10:23 PM
To: Raj Sharma
Cc: Taylor Salman; Raj Sharma; 'Khaled Amer'; Reflector RPRSG; Charles Barry; Atul Shinde
Subject: RE: Traffic characteristics

Raj,

Thanks for the clarification.  It seems that my previous email may have given
the wrong impression that RPR should only support IP services by emphasizing
the TCP/IP traffic modeling.

It would be great if there could be a slot assigned at the next week's meeting
for more discussions on this topic (e.g., RPR vs Gigabit Ethernet, how to
support services without requiring IP functionality, pt2pt ethernet over WDM on a
physical ring, what additional value we can bring with RPR, etc).

Cheers,
BJ

At 01:16 PM 08/24/2000 -0700, Raj Sharma wrote:
BJ and RPRSGers,
 
I perhaps come across garbled in my previous email.
I really dont mean to imply to standardize services.
What my concern is that we have to account for
applications where Data, Voice and Video that might
get be carried over this protocol without requiring
IP layer functionality. Each of these services has a
specific traffic pattern and boundary conditions. This must
be our top level requirement. Then we can drill down into each
of these broad service categories and get more granular.
 
My concern with RPR is that the feedback I am getting from
our customer base. They feel that if RPR deals primarily
with IP traffic and they are expected to figure out the utopian
path of voice or video over IP then this is a far stretch
of an imagination from the reality that exists out there.
Hence, they are interested in provisioning more than IP
traffic over a layer 2 technology. If RPR does not make
that diffrentiation than why would they not use point-to-point
ethernet over DWDM on a physical ring? If RPR is just an
evolution of POS we can pretty much pack our bags and go
home.
 
Further, from a service perspective it is the simple provisioning
paradigm that is of great interest to them. Packet switch based
technology provides a simple "LAN-like" provisioning paradigm.
However, this can only become a true benefit if it applies to all services
and not just to one service.
 
I urge everyone to take a look at how gigabit ethernet is a viable
metro solution and what additional value we can bring with RPR.
The traditional problem that buffer insertion rings have faced
in the past is the compelexity of administering fairness or
deliberate unfairness over the ring in store-forward, multi-hop
network. If all we are worried is about TCP/IP traffic than
point-to-point ethernet resolves that very elegantly. BTW,
point-to-point ethernet can also use a physical ring deployment.
 
There has been a lot of talk of the speed of protection
switching. Realistically, if you are only carrying TCP/IP
traffic than you will get a glazed look in these customers
when sub 50 millisecond protection switching is touted.
The reason for this is fundamentally, protection mechanism
as to at the same layer the service gets provisioned. For SONET
it made sense because all services were provisioned over
SONET.  This is another reason why we must understand
what services (other than IP) will RPR provide to make its
protection mechanism more meaningfull than the lack of it
in Ethernet.
 
This is an issue for RPR in general to gain acceptance as
a serious upcoming paradigm. We must clearly make a
difference over Ethernet or E-over-SONET, etc.
 
comments?
 
 
-----Original Message-----
From: Byoung-Joon (BJ) Lee [mailto:bjlee@xxxxxxxxx]
Sent: Thursday, August 24, 2000 12:36 PM
To: Taylor Salman
Cc: Raj Sharma; 'Khaled Amer'; Reflector RPRSG; Charles Barry; Atul Shinde
Subject: RE: Traffic characteristics

Hi,

Among many web sites, the following two could be useful for traffic traces
and packet size distributions:
   traffic traces:
        http://www.acm.org/sigcomm/ITA/
   packet size distribution:
        http://oceana.nlanr.net/NA/Learn/packetsizes.html

However, the problem with the trace-driven or self-similar traffic simulation is that
we do not explicitly model the dynamic nature of the TCP/IP traffic reacting to
congestion.  Unless for the one-way UDP or background traffic modeling, such
aggregate traffic characterization may not be useful.

As for the Raj's argument for the service modeling, I would say we are expanding
the scope too much at this time?  It could be argued that we are not defining or
standardizing the "services"?

Cheers,
BJ

At 10:08 AM 08/24/2000 -0400, Taylor Salman wrote:
Expanding on some of these points.  Yes, self-similar traffic is a good way to go to simulate large amounts of IP traffic.  Self-similar traffic models are already available that allow you to bypass modelling of the TCP/IP stack (although, I believe that BJ has some concerns in this regard).  Bo Ryu from Hughes Research Labs has done extensive research in this area.  I can put people in touch with him if they are interested.

In my opinion, there should be a series of scenarios that test different aspects of the system.  Perhaps, self-similar traffic is one test for the performance of TCP/IP traffic.  It is also possible to get models of live traffic to play over the models, as well as, standard application traffic models such as HTTP, FTP, e-mail, video, voice, etc.  I agree with Raj in that a determination of typical traffic mixes needs to be made, and different scenarios constructed to model these mixes.  My opinion does differ from Raj, however, in that I think pathological cases can provide as much insight into system behavior as typical cases and scenarios should be constructed for both, particularly for a protocol where one of the main goals is resiliency in the face of link failure.

Taylor

At 06:25 PM 8/23/00 -0700, Raj Sharma wrote:
You could use self similar traffic patterns.
There is quite a bit of publication on such traffic
patterns. William Stallings had a book that spent an entire
section on it. (High performance networks ---- or something
like that)
 
However, isnt this more of a "regression" parameter
rather than "main" parameters we want to test our
Layer 2 protocol against. Much of the issue that
we need to sort out for these main parameters
are the ranges of value. Cooking up these values must
make practical sense and they cannot be pathological
cases to test boundary conditions.
 
Perhaps what we need to do is to move one level up
and ask ourselves what services will get provisioned
over (layer 2) RPR and hence what are the main parameters.
Remember you cant design a layer 2 protocol only
based on a specific layer 3 protocol. If that is the case, I rather
do MPLS switching over rings.
 
Thoughts?
 
 -----Original Message-----
From: Khaled Amer [mailto:khaledamer@xxxxxxx]
Sent: Thursday, August 24, 2000 3:23 AM
To: Reflector RPRSG
Subject: Traffic characteristics

RPR'ers,
I raised this question about 3 weeks ago and got the familiar underwhelming response!
Several people indicated availability of traffic patterns that were taken off the Internet or other networking environments that would be useful for us to use in the simulations. In my mind, we need things like packet size distributions, interarrival times, burstiness, application and protocol distributions, and any other relevant characteristics of network traffics that we may want to consider in the simulations.
Does anyone have such information that they can share with us in the meeting next week?
 
Khaled Amer
President, AmerNet               
Architecture Analysis and Performance Modeling Specialists
Phone: (949)552-1114             13711 Solitaire Way, Irvine, CA 92620
Fax:     (949)552-1116             e-mail: khaledamer@xxxxxxx

**********************************************************************
P. Taylor Salman                Phone:  202-364-4700x2297
Manager, Market Research        Mobile:  202-427-3319
OPNET Technologies, Inc.        E-mail:  tsalman@xxxxxxxxx
3400 International Drive, NW    Web:     www.opnet.com
Washington, DC 20008
**********************************************************************