50K pedigree REML qualification against DMU

This benchmark compares gsuite 0.29.0 with DMU 5.6 on one frozen 50,000-animal pedigree using single-trait, multitrait, and rank-2 longitudinal animal models. It is a practical cross-software qualification, not a repeated performance benchmark or an exact simulation oracle.

The public gsim 0.2.0 generators are used unchanged. Their latent effects are not exact inbreeding-aware draws from the fitted relationship model, exact_relationship_draw is false, multitrait missingness follows gsim’s native observation patterns, and the realized phenotyped count is 42,497. Truth comparisons are therefore descriptive. The primary evidence is agreement between gsuite and DMU on the same frozen observations and pedigree.

Results

Back to all benchmarks · Public scope

The saved results below compare gsuite 0.29.0 and DMU 5.6. Both used direct pedigree REML on identical frozen data and starts. gsuite used CHOLMOD selected inversion; DMU used dmu1 followed by direct dmuai. Across the three models, the largest absolute covariance-estimate difference is below \(4.3\times10^{-8}\).

Prediction metrics use the predeclared first 1,000 pedigree-coded animals (856 phenotyped and 144 unphenotyped), not all 50,000 animals. The tables show Pearson correlation and normalized RMSE for each effect on that subset.

Timings are single executions on the same Intel Core i7-1365U Windows machine. gsuite times cover the fit; DMU times cover the dmu1 plus dmuai pipeline. They include different work boundaries and do not establish general speedups. See the selected hardware, software and thread settings and the convergence table below. The original development PDF is not part of this public snapshot.

Single-trait results

Covariance parameter gsuite DMU
G[1,1] 0.9912148142 0.9912148200
R[1,1] 1.1348505615 1.1348506000

\(G\) denotes the genetic covariance matrix; \(R\) denotes the residual covariance matrix.

Prediction effect Animals Pearson correlation Normalized RMSE
animal 1000 0.999999999998527 1.721e-06
Program Observed wall time (seconds)
gsuite fit 80.920
DMU pipeline 178.421

Multitrait results

Covariance parameter gsuite DMU
G[1,1] 1.0513593430 1.0513593000
G[2,1] 0.4060368505 0.4060368500
G[2,2] 0.8012559231 0.8012559300
R[1,1] 1.0986902326 1.0986902000
R[2,1] 0.2413718588 0.2413718600
R[2,2] 0.8975498573 0.8975498500

\(G\) denotes the genetic covariance matrix; \(R\) denotes the residual covariance matrix.

Prediction effect Animals Pearson correlation Normalized RMSE
trait1 1000 0.999999999998714 1.604e-06
trait2 1000 0.999999999998806 1.552e-06
Program Observed wall time (seconds)
gsuite fit 105.640
DMU pipeline 818.460

Longitudinal results

Covariance parameter gsuite DMU
Q[1,1] 1.0057441325 1.0057441000
Q[2,1] 0.1463268382 0.1463268400
Q[2,2] 0.2475494863 0.2475494800
R[1,1] 0.7974774374 0.7974774400

\(Q\) denotes the genetic random-regression coefficient covariance matrix; \(R\) denotes the residual covariance matrix.

Prediction effect Animals Pearson correlation Normalized RMSE
coefficient1 1000 0.999999999998514 1.734e-06
coefficient2 1000 0.999999999999474 1.081e-06
Program Observed wall time (seconds)
gsuite fit 386.860
DMU pipeline 1182.680

Saved evidence

Single trait: gsuite 80.9 seconds, DMU 178.4. Multitrait: gsuite 105.6, DMU 818.5. Longitudinal: gsuite 386.9, DMU 1182.7.

Observed wall times on the frozen pedigree models. Each bar is one execution; gsuite fit and DMU preprocessing-plus-fitting timings have different work boundaries.

The website figure is drawn from the saved timing CSV.

These tables summarize the committed evidence. No model fitting or new raw-output reparse is performed when rendering this page. Likelihood values are not compared because the DMU additive constant remains unreconciled.