Every Packet Counts: Fine-Grained Delay and Loss Measurement with Reordering

Abstract

Delay is an important metric to understand and improve system performance. While existing approaches focus on aggregate delay statistics in pre-programmed granularity, providing only statistical results such as averages and deviations, those approaches fail to provide fine-grained delay measurement at a flexible level and thus may miss important delay characteristics. For example, delay anomalies, which are critical system performance indicators, may not be captured by existing coarse grained approaches. In this work, we propose a fine-grained delay measurement approach based on a new measurement structure design called order preserving aggregator (OPA). OPA can efficiently encode the ordering and loss information by exploiting inherent data characteristics. Based on OPA, we propose a two layer design to convey both ordering and time stamp information, and then derive per-packet delay/loss measurement with a small overhead. We evaluate our approach both analytically and experimentally with widely used real-world data sets. The results show that our approach can achieve accurate per-packet delay measurement with an average of per-packet relative error at 2%, and an average of aggregated relative error at 10-5, while introducing less than 4 × 10-4 additional overhead.