Many web businesses are hosted through a datacenter, and there are many pricing plans depending on your bandwidth needs. If your site is heavily text based without a lot of images and steady traffic, then a flat rate might be best for you . If you are like most web businesses, then you need dynamic bandwidth that can handle spikes, but don’t want to overpay for unused traffic.
Much like cell phone plans, you don’t want to overpay for minutes you’re not using but don’t want to guess too short and have to pay overage charges. A 95th percentile pricing schedule can dynamically bill you based on your bandwidth needs, but is a little complicated to follow.
I discovered the pitfalls of this pricing schedule last month when I posted commercials from SuperBowl 40. Since the OCmodshop server has 100Mbit port to the internet, I figured that I could use up to 95% of that and be within my billable limit. I severely misunderstood how this pricing schedule works, as probably do most datacenter customers.
95th Percentile is a method of measuring bandwidth that bases your bill on peak utilization. Your bandwidth is measured from the switch or router and recorded in a log file. At the end of the month, your usage statistics are sorted, and the top 5% (approximately 36 hours of a 30-day billing cycle) of data is thrown away, and that next measurement becomes your ‘billable utilization’ for the month.
Usually only a few services are billed using 95th percentile pricing: Dedicated T1 and Colocated servers. DSL and Wireless customers are usually billed on the total bandwidth used over the entire billing period.
Some will argue that the 95th percentile billing method favors the datacenter, others will counter that it benefits the customer. It is true that there is potential for paying for bandwidth you didn’t use, but it also allows you to dip into peak bandwidth when you need it. Based on the 95th percentile model, the top 36 hours (top 5% of 720 hours) of peak traffic aren’t considered when billing for an entire month.
If your website was slashdotted for an entire weekend (48 hours) then you would be billed based on a higher rate just because of the one weekend. A good datacenter will allow you to upgrade or change your plan if you notice these type of peak patterns.
Every 5 or 6 minutes (or whatever increment your datacenter has set), two readings are taken of the total bandwidth used by your network. One outgoing and one incoming reading are taken on synchronous connections. The higher of these two readings is recorded as the bps transferred during this time. At the end of the billing cycle, these readings are sorted and the top 5% are discarded. The remaining readings are what you are billed on.
For these services, the data quota is based off a 95th percentile threshold. Billable quota is based on an entire month of readings at the 95th percentile. Many datacenters will report hourly data, but is only useful to get a general idea of how much bandwidth is being used. Billing is usually determined using an entire month of recorded measurements.
For example, if your datacenter takes 10 readings in an hour, then 240 readings will be taken in one 24-hour period. This adds up to 7200 readings in a 30-day month. In a one hour period the 95th percentile cannot be calculated into a round number, since removing 1 entry will leave 90 percent.
In order to be able to display the hourly data, your datacenter may round down to the 90th percentile. There are many more data points in a day or a month than in an hour, so 95th percentile can be calculated more accurately.
The graphs are examples of bandwith useage for a partial month, and consist of 4000 readings. The first graph shows an example of bandwidth with lots of sporadic peaks and spikes.
The second graph shows a much smoother curve once the 95th percentile billing model filters out the top 5% of infrequent peaks. You can easily see the average network trend and bill the customer based on this average. In this instance, the customer is billed for the highest remaining entry, which is 48.3Kbps. This places the customer in a 32Kbps – 64Kbps tier.
Some datacenters have a particular bandwidth commit, which means you pay for a minimum tier whether you use it or not. If you have a 1Mbit commit and only use 64Kbps for the month, then you are billed for 1Mbit even though you haven’t used the rest of your available bandwidth.
Over a period of 600 minutes, 100 readings are taken of both incoming and outgoing traffic. If the billing period was also 600 minutes, the top 5% of these readings (5 in this case) are ignored, and you are billed at the value of the 95th reading.
If the top 10 readings of this set of 100 were:
| rank | bandwidth |
| 100 | 120Kbps |
| 99 | 109Kbps |
| 98 | 98Kbps |
| 97 | 80Kbps |
| 96 | 74Kbps |
| 95 | 30Kbps |
| 94 | 28 Kbps |
| 93 | 27Kbps |
| 92 | 24Kbps |
| 91 | 21Kbps |
The customer would be billed using the 30Kbps entry, falling within a 16Kbps tier.
The 95th percentile billing method is a tool many datacenters use to leverage bandwidth to their customers. Understanding this billing is very complex, and you will have to analyze if this billing method is right for you.
Every datacenter expects to have some spikes, and the 95th percentile allows customers to have alot of potential bandwidth without charging the customer for every single peak that goes over their normal commit. It is possible to be over-charged using this billing method, which is why it is important to monitor your network usage and adjust your billing plan accordingly.