Buffers, Congestion, Jitter, and Shapers « ipSpace.internet weblog


Béla Várkonyi left a nice remark on a weblog publish discussing (amongst different issues) whether or not we want massive buffers on backbone switches. I don’t know the way many individuals learn the feedback; this one is simply too invaluable to be misplaced someplace under the fold

You would possibly need to add one other consideration. In case you have a whole lot of visitors aggregation even when the ingress and egress port are roughly on the similar velocity or when the egress port has extra capability, you can nonetheless have congestion. Then you might have two methods, buffer and undergo jitter and delay, or drop and hope that the higher layers will detect it and scale back the sending by shaping.

However you’re proper, deep buffers imply excessive jitter that may be transformed into excessive delay sooner or later by dejitter buffers. In case you have a visitors sample that may be a type of distribution into a number of instructions, then you don’t want to concern about congestion and deep buffers. Even when the distributing interfaces might need much less velocity however they nonetheless match to the disaggregation patterns. For instance, quick ethernet entry ports and gigabit uplinks. There are a lot of conditions like that. As common, an optimum community design requires the information of your visitors patterns. In any other case, you may simply overdesign as a lot as allowed by your budgets.

Additionally it is essential to have in mind the policer/shaper chaining guidelines. If you wish to scale back congestion, then there have to be a shaper on the sender for complying with the policer on the receiver. The bodily capability of the hyperlink is a pure policer, even when you have not configured a policer explicitly. Contained in the router/swap you might have an analogous pairing, you simply should depend with the aggregation issue, too.

Such a sequence of shapers could be very troublesome to handle, so a backpressure mechanism edge-to-edge or end-to-end could be wanted for the rescue. Sadly, some visitors doesn’t have backpressure, corresponding to sure UDP streams. Then you might have a foul luck and both configure all of the shapers based mostly on a visitors patterns, or make a compromise by tolerating packet loss and jitter.

I’ve seen a whole lot of issues in networks when shapers have been forgotten… You can’t make wonders, so it’s a must to think about them…


Supply hyperlink

Leave a Reply

Your email address will not be published.