The deployment of Asynchronous Transfer Mode (ATM) networks is a
recent development in the field of computer communication. When we
attempt to use these networks as a part of the global Internet,
running the Internet Protocol (IP), we see a number of differences
between the data forwarding models of ATM (virtual circuits supporting
performance guar-antees) and IP (datagrams, usually best-effort).
In our research, we have evaluated different policies for IP-over-ATM
networks to bridge the gaps between these two networks and to make
them function more efficiently together.
The differences between IP and ATM raise three issues. First is the
question of how Internet applications can take advantage of ATM
quality of service facilities, without support from other portions
of the Internet. A second issue is that of determining which IP
conversations should be multiplexed onto a single ATM virtual
circuit. Last is the problem of virtual circuit management, which
determines when ATM connections should be established and torn down.
We have examined different quality of service, multiplexing, and
virtual circuit management policies, and evaluated their relative
merits from the standpoint of the performance of typical Internet
applications. Our evaluation used a simulation of a large IP
internetwork with a wide-area ATM backbone and a synthetic workload
modeling the traffic generated by common Internet applications. For
this purpose, we implemented a new network simulator, the Internet
Simulated ATM Networking Environment (INSANE).
Our results show that the use of different scheduling algorithms
and QOS parameters can be used to express preference for certain
applications, although some care must be taken to avoid starvation
effects. The use of jitter controlling schedulers in the ATM network
can be efficacious in reducing packet loss in long TCP bulk transfers.
We see that multiplexing can improve application performance due
to a reduced need to set up ATM virtual circuits, although interactions
with some network service disciplines can negate these effects.
Finally, we show that caching idle virtual circuits for reuse is,
in general, beneficial for both network and application performance.
|