Speaker
Description
Operators expect DNS servers to respond within microseconds if all the data to answer a given query are locally available. Some BIND operators have reported suspicions that their production servers sometimes pause query responses.
When we attempted to reproduce this in a lab environment, we found that standard benchmarking tools like dnsperf, resperf, and flamethrower do not provide sufficient granularity for latency measurements.
In this talk, we present a new feature in dnsperf, which allows more fine-grained latency measurements, and we also present a new way to post-process dnsperf data into latency plots using DNS Shotgun toolchain.
Using these new features, we were able to measure latency spikes in BIND servers during server management operations.
We confirm that some operations can cause answer latency to spike, and we present recommendations for BIND operators.
