RE: Network Sizing Guidelines?

I'd recommend taking a look at: Wide-Area Data Network Performance
Engineering By Cole, Robert G. / Ramaswamy, Ravi
ISBN: 0890065691

Here is a guideline based on Ravi's methodology that I wrote for our
customer newsletter some time ago:

The bandwidth size that is required for any given connection is a function
of the following three factors, number of users, requirements of specific
applications, and how the application is used. For example, a site with five
users that all access a highly interactive application for twelve hours per
day may require more bandwidth than a site in which a dozen users
sporadically access a client-server application in which most of the
processing is performed by the remote server.

In addition, another concern in the bandwidth selection process is delay.
Certain applications such as voice and video may require a low level of
delay (latency) as well as a low variability in delay (jitter). These
requirements may add significant complexity to the design process.

The first step in sizing bandwidth is to determine the requirements for the
specific applications that will be deployed. During this step, a sniffer is
useful in tracing application sessions to determine the average packet size
and the average number of packets for a given transaction. Once you have
these values, the next step is to factor in the number of users, the
required latency, and the amount of time that typically exists between

Once you have obtained these values, you can use the following formula
(created by Ravi Ramaswamy of AT&T Solutions) to determine bandwidth

8 x N x K x M / (K x P + T) Where:

N = number of active users at a location (the number of users that will
simultaneously use an application)

T = User think time (how much time typically exists between inquiries

K = number of packets per transaction in any given direction

M = number of bytes per packet in any one direction

P = one-way network latency

Note that this calculation must be performed for both directions of the
connection. The required bandwidth is then the maximum bandwidth estimated
by this formula (unless you are deploying a technology such as Frame Relay
which allows for different bandwidth allocations for each direction of the

Note also that this formula only applies to client-server type applications
in which there is a substantial amount of two-way traffic. For additional
information on bandwidth sizing, please see "Optimizing Client-Server
Application Performance on the WAN" in the November issue of "Network

These guidelines will be badly misleading if you have serious
streaming use, where K, T and P are basically undefined or irrelevant.

BTW, I have asked numbers of people about the underprovisioning
(or oversubscribing) they use, and have received responses
ranging from a factor of 5 to a factor of 20 or more.

                                   Marshall Eubanks

Irwin Lazar wrote:


Thanks. Those are the kind of ballpark numbers I was hoping to get by asking this list. I know it doesn't seem like a range like that would be helpful, but that's the real world. I have also received some help that says depending on what your users do with the network that you can handle 100 users on a 128K ISDN and sometimes a T1 can only support 3 users. That at least helps me bound the problem. And, of course, there are wild cards like P2P that require predicting the future.

I could try to build some fancy models to make fancier estimates, but in the end it still comes down to guessing. I figured nanog would be the best place to ask for a good guess tempered by real world experience. Thanks again for some numbers.