Skip to content

parse_speed.md

Latest commit

 

History

History
13 lines (7 loc) · 3.18 KB

parse_speed.md

File metadata and controls

13 lines (7 loc) · 3.18 KB
Raw

Parse Speed

Parse speed varies significantly by parse mode, dataset characteristics, and hardware/compiler settings. The results below reflect reported, reproducible measurements and should not be compared without matching mode, corpus, and platform.

  • Boost.JSON: On mixed/escaped corpora, reported DOM throughput ranges from roughly 567–1120 MB/s depending on content and options [EVID:EVID-20250824-ParseSpeed-Boost-004][EVID:EVID-20250824-ParseSpeed-Boost-005]. Numeric-only arrays can exceed 1 GB/s (e.g., 1105 MB/s on a 147 KB numeric array with a monotonic memory resource) [EVID:EVID-20250824-ParseSpeed-Boost-002]. A specialized strings-only case (long unescaped strings) shows ~6000 MB/s on historical hardware/docs and should be treated as an outlier scenario rather than typical mixed JSON [EVID:EVID-20250824-ParseSpeed-Boost-003]. Using a monotonic memory resource and parser reuse materially improves throughput (e.g., 740 MB/s vs 290 MB/s on the same corpus) [EVID:EVID-20250824-ParseSpeed-Boost-001].

  • simdjson: Single-core DOM parsing has been reported at up to ~3.0 GB/s on Skylake-era CPUs [EVID:EVID-20250824-ParseSpeed-simdjson-001][EVID:EVID-20250824-ParseSpeed-simdjson-002]. On-Demand mode (lazy parsing without full DOM materialization) reaches ~6.8–7.0 GB/s on Sapphire Rapids for tweet-like corpora [EVID:EVID-20250824-ParseSpeed-simdjson-004][EVID:EVID-20250824-ParseSpeed-simdjson-006]. For NDJSON, document-stream parsing can reach ~3.5 GB/s using multi-core pipelines [EVID:EVID-20250824-ParseSpeed-simdjson-003]. Comparable DOM figures on the same Sapphire Rapids platform are around ~4.8–4.9 GB/s for the tested datasets/queries [EVID:EVID-20250824-ParseSpeed-simdjson-005][EVID:EVID-20250824-ParseSpeed-simdjson-007].

  • nlohmann/json: Reported DOM throughput on small-to-medium corpora typically falls in the tens to low hundreds of MB/s (e.g., ~55–137 MB/s across canada/random/twitter), with some reports around ~0.1 GB/s on twitter.json under different platforms [EVID:EVID-20250824-ParseSpeed-nlohmann-002][EVID:EVID-20250824-ParseSpeed-nlohmann-003][EVID:EVID-20250824-ParseSpeed-nlohmann-004][EVID:EVID-20250824-ParseSpeed-nlohmann-005][EVID:EVID-20250824-ParseSpeed-nlohmann-006][EVID:EVID-20250824-ParseSpeed-nlohmann-001].

  • Glaze: Author-reported microbenchmarks on an Apple M1 (single thread) show ~1.08–1.22 GB/s on a 670 B custom JSON and related tests, with consistent results over several dates/commits; these are in-memory struct parses on tiny payloads and should be interpreted accordingly [EVID:EVID-20250824-ParseSpeed-Glaze-001][EVID:EVID-20250824-ParseSpeed-Glaze-002][EVID:EVID-20250824-ParseSpeed-Glaze-003][EVID:EVID-20250824-ParseSpeed-Glaze-004][EVID:EVID-20250824-ParseSpeed-Glaze-005][EVID:EVID-20250824-ParseSpeed-Glaze-006][EVID:EVID-20250824-ParseSpeed-Glaze-007].

Key caveats: Cross-library comparisons require parity in mode (DOM vs On-Demand vs SAX/NDJSON), dataset structure (numeric-heavy vs escaped strings vs mixed), and hardware/compiler configuration. Treat mode/dataset mismatches as non-comparable without appropriate normalization [EVID:EVID-20250824-ParseSpeed-simdjson-003][EVID:EVID-20250824-ParseSpeed-Boost-005][EVID:EVID-20250824-ParseSpeed-Glaze-006].