Name: SALIM SUHET MUSSI
Type: MSc dissertation
Publication date: 15/12/2011
Advisor:

Name	Role
MOISÉS RENATO NUNES RIBEIRO	Advisor *

Examining board:

Name	Role
MAGNOS MARTINELLO	External Examiner *
MOISÉS RENATO NUNES RIBEIRO	Advisor *

Summary: Internet traffic is dominated by short data transfers. However, short flows account for a
small portion of the total link capacity. In addition, short flows unfairly compete for resour-
ces with connections that carry large volumes of data. The performance of TCP (Transmission
Control Protocol) operating at slow-start phase (or under the small transmission windows) are
impaired when sharing buffers and with long bursts coming from sessions at the stage of con-
gestion control. One way to mitigate this inequality is to treat short and long flows differently.
In this work we investigate stateless techniques in order to achieve a good compromise between
efficiency and complexity of service differentiation between short and long flows. From the
methodological point of view, we adopt an experimental approach using PC-based routers un-
der real traffic. The goal is to achieve trustworthy to the results, since they are not tied to parti-
cular features of simulation models, which are often overlook the actual behavior of a network.
We use the Click platform for routers’ implementation. The evaluation methodology for the
different mechanisms is based on a controlled environment considering a backup of a hard-disk
via FTP (File Transfer Protocol). We present what is, in the authors’ best knowledge, the first
physical implementation of a router with the RUN2C mechanism (Running Class Number 2).
In addition, we investigate its performance for different scheduling techniques. Our main con-
tribution is a new method for flow differentiation called RAFLE (Random Lengths Assorter of
Flow). In contrast with RuN2C, RAFLE requires no change of current network protocols. The
classification of packets belonging to long and short flows is inferred from a small table with
identification information of the last forwarded packets, but with no need for keeping the full
flow states. Results show that RAFLE’s performance exceeds RUN2C’s. Moreover, RAFLE
is able to virtually reach the performance of systems with full knowledge flows (full-state) in
different traffic scenarios.

Access to document