Dr. Roland Bless

Description

The aim of this work is to carry out fundamental investigations into the dimensioning of packet buffers in routers and switches. The question of the size of buffers has already been the subject of heated and controversial debate in the past. The basic function of a packet buffer is to improve utilization by absorbing packets in short-term overload situations. However, the previous rule of thumb was derived from the behavior of classic TCP congestion control procedures. However, there are new procedures that behave differently, as well as increased requirements for lower latencies, which are, however, increased by excessively large buffers. In this respect, a renewed discussion about the correct dimensioning of buffers is necessary (see also https://buffer-workshop.stanford.edu/). It can be assumed that this also depends on the location of the router/switch, i.e. different rules may apply for an Internet core router than for a router in a data center or a DSL router at home.

Task definition

Within the scope of the work, various studies are to be carried out on the necessary dimensioning of buffers. The main focus here is on the characteristics of the incoming data packets (e.g. arrival process, i.e. average arrival rate, packet sizes and variances or burstiness of arrivals). Therefore, recordings of real Internet traffic will serve as a basis for developing an ML-based prediction of buffer occupancy quantiles. For this purpose, a suitable ML method is to be selected, trained accordingly and then evaluated with previously unseen recordings. Part of the investigations will be carried out simulatively (using the simulator OMNeT++).

Prerequisites

Basic programming knowledge, ML knowledge, Python/C++ knowledge is advantageous