Contributor · How-to

Measure memory and speed on a Pi#

The server’s budget is a Raspberry Pi 1 (The shape of the system). Two scripts check a change against it.

The footprint: memory and threads#

make footprint

Builds a ReleaseSafe server and runs scripts/footprint.py: it starts the server, holds idle connections, and reads the process’s memory; then it seeds a library at two sizes and runs a backup beside the serving server. It fails when a number crosses its ceiling. CI runs it on the x86_64 musl binary that ships (make footprint SERVER=path measures a binary already built). Ceilings are raised only with a note in the script saying what bought the bytes.

The numbers are measured on x86_64 as a stand-in for the Pi: the structure they pin (threads don’t grow with connections) is the same there.

A household’s evening: load#

scripts/household_load.py plays a household against a server: people browsing, opening and searching, and tablets syncing.

scripts/household_load.py zig-out/bin/inkhold --seconds 120 --people 4

To measure the Pi itself, run the server on it and the load from another machine, so the load generator doesn’t share the Pi’s one core:

scripts/household_load.py --ssh pi --remote-binary inkhold-arm --seconds 300

On the devlab’s Pi 1 B+, the server runs as a job (devlab-target), which answers only the NUC, so the load comes from the NUC’s scratch/inkhold-builder instance. From the laptop, with a release’s armhf archive:

scripts/pi-load.sh inkhold-arm-linux-musleabihf.tar.gz --seconds 300

It submits the archive and a config as the job, signs in with the job’s first-run password, and seeds the libraries from the builder. The evening starts once indexing is done, read from the job’s CPU time (under 5% of the core for 3 s), and the script ends with the job’s own accounting: its peak memory (the unpacked archive included, since on the Pi’s read-only root files are RAM) and its CPU time.

--think 0 removes the pause between clicks (a stress test); --tls CERT KEY serves HTTPS, and --suite limits the cipher suites, to compare them.

Measured on a Pi 1 B+ in a household evening (four people, two tablets): about 15% of its core and 5 MB of anonymous memory, a folder listing in 68 ms at the median. A change worth measuring is one that should move these.