"Packet arrivals are not Poisson
.... but some events are, such as web requests and new flow arrivals"
I know since the network traffic is very bursty, they are not Poisson.
But I am unable to understand what "new flow arrivals" mean here and how web requests and new flow arrivals can be considered as Poisson processes.

  • 1
    $\begingroup$ Understand that the Poisson distribution is only a model here, one which apparently matches real traffic somewhat closely in practice, sometimes. I would not expect any real scenario to actually follow a Poisson distribution. $\endgroup$ – Raphael Nov 21 '15 at 11:03

"new flow arrivals" means "arrivals of new flows". A flow is a TCP connection (roughly); each individual TCP connection is a separate "flow". So, this is talking about new TCP connections, and the rate/time at which the server receives new TCP connections.

Contrarily to the statement you quoted, web requests won't necessarily be Poisson. There are many factors that can cause web requests to be other-than-Poisson-distributed. For instance, the Olympics web site receives a lot more traffic when the Olympics events are happening than when they aren't. Many web sites receive more traffic during the time of day when their primary user population is awake than when they're not. In addition, visiting one page on a site typically triggers many requests for other related resources (e.g., images that are loaded on that page).

However, during any sufficiently short time period, if their user base is sufficiently large, the arrival process for a particular type of web request might well be approximately Poisson.

See, e.g., https://en.wikipedia.org/wiki/Poisson_distribution#Occurrence.

  • 1
    $\begingroup$ If the requests are Poisson-distributed, the packets will not be because each web request will involve multiple packets. (As noted, an initial GET request for an html page will often be quickly followed by requests for related resources such as Javascript, CSS, and images.) On an intermediate time scale, a Poisson model seems problematic because hypertext links to site-local resources would seem to encourage violation of event independence, but for a simple model it might be adequate for deriving useful observations. $\endgroup$ – Paul A. Clayton Nov 21 '15 at 13:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.