feat(bench): precompute-engine throughput vs CPU microbenchmark

[milindsrivastava1997]

Summary

Microbenchmark of the standalone precompute_engine: throughput (samples/s) vs engine CPU for several StreamingConfigs and input cardinalities, for the paper's microbenchmark section. No Prometheus in the loop.

• bench_ingest_sender (new bin): remote-write load generator run as a separate process on separate cores. Pre-encodes bodies off the send path, 16 senders each owning a disjoint slice of series (one request in flight, so per-series order holds), lockstepped so per-group watermarks don't drop a lagging sender's samples, optional target rate. Reports acked samples/s, engine + sender CPU (from /proc), engine peak RSS.
• precompute_engine --num-aggregates-to-retain: without an inference config the standalone binary retained every window, so memory grew without bound (34 GB for KLL at 100K series in a 45 s run).
• asap-tools/experiments/microbench/precompute_throughput/: pipeline configs (KLL, DDSketch, CMS+Heap, HydraKLL), sweep.py (main / fanin / window / scaling experiments), plot.py, results from a CloudLab run, and HANDOFF.md with the exact run procedure and why each setting is what it is.

Results (8 cores, saturation, M samples/s)

Pipeline	1K	10K	100K	1M
KLL	1.40	1.36	0.71	0.89
DDSketch	1.24	1.27	0.57	0.83
CMS+Heap	1.93	2.44	2.33	2.31
HydraKLL	0.64	1.73	1.94	2.12

Keyed sketches use all 8 cores and are flat in cardinality; per-series sketches saturate at ~5 of 8 cores at 100K+ series (not yet profiled; suspect the store's one RwLock per aggregation id). Caveats are listed in HANDOFF.md.

Test plan

• Built and clippy-clean (bench_ingest_sender, precompute_engine); pre-commit (fmt/check/clippy/test) passed
• Full sweep ran on CloudLab node5; sanity checks hold (sender CPU well under its cores, rate-limited points hit target)
• Repeat runs (median of 3) before using numbers in the paper

🤖 Generated with Claude Code