Accuracy Evaluation of Application-Level Performance Measurements

Abstract

In many cases, application-level measurements can be the only way for an application to evaluate and adapt to the performance offered by the underlying networks. Applications perceive heterogeneous networking environments spanning over multiple administrative domains as "black boxes" being inaccessible for lower-level measurement instrumentation. However, application-level measurements can be inaccurate and differ significantly from the lower-level ones, amongst others due to the influence of the protocol stacks. In this paper we quantify and discuss such differences using the Distributed Passive Measurement Infrastructure (DPMI), with Measurement Points (MPs) instrumented with DAG 3.5E cards for the reference link-level measurements. We shed light on various impacts on timestamp accuracy of application-level measurements. Moreover, we quantify the accuracy of generating traffic with constant inter-packet-times (IPTs). The latter is essential for an accurate emulation of application-level streaming traffic and thus for obtaining realistic end-to-end performance measurements.