0

Is there any common method (usable in practice) for estimating the uplink capacity of switch based on known number of users connected to this switch and the maximum internet speed they are allowed.

This is an example topology:

core switch ==== aggregation switch ---- access switch 1 ---- 10 users
                                  |----- access switch 2 ---- 30 users
                                  |----- access switch 3 ---- 50 users

I want to find the needed capacity (or maximum traffic in the busiest hour) for the links between aggregation switch and access switches 1, 2 and 3. I know they will be allowed to have maximum internet speed of 50 Mbit/s and no transfer between users.

In the worst case they will all download at 50 Mbit/s, so 50 users * 50Mbits/s = 2500Mbit/s but this case will almost never happen as the traffic comes on bursts and it will be waste of resources.

So I need a general method, best practices or any guide for this job. All I find in the Internet is theory with almost scientific value, but no information how they do it in IPSs and telecommunication companies in practice.

vladiz
  • 111
  • 4
  • Maybe. It is a specific and very common case, but I am sure it has been answered as something else before. – Falcon Momot May 07 '14 at 07:30
  • My question is about network capacity planning and the other questions are about server capacity planning which is slightly different, respectively the answers are different – vladiz May 07 '14 at 19:23
  • Yes, it's not an exact duplicate of that specific question at all. I also can't find anything this is an exact duplicate of, so I've nominated it for reopening. – Falcon Momot May 07 '14 at 21:26
  • This may not be an exact duplicate of the linked question, but it's close enough, IMO. The easiest way to tell is through testing and gathering real world data... in which case it's a duplicate of the linked question. Otherwise you would need to do statistical analysis of a "normal" network's peak demand versus users and allowed bandwidth speed... which would require testing and gathering real world data (still a duplicate) or going purely off assumptions, educated guesses and network bandwidth modelling, in which case, too broad/subjective, and also better suited for the stats StackExchange. – HopelessN00b May 07 '14 at 22:06

2 Answers2

2

The reason you're only seeing theory without many results is that queuing models are tricked by "packet trains", and TCP does congestion handling, which prevents queues. The theory says "I don't do that kind of stuff" (;-))

You need to find the actual number of users for your busy periods on a comparable link: same technology, not necessarily the same configuration. That and your population will give your the over-subscription, and then you can work out an initial estimator.

Having done all that work, then measure the results, figure out the margin of error, and save that away for next time.

A caveat: if this link is a bottleneck or is between two networks with different speeds, you're about to learn all about "bufferbloat". You won't enjoy it (;-))

Falcon Momot
  • 24,975
  • 13
  • 61
  • 92
davecb
  • 211
  • 2
  • 5
1

You will have to determine your oversubscription rate, which will depending on your application. It might be 20:1 (which is common for ISP's) or higher, like 4:1.

A single gigabit link will provide a 20:1 ratio for 50mb/s of traffic.

Bert
  • 2,733
  • 11
  • 12