Unformatted text preview: to have network
provide explicit per ow feedback to ingress nodes and have edge
nodes police the tra c on a per ow basis. We assume this is a too
heavyweight a mechanism for the Internet. CSFQ with this punishment extension may be seen as a
marriage of the allocation and identi cation approaches; the
di erence between 8 and our approach is largely one of
the relative importance of identi cation and allocation. We
start with allocation as fundamental, and then do identi cation only when necessary; 8 starts with identi cation, and
then considers allocation only in the context of managing
the bandwidth of identi ed ows. 5 Summary
This paper presents an architecture for achieving reasonably
fair bandwidth allocations while not requiring per- ow state
in core routers. Edge routers estimate ow rates and insert
them into the packet labels. Core routers merely perform
probabilistic dropping on input based on these labels and
an estimate of the fair share rate, the computation of which
requires only aggregate measurements. Packet labels are
rewritten by the core routers to re ect output rates, so this
approach can handle multihop situations.
We tested CSFQ, and several other algorithms, on a wide
variety of conditions. We nd that CSFQ achieve a signi cant degree of fairness in all of these circumstances. While
not matching the fairness benchmark of DRR, it is comparable or superior to FRED, and vastly better than the baseline
cases of RED and FIFO. We know of no other approach that
can achieve comparable levels of fairness without any perow operations in the core routers.
The main thrust of CSFQ is to use rate estimation at the
edge routers and packet labels to carry rate estimates to core
routers. The details of our proposal, such as the estimation
algorithms, are still very much the subject of active research.
However, the results of our initial experiments with a rather
untuned algorithm are quite encouraging.
One open question is the e ect of large latencies. The
logical extreme of the CSFQ approach would be to do rate
estimation at the entrance to the network at the customer ISP
boundary, and then consider everything else the core. This
introduces signi cant latencies between the point of estimation and the points of congestion; wh...
View Full Document
- Fall '10
- Eugene Ng