Benchmarks

Benchmarks

The measurements below describe specific tested workloads. Runtime depends on panel size, annotations, context dimension, storage, and CPU configuration. For estimated genetic variance in UK Biobank, see Real-data results.

PGS synthetic benchmark

The PGS benchmark uses 2,048 generated samples and 4,096 variants, two trait masks, ten models per trait, five context columns, and one CPU thread. Each process performs setup, one warmup, and three operator repetitions.

Genotype storage	Setup, seconds	Median operator, seconds	Process peak RSS, MiB
Stream	0.1738	0.3730	222.4
Compact	0.1780	0.3296	224.1
Standardized FP64	0.1708	0.1788	261.1

These Linux measurements used pthread BLIS, SNP blocks of 256, and RHS width 64. Peak RSS includes fixture generation. The timing covers a covariance operation, not a complete fit. Full-cohort PGS memory and throughput have not been measured.

Run each mode in a fresh process:

python scripts/prediction/checkout.py benchmark --mode stream --threads 1
python scripts/prediction/checkout.py benchmark --mode compact --threads 1
python scripts/prediction/checkout.py benchmark --mode standardized --threads 1

Cached modes read the source once during setup. Streaming reads it on every operator call. Repeated outputs were identical in the recorded runs.

Batch h² timing

A recorded four-trait comparison used 7,774,235 SNPs, 59 annotation columns, chromosome jackknife, and four threads:

Mode	Total wall time, seconds	Peak RSS, GB
Regular	689.8	31.92
Fast, uncached text	724.8	32.03
Fast, cached binary	532.4	31.93

The 480 serialized component values differed by at most 2.7e-12. Cached trait processing took 6.43 seconds for four traits, but the common model load dominated the end-to-end run. Uncached batching was not faster in this comparison.

Measuring a new workload

Measure complete wall time through result writing and reload, peak RSS, and input traversals. Compare estimates at the same solver tolerance or random-vector count. For PGS, include held-out scoring; for randomized references, check more than one seed and vector count before choosing a production setting.