The following excerpt is from Datacate’s Colocation Survival Guide – get the full Guide here.

When shopping for Internet bandwidth, you’ll likely be faced with a decision: do you provision a flat-rate (unmetered) connection, or opt instead for a fractional (metered) commit? The answer depends upon the deal you are getting, and what your usage profile looks like. 100Mbps flat-rate has become an increasingly common bandwidth allocation, and often times the financial increase going from a fractional commit to a flat-rate commit on a 100Mbps line is marginal, which makes taking 100Mbps flat-rate a no-brainer. If you are truly counting pennies, you may be able to save a few by dropping down to a fractional commit – and if you are looking are a higher port speed, such as 1Gbps, the financial differences between fractional and flat-rate commits can be much more pronounced. In those cases, a fractional commit may make sense, but you must consider closely how your billable usage will be calculated.

How “Nth Percentile” Calculations Work

If you’ve inquired, you may have been told that your usage on a fractional commit bandwidth port will be calculated using the “95^th percentile method” (some carriers may instead use another number, such as the 90^th percentile method, but 95^th percentile is the most common method). This type of calculation is routinely used by dedicated carriers as well as colocation providers whom have a “house” offering. This is how the method works: samples for your data rates (speed, not transfer) in both directions are collected and stored in a single list at regular intervals, usually every five minutes, over a 30 day period. At the end of that 30 days, the list is sorted, and (using the 95^th percentile method) the top 5% of the measurements are discarded. The highest of the remaining measurements is designated as your billable usage for that period. From this, you can deduce the following:

• Since data for both directions (in and out of your interface) are stored in a single list, only the direction of greater flow matters for billing purposes, the other direction is ignored;
• Five percent of a 30-day period equates to 36 hours. Since the top 5% of measurements are discarded before calculating usage, you essentially get 36 hours of free bursting to the limit of your interface during each billing period. That can translate to a considerable amount of free usage on a fast connection.
• If your bandwidth usage is characterized by frequent periods of high usage interspersed with periods of very little usage, the Nth percentile method could potentially work against you. It all depends upon the frequency and duration of the peak usage periods. Once you aggregate peak usage crosses that “magic number” (36 hours in the case of 95^th percentile monitoring), each and every additional peak could potentially be your billable usage number for the entire period.

The takeaway from the above is this: since providers typically offer discounted rates on flat-rate connections, you may wish to consider flat-rate if you expect frequent peak usage intervals that with push you past the Nth percentile threshold. In that event, the provider will usually bill at overage rates that are priced at a premium; as a result a substantial overage could end up costing you more than a flat-rate port would have. Alternatively, if you expect your usage to be generally low with infrequent bursts, a fractional commit may be the best use of your bandwidth dollars.

If you are unsure of your usage profile, e.g.: in the case of a new project for which no historical bandwidth usage data is available, consider seeking a bandwidth plan that allows you to get a minimal commit on a port of the desired size as a starting point, with the option to upgrade at any time without penalty. In this way you can gauge bandwidth demand over time, and if it turns out that you need more, you simply increase your commit at preferred rates, rather than paying for pricey overages.