Using Measurements to Validate Simulation Models of TCP/IP over High Speed ATM Wide Area Networks
Overview of TCP/IP over ATM
1 Importance of TCP/IP over ATM
Probably the most succesful idea in data networking over the past twenty years
has been the concept of internetworking [13]. It is a
method for interconnecting networks, regardless of the particular networking
technology (Ethernet, ATM, HIPPI, Frame Relay, FDDI), used by the individual
systems. What makes internetworking possible is the development of
protocols like TCP/IP. The TCP/IP protocol
suite is the internetworking protocol used on the Internet, a global collection
of networks connecting millions of computers and users, and incorporating a
large variety of different network technologies [13]. It allows
computers of all sizes, from many different computer vendors, running
totally different operating systems, to communicate with each other [17].
ATM technology is the emerging standard adopted by telecommunications and
computer vendors for high speed networks. The fast-cell switching technology
employed by ATM helps to provide scalable (in size and speed)
networks [5]. Further, ATM technology
provides bandwidth on demand and enables the integration of
real-time and data traffic over the same physical medium for
wide area networks.
Although TCP/IP and ATM often have been viewed as competitors, their
complementary strengths and limitations form a natural alliance that combines
the best aspects of both technologies [14]. In the near future, a
large portion of the traffic carried by the ATM networks will be generated
by applications written to run over a TCP/IP protocol stack [11].
In fact, many of the existing ATM networks employ TCP/IP over ATM
technology.
2 Previous Studies of TCP/IP over ATM
Numerous simulation and experimental studies have been performed in order to
predict the performance of TCP/IP over ATM under congestion and buffer overflow
conditions that arise from bandwidth mismatches or multiple sources contending
for the same link [2, 5, 8, 11, 15, 16, 19].
The simulation study in [16] examines the performance of
TCP over ATM under conditions of network congestion. The results show that the
TCP/IP over ATM performance is poor when there is congestion caused by
small switch buffers and large TCP segment and window sizes.
This performance degradation is caused by a loss-rate multiplier
effect caused by the switch dropping cells from multiple packets.
In [15], the performance of TCP connections over ATM networks without
ATM-level congestion control is investigated. Simulation results of congested
networks show that the effective throughput of TCP over ATM can be quite low
when cells are dropped at the congested ATM switch. To improve the performance,
a mechanism called early packet discard (EPD) which brings throughput
performance to its optimal level is proposed.
The work in [11] considers some undesirable interactions between the
congestion control scheme used in TCP and the policing mechanisms used in
ATM networks that can significantly degrade the throughput of TCP traffic. It
is shown that in the presence of policing, once a TCP connection has increased
its window size beyond the sustainable cell rate (SCR) times the round-trip
time and if the bottleneck capacity exceeds the SCR, the buffer at the
site providing the
policing mechanism fills up quickly and most of the packets in the TCP
window are dropped. This causes the average throughput to be significantly
lower than SCR value. In order to improve the performance,
the use of smarter policing or cell-level traffic shaping schemes
is suggested.
The work in [2,12] describes performance measurements taken
from the MAGIC gigabit testbed relating to the performance of TCP in wide area
ATM networks. Results show that the TCP rate control mechanism alone is
inadequate for congestion avoidance and control in wide area gigabit
networks. It is also illustrated that TCP augmented by cell-level
pacing allows the full link capacity to be utilized. Cell level
pacing is necessary because the TCP rate control mechanism does not
control traffic burstiness sufficiently to avoid congestion-induced
cell losses in wide area networks.
These studies separately present simulation and measurement results.
However, comparison of performance predictions from
measurement techniques and simulation models of TCP/IP-ATM networks is
needed to refine our understanding of the network operation.
Validation of models against measurements is also required so simulation models
can be used with confidence to capture the effects of network control,
and to accurately characterize the performance of ATM WANs.
However, in order to achieve that goal, the precision level of
the models (which is directly proportional to the model
execution time) needs to be determined because of the computational
complexity of the simulation.
Return to
