Performance

Historical snapshot (v1.4). The numbers and suite list below are frozen as captured for v1.4 and are kept for reference. The pipeline has since changed: the core-engine, full-cv, and scalability suites were retired, and current numbers come from the current-speed / comparative / stress pipeline plus the JMH suite. See docs/operations/benchmarks.md.

All numbers below were captured from scripts/run-benchmarks.ps1 — the full local benchmark workflow that built the test classpath once and ran current-speed, comparative, core-engine, full-cv, scalability, and stress suites in sequence. They were captured on a developer laptop; CI machines are typically 1.5–2× slower. The benchmark methodology, profiles, GC stabilisation, percentile rule, and how to compare two runs locally are documented in docs/operations/benchmarks.md.

End-to-end latency

current-speed full profile — 12 warmup + 40 measurement iterations.

Scenario	Avg ms	p50 ms	p95 ms	Docs/sec
engine-simple	3.00	2.73	4.86	333.83
invoice-template	17.74	17.44	25.13	56.38
cv-template	10.16	9.91	14.08	98.46
proposal-template	18.21	16.93	23.57	54.91
feature-rich	36.02	34.18	41.79	27.76

Per-stage breakdown (median ms per stage)

Scenario	Compose	Layout	Render	Total
invoice-template	0.33	2.55	5.76	8.63
cv-template	0.27	2.77	1.60	4.72
proposal-template	0.34	9.54	5.66	15.65

Render time is dominated by PDFBox serialization (36–67 % of total), so engine-side optimisations look smaller in the end-to-end avg than they do in the layout column. Page-background injection is a constant 1 fragment per page; column spans, layer stacks, and themes do not change the number of fragments emitted.

Parallel throughput

Invoice template, 12 docs per thread.

Threads	Total docs	Throughput	Avg doc ms
1	12	89.56/s	11.17
2	24	143.53/s	6.97
4	48	245.26/s	4.08
8	96	328.78/s	3.04

Near-linear scaling through 4 cores, ~2.7× throughput by 8 threads on a hyper-threaded CPU.

Linear scalability

scalability suite, simple docs.

Threads	Total docs	Throughput
1	100	807.41/s
2	200	1,960.75/s
4	400	3,839.64/s
8	800	7,394.56/s
16	1,600	11,164.76/s

13.8× throughput at 16 threads — the engine has no global synchronisation in the hot path.

Stress test

50-thread pool, 5,000 documents, single run:

Successful: 5000
Errors:     0
Time:       2499 ms

~2,000 docs/sec sustained under contention, zero failures.

Comparative benchmark

Simple invoice-class document, 100 measurement iterations.

Library	Avg ms	Avg heap MB	Notes
iText 5	1.57	0.16	low-level page primitives
GraphCompose v1.4	2.45	0.16	semantic DSL + pagination
JasperReports	4.45	0.19	XML-template based engine

GraphCompose sits between low-level PDF generators (iText 5) and template engines (JasperReports): close to iText latency on a per-doc basis while exposing a fully semantic Java DSL with deterministic snapshots.

Engine-only timings

The GraphComposeBenchmark and FullCvBenchmark mains below were retired after v1.4. Equivalent timings now come from the CurrentSpeedBenchmark engine-simple scenario and the JMH TemplateCvJmhBenchmark.

• GraphComposeBenchmark (engine-only, no PDF render): avg 1.04 ms, p50 0.97 ms, p95 1.64 ms.
• FullCvBenchmark (full CV template, including render): avg 4.14 ms, p50 3.80 ms, p95 6.37 ms.