Cisco – Determining optimal bandwidth needs for the office

Expanding upon @javano's answer...

I would change one thing, instead of policing the email traffic (which will drop traffic exceeding that rate) I would recommend just setting aside a small amount of bandwidth for that traffic. Therefore, if nobody else is using the link the email can utilize the unused bandwidth

policy-map PM-Limit-Emails
 class CM-Match-Critical-Apps
  bandwidth percent 50
 class CM-Match-Scavenger
  bandwidth percent 1
 class class-default
  bandwidth percent 20
  fair-queue

I didn't get a sense of if the issue was the sending (outbound) or receiving (inbound) of large attachments so, since you are just connected to the internet, there will be the potential that this issue will continue as the ISP is not throttling the traffic (inbound) based on your classifications. If you have a business class internet service there may be the possibility for them to install a similar policy on their end of your link but you will need to talk with your ISP to determine what, if anything, they will support.

Assuming you're untagged, your current stack on the wire is:

• 7B preamble
• 1B SFD
• 6B DMAC
• 6B SMAC
• 2B type
• 20B IPv4
• 20B TCP
• nB application data
• 4B CRC
• 12B IFG

First thing is very important, you are looking at L2 overhead, you must consider L1 overhead also.
For ethernet preamble, sfd and ifg are L1, they are not really bytes but ethernet defines them strictly at byte sized amount of time.

For ATM you can't get very accurate reading with your methodology, because it assumes static overhead. If you try send payload of 49B you'll fit it in single ethernet frame, but you'll need two ATM cell, second cell has 52B of overhead and 1B of data.
This problem does not exist in HDLC, PPP or FrameRelay, and your methodology will give decent approximation on those.

If you want really accurate data, you need to have size of each packet send, then calculate what they are when serialized to given L1 technology with given L2 technology.
The larger the packets are that you send, the less sensitive your calculations are to the approximation method as overhead will be smaller contributor.