Networking notebook

Look beyond the average.

Latency changes with distance, congestion and the amount of work a service performs. A median can describe a typical request while hiding the slower requests that people actually notice.

Keep the conditions comparable

Use the same destination, request size and time window. Record the range of results, not only the fastest sample. Separate connection setup from the time needed to deliver the response.

Measure before adding capacity

When a path is congested, a faster processor may not help. When the service is CPU-bound, extra bandwidth may have little effect. A small controlled comparison is often more useful than a large configuration change.