As the Internet develops support for a wider variety of applications and offered service qualities, tracking the composition of network traffic will be important to both meet current demand as well as evolve toward future applications. In particular, as many new applications consume considerably more bandwidth, and for longer continuous periods of time, than those with which the Internet has fundamentally grown to where it is today, it will be important to measure traffic in order to determine the extent and range of different user requirements.
We thus consider it critical to ensure that the follow-on NSFNET activities, including the NAPs, recognize the need for and in fact expand on the operational statistics collection paradigm during the lifetime of the NSFNET. The architecture will likely alter requirements for earlier NSFNET statistics collection techniques and requirements. In particular we think it will be important to foster a multi-agency effort to support the aggregation of network statistics data from multiple service providers.
In pursuit of this agenda we have proposed to deploy operational software for characterizing Internet flows at the four NSF-sponsored network access points (NAPs) which were planned to be operational as of 1 November 1994. The Sprint NAP was operational at that time; it was not clear as of February 95 when the others would be operational. We also suggested other strategic locations such as the FIXes and their successors, but required appropriate equipment that could keep up with the traffic rate in order to perform accurate workload characterization. NSF has specified that they want statistics collected at the NAPs, but without very clear outline of deliverables, and we are working with both NSF and the NAPs to clarify the objectives.
Claffy visited MCI in Reston, Virginia at the end of December 1994 to meet with the MCI engineers responsible for statistics collection and general operation of the vBNS (Rick Wilder, John Jamison, Joe Lawrence, and Dennis Ferguson). MCI will provide each vBNS site with a DEC Alpha 900 server to support vBNS statistics, and delivered one to SDSC in December 1994 for preliminary testing and development of statistics tools.
We discussed MCI's plans for collecting statistics for InternetMCI, which uses a model similar to that of the NSFNET. MCI intends to extend this same architecture to the vBNS to the extent possible, but it is not yet clear how to deal with statistics on the newer switches that are not so conducive to statistics collection. We have made available to MCI tools that we have developed and used for the last three years to process and analyze NSFNET data, including SNMP and ARTS, as well as tools to analyze dedicated packet traces to perform flow characterization.
We hope to continue to collaborate, and offer MCI and vBNS users counsel and support for traffic characterization on the vBNS service when it becomes operational, both for individual vBNS flows as well as the aggregate vBNS workload. Other vBNS sites also plan to contribute to the determination of what vBNS service providers should track regularly, and to coordinate placement of these statistics into a database to facilities studies similar to those of supercomputer usage, e.g., workload burstiness characterization (see sections 2.3 and 10). An important objective is to ensure that metrics and models derived from them are well documented so they can become applicable for the Internet community.