The 95th Percentile and Bandwidth Contention
Posted on June 20, 2021 — 3 Minutes Read
The irregular traffic pattern by the time of day and by the day of week is one of the many difficulties of measuring bandwidth usage for connectivity service in a business environment. A typical office works 8 hours a day and 5 days a week, which implies that, given minimum usage in non-office hours, the theoretical maximum for bandwidth usage in any given week goes down to 8 hours divided by 24 hours, times 5 days divided 7 days a week, that comes to about 23.80%. This means if during the office hours the bandwidth usage is about 70Mbps on a 100Mbps subscribed bandwidth, at the end of a month the average bandwidth usage would be 23.80% of 70Mbps i.e. 16.66 Mbps. If one then bases a bandwidth downgrade decision on this meter, productivity will without a doubt suffer.
One metering mechanism that takes into account of this irregular traffic pattern is the 95th percentile, which instead uses the bandwidth meter sample that falls on the 95th percentile as a representative of the bandwidth usage for the period measured. If a bandwidth usage sample is taken every 5 minutes, in a month of 30 days there will be 30 days times 24 hours times 60 minutes divided by 5 minutes = 8,640 samples. With these sorted and ranked by percentile, 8,208 of those measurement samples will fall under the 95th percentile and will yield a more representative measure for the highly and positively skewed distribution of bandwidth usage in a business context. This neat trick of the 95th percentile provides a billing mechanism that allows an enterprise to subscribe to a connectivity service that permits bursting above the committed base information rate i.e. occasional and uncommitted bandwidth above the committed base bandwidth, without having to instead subscribe to a service with a fixed bandwidth of the peak usage for the entire contract term, which will be cost-prohibitive.
Given a diverse pool of bandwidth subscribers from various sectors and industries, each with a distinct traffic pattern, e.g. concentrated during business hours for enterprises, surging after-hour for gaming and entertainment, and round-the-clock low and stable bandwidth usage for trading and manufacturing, these distinctly irregular traffic patterns will cancel out each other, yielding a rather smooth and consistent bandwidth usage that will be necessary to an efficient transmission and wavelength resources planning. Harnessing then the secret weapon of a telecommunications carrier i.e. bandwidth contention or oversubscription, which by experience starts at 3x to as high as 5x for often the actual bandwidth usage is about 33% or in cases even 20% of the bandwidth subscribed, this allows for a lucrative margin for a well planned and managed network service. Glorious this may seem, as any insider would know, the telecommunications sector itself is not without threat, it is nonetheless a topic that deserves a dedicated discussion.