[zeenix]

The README's numbers were measured with dbus-bench's client on zbus 4.3, which was so slow that, with 20 clients, it held every broker but dbus-daemon to about the same timings. This takes them from the run of its client on zbus 6 instead, on the same laptop and with as many rounds, which zeenix/dbus-bench#3 adds as results/2026-10-20-zbus-6-ryzen-7840u:

Broker	1 client	20 clients	CPU time with 20 clients	Peak memory
busd	0.088 → 0.070 s	0.926 → 0.475 s	0.42 → 0.37 s	6.3 → 6.4 MiB
dbus-broker 37	0.082 → 0.062 s	0.930 → 0.451 s	0.41 → 0.31 s	7.9 → 8.1 MiB
dbus-daemon 1.16.2	0.160 → 0.142 s	1.185 → 1.195 s	1.14 → 1.17 s	6.0 → 6.1 MiB

With the faster client, dbus-broker comes out ahead of busd: busd takes 5 to 13% longer, and spends 19% more CPU time with 20 clients, but still needs 21% less memory. So the README no longer says that busd keeps up with dbus-broker, but that it's close behind it, with less memory.

• The numbers of busd are those of a982812, the main branch before 🐛 Queue messages for each client rather than wait on it #376 gave each client a queue of its own. On a 4-vCPU VM, the queues cost 2.1% more CPU time with 20 clients and 1.5% more memory, and nothing measurable in the timings.
• The values are the medians in the run's results.md: single_s, multi_s, multi_broker_cpu_s and broker_maxrss_kib.
• The README links to the results directory of dbus-bench's main branch, which gets this run once Refresh the results on zbus 6, and add a look into a busd stall zeenix/dbus-bench#3 is merged.

Generated by Claude Opus 5.5 (claude-opus-5-5).

🤖 Generated with Claude Code