Evaluating Distributed Systems: Does Background Traffic Matter?

  • Kashi Vishwanath

Annual Technical Conference |

Published by USENIX

Evaluating novel networked protocols and services requires subjecting the target system to realistic Internet conditions. However, there is no common understanding of what is required to capture such realism. Conventional wisdom suggests that competing background traffic will influence service and protocol behavior. Once again however, there is no understanding of what aspects of background traffic are important and the extent to which services are sensitive to these characteristics. Earlier work shows that Internet traffic demonstrates significant burstiness at a range of time scales. Unfortunately, existing systems evaluations either do not consider background traffic or employ simple synthetic models, e.g., based on Poisson arrivals, that do not capture these burstiness properties. In this paper, we show that realistic background traffic, has qualitatively different impact on application and protocol behavior than simple traffic models. One conclusion from our work is that applications should be evaluated under a range of background traffic characteristics to determine the relative merits of applications and to understand behavior in corner cases of live deployment.