{ "_note": "Raw per-run output behind docs/efficiency.md. Produced by benchmarks/nexmark_compare/cloud/split-run.sh; not hand-edited. Each row is one measured run: one engine, one query, one trial, on the same engine node against the same isolated broker.", "measured_utc_date": "2026-07-26", "clink_commit": "1eb6fae", "flink_version": "2.2.0-scala_2.12-java21 (upstream image, unmodified)", "rig": { "engine_node": "Hetzner ccx23, 4 dedicated vCPU, AMD EPYC Milan, fsn1", "broker_node": "Hetzner ccx23, 4 dedicated vCPU, fsn1, runs the broker alone", "network": "private network between the two nodes", "note": "one engine at a time on the engine node; fresh cluster per measured run" }, "workload": { "source": "Kafka topic nx-bid, 4 partitions, 9.2M nexmark bid records", "queries": "nexmark q0 (stateless projection), q12 (windowed GROUP BY)", "parallelism": 4, "sink": "blackhole (output discarded, so the engine is measured and not a sink)" }, "how_read": { "cpu": "cgroup v2 cpu.stat usage_usec across every container of the engine under test, sampled before and after the drain", "memory": "cgroup v2 memory.stat anon (engine heap and stacks; excludes page cache), memory.peak for the high-water mark", "rate": "the engine's own records-processed counter polled at 100ms; sustained_slope is the steepest 0.5s slope, drain_rate the whole-run average" }, "correctness_gate": { "harness": "benchmarks/nexmark_compare/run.sh (separate run, parallelism 4, 500k events)", "q0_output_rows": { "clink": 460000, "flink": 460000, "identical": true }, "q12_output_rows": { "clink": 184767, "flink": 184767, "identical": true } }, "not_measured": [ "wall power, PUE, grid carbon intensity, embodied carbon", "workloads beyond nexmark q0 and q12", "hardware generations other than AMD EPYC Milan", "a tuning effort on either engine", "long-run (days/weeks) memory behaviour" ], "caveat_throughput_variance": "The same clink commit measured 1.56x and 1.30x against Flink on two different machines of this same type. events_per_cpu_sec and anon_mb held steady across those runs; the raw throughput ratio did not.", "caveat_q12_memory": "q12 memory is dominated by un-fired window state, which any engine computing that query must hold, not by engine overhead. Verified on the rig: jemalloc changed it not at all, disabling columnar execution RAISED it to 4663 MB, and it tracks records consumed then goes flat when ingest stops.", "runs": [ { "query": "q0", "engine": "clink", "trial": 1, "par": 4, "input_events": 9200000, "sustained_slope": 2977002.1, "drain_rate": 2791461.6, "drain_seconds": 3.296, "reached_target": true, "cpu_seconds": 13.04, "wall_seconds": 12.03, "cores": 1.08, "events_per_cpu_sec": 705521, "anon_mb": 74.1, "peak_mb": 86.8, "mem_fresh_stack": true }, { "query": "q0", "engine": "clink", "trial": 2, "par": 4, "input_events": 9200000, "sustained_slope": 2937084.7, "drain_rate": 2730722.3, "drain_seconds": 3.369, "reached_target": true, "cpu_seconds": 12.96, "wall_seconds": 11.08, "cores": 1.17, "events_per_cpu_sec": 709877, "anon_mb": 69.9, "peak_mb": 85.5, "mem_fresh_stack": true }, { "query": "q0", "engine": "flink", "trial": 1, "par": 4, "input_events": 9200000, "sustained_slope": 2088165.4, "drain_rate": 1021201.0, "drain_seconds": 9.009, "reached_target": true, "cpu_seconds": 39.45, "wall_seconds": 15.27, "cores": 2.58, "events_per_cpu_sec": 233207, "anon_mb": 1146.4, "peak_mb": 1513.3, "mem_fresh_stack": true }, { "query": "q0", "engine": "flink", "trial": 2, "par": 4, "input_events": 9200000, "sustained_slope": 2289062.0, "drain_rate": 1057957.7, "drain_seconds": 8.696, "reached_target": true, "cpu_seconds": 40.35, "wall_seconds": 14.78, "cores": 2.73, "events_per_cpu_sec": 228005, "anon_mb": 1160.1, "peak_mb": 1537.9, "mem_fresh_stack": true }, { "query": "q12", "engine": "clink", "trial": 1, "par": 4, "input_events": 9200000, "sustained_slope": 1872541.7, "drain_rate": 1587131.9, "drain_seconds": 5.797, "reached_target": true, "cpu_seconds": 29.95, "wall_seconds": 13.84, "cores": 2.16, "events_per_cpu_sec": 307179, "anon_mb": 1230.0, "peak_mb": 1275.0, "mem_fresh_stack": true }, { "query": "q12", "engine": "clink", "trial": 2, "par": 4, "input_events": 9200000, "sustained_slope": 1837941.0, "drain_rate": 1536769.6, "drain_seconds": 5.987, "reached_target": true, "cpu_seconds": 29.6, "wall_seconds": 13.67, "cores": 2.17, "events_per_cpu_sec": 310811, "anon_mb": 1361.5, "peak_mb": 1378.1, "mem_fresh_stack": true }, { "query": "q12", "engine": "flink", "trial": 1, "par": 4, "input_events": 9200000, "sustained_slope": 1601015.3, "drain_rate": 702772.9, "drain_seconds": 13.091, "reached_target": true, "cpu_seconds": 59.32, "wall_seconds": 19.64, "cores": 3.02, "events_per_cpu_sec": 155091, "anon_mb": 1555.2, "peak_mb": 1850.5, "mem_fresh_stack": true }, { "query": "q12", "engine": "flink", "trial": 2, "par": 4, "input_events": 9200000, "sustained_slope": 1585431.9, "drain_rate": 707311.4, "drain_seconds": 13.007, "reached_target": true, "cpu_seconds": 59.39, "wall_seconds": 19.85, "cores": 2.99, "events_per_cpu_sec": 154908, "anon_mb": 1504.8, "peak_mb": 1853.9, "mem_fresh_stack": true } ], "derived_ratios": { "q0": { "clink_events_per_cpu_sec": 709877, "flink_events_per_cpu_sec": 233207, "cpu_efficiency_ratio": 3.04, "clink_anon_mb": 69.9, "flink_anon_mb": 1146.4, "memory_ratio": 16.4 }, "q12": { "clink_events_per_cpu_sec": 310811, "flink_events_per_cpu_sec": 155091, "cpu_efficiency_ratio": 2.0, "clink_anon_mb": 1230.0, "flink_anon_mb": 1504.8, "memory_ratio": 1.2 } } }