Resources

The cost of a run, captured beside it over its measured phase:

• each server JVM (processes, name to pid): its CPU time and resident memory every second (/proc on Linux, ps elsewhere), its garbage collector and heap from jstat -gc, one sampling process per JVM, and the bytes it allocated (Allocations), so the results give bytes per operation beside CPU;

• each container (containers): its CPU and memory from one docker stats stream;

• the stores' operation counters: MongoDB's serverStatus().opcounters (mongoContainer) and PostgreSQL's pg_stat_database of postgresDatabase with pg_stat_wal (postgresContainer), each from one session that prints them every second, and the end offsets of every Kafka topic (kafkaContainer).

Nothing starts inside the measured phase (B.O-2): prepare, at the warm-up, starts every sampling process, opens the allocation connections and reads the Kafka offsets once (kafka-get-offsets is a JVM of its own); start and stop only read what is already running, and complete, after the drain, reads the Kafka offsets again. The stores' counters at the phase's edges are the samples nearest to them; PostgreSQL's also give a per-second series of commits and WAL writes and flushes (B.O-3). The samplers are a few percent of a core; the results say what they were.