Theory and practice of non-intrusive active network measurements

摘要: Today's data networks are highly distributed and enormous in scale. The ability to measure them is vital both network operators end-users. Network measurement methods can broadly be classified into passive that rely on collected at routers, active based observations of actively-injected probe packets. Active measurements, the focus this dissertation, attractive end-users who, under current architecture, cannot access any routers. use measurements because they easy conduct, have low overhead and, contrast collection methods, exactly what normal packets experience. most significant disadvantage limited accuracy has typically been achievable using them. One main reasons for need non-intrusive, thus leaving measured systems uninfluenced by observation, fundamentally affecting accuracy. In we rigorous theoretical analysis understand impact non-intrusiveness investigate how theory translates practice. Our investigation consists three parts. In first, sampling-related issues, i.e., when do send why? starting point conventional wisdom says "Poisson Arrivals See Time Averages (PASTA)" principle implies Poisson probing. We show PASTA does not imply probing optimal it ignores bias caused intrusiveness estimation variance. Using simulations, motivate rare probing, preferably so-called mixing epochs, as a sound practical strategy. second part, if observed delays (non-intrusive) pairs used estimate cross-traffic properties single-hop case. inability prior works [SKK03] without hard-to-achieve timing control. derive estimated, theory, that, well-motivated assumption, non-intrusive entire distribution an intra-pair interval. third part motivated apparent difficulty designing robust multi-hop queueing effects; experience first-hand with our estimators. novel hop-dependent priority primitives design Measurement-Friendly Networks (MFNs), which accurate effects, performed. only simplify management tasks but also easily deployable. exploring MFNs, find nonpreemption persistence cause unavoidable inaccuracies represent, sense, fundamental limitations measurements.