4.5.3 Traffic Measurement/Characterization

From a network technology point of view, the presence of real application traffic in the testbeds had two major benefits: it allowed network problems to be discovered during the debugging stage which would not otherwise have been found, and it allowed measurement of the traffic properties which result when real high speed networks are used.

Essentially all of the testbeds experienced problems when real traffic was used which did not occur with artificial traffic sources. Many of the problems involved equipment design subtleties, for example timing effects due to traffic burstiness. The use of real applications was especially important for this, since the traffic characteristics seen by the network are shaped by the interaction of the application and the network, and so for new contexts such as gigabit networks the traffic cannot be realistically generated a priori using artificial means.

One of the research goals of the Vistanet testbed was an investigation of traffic characteristics using its dynamic radiation therapy planning application. To allow traffic measurements at the testbed's 622 Mbps OC-12c SONET link rates, a special control and monitoring subsystem (CMS) was developed by BellSouth as part of each Network Terminal Adapter (NTA) used to interface the testbed's HIPPI hosts to the ATM/SONET network. The CMS allowed information about ATM cells flowing through an NTA to be captured in real-time and saved for subsequent analysis.

The CMS was used by Vistanet researchers to capture and analyze radiation dose traffic generated under a number of different usage conditions, with data captured both at the transmitting NTA connected to the Cray source at MCNC and at the receiving NTA collocated with the Pixel Planes 5 destination on the UNC campus. For the testbed experiments, flow control was used between NTAs communicating across the network and between each NTA and its local HIPPI host to prevent overflowing NTA buffers, but no switch-NTA flow control or traffic shaping was applied to the ATM cells entering the network.

The resulting histograms showed the application's ATM cell traffic to be highly bursty -- mean link utilization ranged from only 0.27% to 1.2%, with each new dose transfer lasting for 40 milliseconds at an aggregate data rate of 80-90% of the available ATM cell rate (the latter was approximately 600 Mbps after subtracting out SONET overhead on the 622 Mbps link). Inspection of detailed histograms of dose bursts showed an oscillatory pattern of peaks and valleys, attributed to the end-to-end HIPPI flow control used in the testbed and possibly also due to Cray computation effects.

The different mean utilizations reflect the interburst idle times measured for different usage conditions. When an experienced user was using the system for therapy planning, idle times were approximately 2 seconds long and fairly periodic, reflecting user think time before initiating a new dose calculation and its resulting 40 millisecond burst transfer. At other times interburst idle times were as long as 28 seconds and highly variable, and were associated with demonstrations of the system.

Simulations were carried out using the measured data and a model of the testbed's Fetex 150 ATM switch to investigate the impacts of the traffic burstiness on cell loss. The results showed that a high peak cell loss rate occurred in the switch for highly bursty traffic, and that increasing the amount of cell buffering in the switch provided only a relatively small improvement over a wide range of buffer sizes. The conclusion drawn from these and other results was that either traffic shaping must be used for the ATM cell traffic, or a near peak-rate virtual circuit bandwidth allocation must be used to avoid significant cell loss rates.

As part of their Vistanet work, MCNC researchers developed the HIPPI Link Data Analyzer (HILDA) to capture 800 Mbps HIPPI traffic statistics. Configured as a single VME-bus board for operation on a Sun4 or SGI workstation, HILDA could also serve as a continuous 800 Mbps traffic source for network testing and as a standard HIPPI host interface (with data rates constrained in the latter mode by the VME bus). For data capture, the HILDA board provided a passthrough connection for a HIPPI link, analyzing and displaying the resulting traffic statistics on its host workstation or on a remotely located X-windows node. It also provided a capability for programmable error insertion onto a HIPPI link.

HILDA was used to capture application traffic data on local HIPPI networks and for the evaluation of transport protocols operating over HIPPI links, with analysis of the resulting data carried out in collaboration with NCSU researchers. More generally, HILDA proved invaluable for testing and problem diagnosis of Vistanet HIPPI/ATM/SONET facilities, and was also used for these purposes by other testbeds. A technology transfer of HILDA took place during 1992, when MCNC entered into an agreement with the Avaika Corporation to make HILDA available commercially.