Field notes: measuring jitter on a shared research backbone

We spent the last quarter instrumenting every hop of the cooperative backbone with hardware timestamping probes. The goal was modest: understand where jitter actually comes from, rather than guessing.

Three findings stood out.

First, most jitter on a research network is not caused by congestion but by queueing in transit devices that were never designed for science traffic. Second, microbursts below one millisecond are invisible to five-minute interface counters and can persist for weeks. Third, the largest single improvement came from something unglamorous — tidying up equal-cost multipath hashing so that a single large flow stops flapping between links.

We published the raw probe data, the collection configuration and the analysis notebooks. Members are free to reuse the methodology on their own campus fabrics.

The next phase adds one-way delay measurement between all edge sites, which should let us localise asymmetry to a specific fibre span.