- RED gateways offer a practical approach to congestion avoidance, and
appear to be effective in Floyd and Jacobson’s tests. RED gateways are a
practical solution because they are easily implemented, require no changes
to existing protocols on the hosts, and do not violate protocol layering.
RED gateways should work well with TCP, which incrementally shrinks its
sending window size when it sees dropped packets, but should theoretically
also work well to proportionally control the bandwidth consumed by an ill
behaved host.
- The fact that RED gateways retain no state information makes them a more
realistic solution to congestion avoidance than Fair Queuing. Fair Queuing
implementations require that the router keep track of connection
information (host-destination pairs), so that it can fairly send off
packets in a round robin fashion. RED gateways just select packets to drop
or mark at random when the average queue size is getting too large. The
paper does suggest, however, that to control an ill behaved host, RED would
need to keep track of how many often packets from a certain connection were
marked.
- RED gateways sound good in theory, but I did not feel that the paper
provided enough evidence that they will work as well in the real world. It
would be interesting to see results from tests of running different kinds
of applications sending packets through the same router. If telnet and
streaming audio applications are competing for buffer space, probability
suggests that the streaming audio packets are more likely to get marked, so
that application will back off and allow telnet to get packets into the
queue. In reality, the packets picked at random could always end up being
the telnet’s packets, which would kill telnet’s throughput and give all of
the bandwidth to the streaming audio. Plus, dropping packets will only
make a real difference in avoiding congestion if the hosts react as TCP
does. If the streaming audio does not care if it is delivered reliably, it
may continue to hog the bandwidth, and telnet will never gain much throughput.
- Even with the above criticism, RED gateways certainly do a better job of
efficiently using the full bandwidth than other schemes. Fair Queuing may
waste a lot of bandwidth by allocating equal sized portions to applications
that send infrequent bursts of packets. Drop Tail gateways notify every
sender at the same time of queue overflow, causing synchronization of
senders, which leads to under utilization of bandwidth while all of the TCP
senders back off.