Unformatted text preview: state, in FRED dropping decisions are based on this ow state. Specifically, FRED preferentially drops a packet of a ow
that has either 1 had many packets dropped in the
past, or 2 a queue larger than the average queue size.
FRED has two variants, which we will call FRED-1
and FRED-2. The main di erence between the two
is that FRED-2 guarantees to each ow a minimum
number of bu ers. As a general rule, FRED-2 performs better than FRED-1 only when the number of
ows is large. In the following data, when we do not
distinguish between the two, we are quoting the results
from the version of FRED which performed better.
DRR De cit Round Robin - This algorithm 20 represents an e cient implementation of the well-known
weighted fair queueing WFQ discipline. The bu er
management scheme assumes that when the bu er is
full the packet from the longest queue is dropped. DRR
is the only one of the four to use a sophisticated perow queueing algorithm, and thus achieves the highest
degree of fairness.
These four algorithms represent four di erent levels of
complexity. DRR and FRED have to classify incoming ows,
whereas FIFO and RED do not. DRR in addition has to
implement its packet scheduling algorithm, whereas the rest all use rst-in- rst-out scheduling. CSFQ edge routers have
complexity comparable to FRED, and CSFQ core routers
have complexity comparable to RED.
We have examined the behavior of CSFQ under a variety of conditions. We use an assortment of tra c sources
mainly TCP sources and constant bit rate UDP sources,4
but also some on-o sources and topologies. Due to space
limitations, we only report on a small sampling of the simulations we have run.5 All simulations were performed in
ns-2 17 , which provide accurate packet-level implementation for various network protocols, such as TCP and RLM
15 Receiver-driven Layered Multicast, and various bu er
management and scheduling algorithms, such as RED and
DRR. All algorithms used in the simulation, except CSFQ
and FRED, were part of the standard ns-2 distribution.
Unless otherwise speci ed, we use the follo...
View Full Document
- Fall '10
- Eugene Ng