New/Changed job control qualifiers

Software Performance

Timing information useful for comparing execution time between models and/or builds of ASReml is available in the .asl file if the !LOGFILE and !DEBUG qualifiers (or command line options -DL) are set on the first line of the job (above the TITLE line). Running the command
grep '>>' job.asl
at the command prompt will extract timing information. (grep is a UNIX/Linux command used to find a specific string from inside a file. For Windows, the grep alternative is findstr.) We give two examples. The first compares the XFA and RRD forms of the Factor Analysis in genomic models. The second example shows the advantage of trimming (see !VPGROUP)

XFA/RRD Example

This is a complex multi environment analysis with 24 experiments and 884 genetic lines linked with a G matrix for a total of 12,354 records. The model fitted is

 yield ~ mu Experiment,
 !r Rep.Experiment Block.Rep.Experiment grm(Line).xfa1(Experiment)

and it fits 22,797 equations. The log-file output for this analysis is

 >> Windows x64   30.0 Gbyte  PhenoG2F8.ni/PhenoG2F  17 Apr 2025 08:59:00.497
 >> >>      ASReml Process        Clock   SumClock
 >> >>     Getting Started: sec       1.04       1.04
 >> >>       Before Order : sec       0.20       1.24
 >> >>        * Ordering : sec       2.23       3.48
 FILLIN 26879705 ==>> 27580197    1.03, #SR: 22772, AvLen: 1211, MxLen: 2023, AvFlops: 818711
 >> >>         C reordered: sec       4.63       5.87
 >> Cabs blocks          407   49.93858              95
 >> >>          C absorbed: sec       3.32       9.19
 >> >>           WV formed: sec       0.06       9.25
 >> >>           AI formed: sec       0.38       9.63
 >> >>           Ci formed: sec       4.22      13.86
         373  >> Ci nodal blocks; average    55.13137
 >> >>        Ci reordered: sec       1.68      15.54
 >> >>         Gscore done: sec       1.20      16.75
 >> >>  Iteration complete: sec       0.02      16.77
 >> >>  Iteration complete: sec      13.24      30.01
 FILLIN 25780785 ==>> 26263304    1.02, #SR: 22772, AvLen: 1153, MxLen: 1781, AvFlops: 772350
 >> >>     Iterations done: sec     152.06     182.07
 >> >>         SLN written: sec       0.12     182.19
 >> >>         YHT written: sec       0.07     182.26
 >> >>            Finished: sec       0.00     182.26

This job required 13 iterations in 182 seconds (that is, 13.2 seconds per iteration), and had a log-likelihood value of -46,808.1. If we fit the following equivalent model

 yield ~ mu Experiment !r Rep.Experiment Block.Rep.Experiment,
 rr1(Experiment).grm(Line) diag(Experiment).grm(Line)

it takes 3.3 seconds per iteration, for a total of 15 iterations with a very similar log-likelihood value of -46,810.2.

 >> Windows x64   30.0 Gbyte  PhenoG2F9/PhenoG2F  17 Apr 2025 10:28:11.087
 >> >>      ASReml Process        Clock   SumClock
 >> >>     Getting Started: sec       1.03       1.03
 >> >>       Before Order : sec       0.07       1.10
 >> >>        * Ordering : sec       1.22       2.32
 FILLIN 9391895 ==>> 10424393    1.11, #SR: 43988, AvLen: 237, MxLen: 862
 >> >>         C reordered: sec       1.69       2.79
 >> Cabs blocks          292   73.28082              95
 >> >>          C absorbed: sec       0.70       3.49
 >> >>           WV formed: sec       0.20       3.68
 >> >>           AI formed: sec       0.12       3.81
 >> >>           Ci formed: sec       0.70       4.51
         233  >> Ci nodal blocks; average    88.08154
 >> >>        Ci reordered: sec       0.38       4.89
 >> >>         Gscore done: sec       0.77       5.66
 >> >>  Iteration complete: sec       0.01       5.67
 >> >>  Iteration complete: sec       3.29       8.96
 >> >>     Iterations done: sec      44.33      53.29
 >> >>         SLN written: sec       0.22      53.51
 >> >>            Finished: sec       0.07      53.58

Typically, as shown above, the major components are: Order found, C absorbed and Ci formed. The FILLIN line reports the size of the C matrix before and after absorption, and the ratio (after/before) of the sizes.

Genetic Model Trimming Example

Consider the analysis of a breeding experiment for wheatgrass, a perennial wheat species. The analysis is of 22 harvests taken from 10 trials planted over 4 years. The selection of genotypes planted in a trial differs considerably, the later plantings being lines (genotypes) derived from earlier plantings. A simplified model for the trait plant height, PTHT, using a genotype relationship matrix (marker-based) to provide genetic links is

 PTHT ~ mu mv !r Harvest rr1(Harvest).grm1(geno) at(Harvest).grm1(geno) female.male
 residual sat(Harvest).ar1v(row).ar1(range)

This corresponds to a factor analytic model where the term provides the overall genetic covariances while the at(Harvest).grm1(geno) term provides the specific genotype by harvest variance components for each harvest.

The coefficient matrix includes 23 instances of the full genomic matrix, which is of order 4,482. An alternative way of setting up the model is just to include the genotypes with data in the GRM used for estimating the 22 specific variance components. The matrices vary in size from 308 to 576 genotypes. The model in this case is

 PTHT ~ mu mv !r Harvest rr1(Harvest).grm1(geno) sat(Harvest).grm1(geno) female.male
 residual sat(Harvest).ar1v(row).ar1(range)

The variance components from this model are the same as for the full model, but this model took 33.3 seconds per iteration compared with 116.7 seconds per iteration for the original model.

Comparison of the model processing details is presented in the table below.

action	using at()	using sat()
Initialize (sec)	1	1
Ordering (sec)	50	7
Absorption (sec)	43	4
Forming AI (sec)	13	13
Sparse Inverse (sec)	40	4
Score (sec)	16	13
Equations	367,587	278,397
C matrix cells	214,560,191	12,453,485

Back to What's New

Return to index