As the NSFNET service provision agreement ceases and Internet connectivity becomes an industrial commodity, metrics for describing the quality of connectivity will be important to market efficiency. One of our goals is to determine how to describe Internet workload using metrics that will enable customers and service providers to agree on a definition of a given grade of service. One common notion of network workload is `burstiness', but there is not yet agreement in the Internet community on the best metric to define burstiness. Network behavioral patterns of burstiness are important for defining service specifications. As such, finding metrics to specify expectations and define observations are critical for the evaluation of services, and the need for them intensifies as services in today's internetworking environments become more commercialized, rather than procured via collaborative undertakings between the federal government and academia and industry.
Several researchers [13,14,15] have explored the failure of Poisson models to adequately characterize both local and wide area Internet traffic. Based on multiple packet traces, we started some work on traffic self similarity considerations, including variance-time plots to compare them to those in the above studies. We have also explored alternative metrics, e.g, peak/average ratio plots, and are still investigating possible metrics that may serve NSF, or Internet service providers, most effectively.
This task relies critically on accurate packet arrival timestamps, and thus has required finding tools adequate for packet tracing at accurate (microsecond) time granularities. We were previously using SGI R4000s for packet collection; however when the traffic level grew beyond loads with which the SGI's could keep up, i.e., at NCSA or even at SDSC during current busy periods, we investigated possible alternative architectures. Jeff Mogul of DEC assisted us in porting our SGI code to the DEC Alpha workstations, which have performed considerably better. (In our measurements, our SGI platforms delivered FDDI packets to the application up to a rate of approximately 3500 packets per second, after which the CPU became the bottleneck for the application to get the data from the kernel.
From our initial measurements and discussions with Jeff Mogul, a higher end DEC Alpha platform will deliver at least 3 to 4 times as many packets, but we have not yet done experiments to conclusively verify the performance limits.)