TKS Ephemeris · developer · performance

Performance Whitepaper & Swiss-Ephemeris Comparison

A reproducible benchmark of the TKS Ephemeris (JPL DE440) against the official Swiss Ephemeris C library, and an end-to-end profile of the Jyotiṣa page. Every figure below is reproducible from §5; measurements are best-of-three full runs on the machine in §1.

Headline. For the 11-body geocentric apparent-longitude workload (10,000 epochs), TKS is 1.31× faster than Swiss Ephemeris in file mode — 0.169 s vs 0.222 s — and the whole Jyotiṣa chart (25 computations) renders in a median 12.5 ms.

1 · Environment

CPU	Apple M3 Max — 14 cores (10 performance + 4 efficiency)
Memory	36 GB unified · Model Mac15,10
OS	macOS 26.6.2 (Darwin 25.6.0, arm64 / T6031)
Toolchain	rustc 1.98.0 · `--release`
TKS engine	JPL DE440 (Chebyshev) · true Chitrāpakṣa
Swiss Ephemeris	pyswisseph 2.10.03 · real `.se1` files (1800–2400), file mode (verified via returned flags, not the Moshier fallback)

2 · Ephemeris throughput — 10,000 × 11 = 110,000 positions

Bodies: Sun · Moon · Mercury · Venus · Mars · Jupiter · Saturn · Uranus · Neptune · Pluto · True-Node. Geocentric apparent tropical longitude, 1-day spacing.

Engine	Total (best-of-3)	per 11-body set	per position	vs Swiss
TKS Ephemeris (DE440)	0.169 s	0.0154 ms	1.54 µs	1.31× faster
Swiss Ephemeris 2.10.03 (files)	0.222 s	0.0202 ms	2.02 µs	1.00×
TKS per-call API¹	0.267 s	0.0243 ms	2.43 µs	1.20× slower
Swiss (Moshier fallback)	1.260 s	0.1146 ms	11.45 µs	—

¹ The public helper apparent_direction() fetches the Earth state per call; the production pipeline shares Earth/Sun per epoch and uses two light-time iterations — the TKS Ephemeris row. Target check: both engines deliver 10,000 × 11 positions well under 3 s. Correctness: true Rāhu = Gemini 10.78° in both (agree < 0.01°).

3 · Jyotiṣa page — end-to-end (25 computations)

The /tks-jyotish page issues 25 parallel computations (vara · chalit · 16 vargas · upagraha · Vimśottarī 3 levels · ṣaḍbala · aṣṭakavarga · maitrī · jaimini · doṣa · rāja/nābhasa-yoga · chara/nārāyaṇa/sudarśana dāśās · yoga-samuh · aṣṭottarī · gochara · kālacakra). Measured over 1,000 page loads (all 25 issued concurrently per load).

Metric	min	p50	p90	p99	max	mean	throughput
Full page (25 endpoints)	10.65 ms	12.45 ms	13.35	17.21	24.65	12.59 ms	79.4 /s

Per-endpoint (warm): sudarśana 3.7 ms · dashas 6.3 ms · graha-sphuṭa 4.0 ms · ṣaḍbala 3.3 ms · bhāva-chalit 3.1 ms · divisionals 4.6 ms. A light endpoint (maitrī) measures 2.65 ms — the HTTP setup floor on this host.

4 · Concurrency

endpoint	req/s	time/req (concurrent)	failed
sudarśana (`ab -n 1000 -c 10`)	1,646	0.607 ms	0
maitrī	2,364	0.423 ms	0

5 · Reproduction

# 1. Ephemeris, 10,000 x 11 (TKS)
cargo build --release --example bench_11 -p tks-ephemeris
./target/release/examples/bench_11

# 2. Swiss Ephemeris (real .se1 files), same 10,000 x 11
TKS_SWISS_EPHE=/path/to/ephe  /tmp/swevenv/bin/python scripts/bench_swiss.py

# 3. Full jyotisa page (25 endpoints, 1000 loads) + concurrency
./target/release/panchanga --serve --port 8099 &
python3 scripts/bench_page.py
ab -n 1000 -c 10 "http://127.0.0.1:8099/v1/vedic/sudarshana?engine=tks&date=2001-08-27&time=06:15&place=Dehradun&lat=30.3165&lon=78.0322&tz=5.5"

6 · Methodology

Warm-up: 30–50 iterations discarded; kernel pages and CPU caches resident.
Iterations: 10,000 × 11 for ephemeris; 1,000 for the page and concurrency.
Statistics: full latency distribution (min·p50·p90·p99·max·mean) for the page; best-of-three for the ephemeris micro-runs (variance < 3 %).
Fairness: identical bodies, epoch grid and iteration count; Swiss in file mode; the engine comparison counts the corrections each engine performs (light-time, aberration, solar deflection, true node).
Scope: server rows include HTTP/TCP setup (HTTP/1.0, Connection: close); ephemeris rows measure pure computation.