A 135-op cross-runtime benchmark for the L database. Every op is written in the shared q dialect, so the same script runs unmodified on any other engine that speaks it (the "reference engine"), giving a direct head-to-head.
l master_compare.q > l.csv # L side
<ref> master_compare.q > ref.csv # same script under any q-dialect engine
python3 compare.py l.csv ref.csv # per-op ratios + geomeansratio = ref_ms / l_ms — >= 1 means L is faster (2.00x = half the time).
Run both on the same idle machine, back to back.
master_compare.q times vector arithmetic, comparison, reductions, scans,
where/index, sort/group, linear algebra, fused compound chains, temporal ops,
compute-on-compressed (int, float, symbol), and 12 queries on a partitioned
on-disk table — one R,section,op,type,n,ms CSV line per op.
Headline: Apple Silicon (M-series), 2026-06-18 — overall geomean 3.57x in L's favor across 125 comparable ops. Full breakdown in results/2026-06-18_apple_silicon.md.
| file | what |
|---|---|
master_compare.q |
Cross-runtime suite (L and the reference engine). |
master_bench.q |
L-only suite: finer size grid, joins, built-in timers. |
compare.py |
Reduces two runs to per-op ratios and section geomeans. |
results/ |
Published head-to-head results. |