0b7dfc66c9f922d96a37066655a01299a3c70de0 angie Tue Nov 2 12:56:59 2021 -0700 Run matOptimize on the cluster. It uses MPI for job management; reserve cluster nodes with a fake parasol job and specify those nodes for MPI. diff --git src/hg/utils/otto/sarscov2phylo/usherClusterRun.sh src/hg/utils/otto/sarscov2phylo/usherClusterRun.sh index 2cd634c..d99b041 100755 --- src/hg/utils/otto/sarscov2phylo/usherClusterRun.sh +++ src/hg/utils/otto/sarscov2phylo/usherClusterRun.sh @@ -1,98 +1,129 @@ #!/bin/bash set -beEu -x -o pipefail # Do not modify this script, modify the source tree copy: # kent/src/hg/utils/otto/sarscov2phylo/usherClusterRun.sh usage() { echo "usage: $0 today" } if [ $# != 1 ]; then usage exit 1 fi today=$1 ottoDir=/hive/data/outside/otto/sarscov2phylo scriptDir=$(dirname "${BASH_SOURCE[0]}") source $scriptDir/util.sh usherDir=~angie/github/usher usher=$usherDir/build/usher matUtils=$usherDir/build/matUtils -matOptimize=$usherDir/build/matOptimize cd $ottoDir/$today # 16 cluster jobs, 16 threads each --> 1/4 capacity of ku cluster. (for many hours) jobCount=16 threadCount=16 sampleCount=$(vcfSamples new.masked.vcf.gz | wc -l) samplesPerJob=$(( ($sampleCount + ($jobCount-1)) / $jobCount )) echo $sampleCount samples, $jobCount jobs, $samplesPerJob samples per job for ((i=0; $i < $jobCount; i++)); do startIx=$(( 10 + ($i * $samplesPerJob) )) if [ $i == $(( $jobCount - 1 )) ]; then endIx=$(( 9 + $sampleCount )) else endIx=$(( 9 + (($i+1) * $samplesPerJob) )) fi echo job $i, genotype columns $startIx-$endIx zcat new.masked.vcf.gz \ | cut -f 1-9,$startIx-$endIx \ | gzip -c \ > new.$i.masked.vcf.gz vcfSamples new.$i.masked.vcf.gz | wc -l done cat > runUsher.sh <<EOF #!/bin/bash set -beEu vcf=\$1 pbIn=\$2 pbOut=\$3 log=\$4 /cluster/home/angie/github/usher/build/usher -u -T $threadCount -A -e 5 --max-parsimony-per-sample 20 -v \$vcf -i \$pbIn -o \$pbOut >& \$log EOF chmod a+x runUsher.sh cp /dev/null jobList for ((i=0; $i < $jobCount; i++)); do echo "./runUsher.sh new.$i.masked.vcf.gz prevRenamed.pb {check out exists preTrim.$i.pb} usher.addNew.$i.log" >> jobList done ssh ku "cd $ottoDir/$today && para -cpu=$threadCount -priority=10 make jobList" # Make a combined log (updateCombinedTree.sh looks there for sequences with excessively high EPPs) cat usher.addNew.*.log > usher.addNew.log # Merge 16 protobufs into 1 time $matUtils merge -T 8 --input-mat-1 preTrim.0.pb --input-mat-2 preTrim.1.pb -o merged.0.1.pb & $matUtils merge -T 8 --input-mat-1 preTrim.2.pb --input-mat-2 preTrim.3.pb -o merged.2.3.pb & $matUtils merge -T 8 --input-mat-1 preTrim.4.pb --input-mat-2 preTrim.5.pb -o merged.4.5.pb & $matUtils merge -T 8 --input-mat-1 preTrim.6.pb --input-mat-2 preTrim.7.pb -o merged.6.7.pb & $matUtils merge -T 8 --input-mat-1 preTrim.8.pb --input-mat-2 preTrim.9.pb -o merged.8.9.pb & $matUtils merge -T 8 --input-mat-1 preTrim.10.pb --input-mat-2 preTrim.11.pb -o merged.10.11.pb & $matUtils merge -T 8 --input-mat-1 preTrim.12.pb --input-mat-2 preTrim.13.pb -o merged.12.13.pb & $matUtils merge -T 8 --input-mat-1 preTrim.14.pb --input-mat-2 preTrim.15.pb -o merged.14.15.pb & wait time $matUtils merge -T 16 --input-mat-1 merged.0.1.pb --input-mat-2 merged.2.3.pb -o merged.0.3.pb & $matUtils merge -T 16 --input-mat-1 merged.4.5.pb --input-mat-2 merged.6.7.pb -o merged.4.7.pb & $matUtils merge -T 16 --input-mat-1 merged.8.9.pb --input-mat-2 merged.10.11.pb -o merged.8.11.pb & $matUtils merge -T 16 --input-mat-1 merged.12.13.pb --input-mat-2 merged.14.15.pb -o merged.12.15.pb & wait time $matUtils merge -T 32 --input-mat-1 merged.0.3.pb --input-mat-2 merged.4.7.pb -o merged.0.7.pb & $matUtils merge -T 32 --input-mat-1 merged.8.11.pb --input-mat-2 merged.12.15.pb -o merged.8.15.pb & wait time $matUtils merge -T 64 --input-mat-1 merged.0.7.pb --input-mat-2 merged.8.15.pb -o merged.pb -# Optimize -$matOptimize -T 40 -r 8 -S move_log -i merged.pb -M 6 -o gisaidAndPublic.$today.masked.preTrim.pb \ - >& matOptimize.merged.log +# Optimize on the cluster too, with Cheng's MPI-parallelized matOptimize +# Reserve parasol nodes with a "cluster run" of sleep commands +mkdir -p reserveKuNodes +cd reserveKuNodes +reserveHours=8 +reserveSeconds=$((3600 * $reserveHours)) +cp /dev/null jobList +for ((i=0; $i < $jobCount; i++)); do + echo "sleep $reserveSeconds" >> jobList +done +ssh ku "cd $ottoDir/$today/reserveKuNodes && para create -cpu=$threadCount jobList && para push" + +# Wait for all jobs to start (i.e. all nodes to be reserved): +echo "Waiting for parasol to assign ku nodes" +while (( $(ssh ku "parasol list jobs | grep $USER" | awk '$2 != "none"' | wc -l) < $jobCount )); do + sleep 10 +done + +# Make hostfile for mpirun +cd $ottoDir/$today +cp /dev/null hostfile +for h in $(ssh ku "parasol list jobs | grep $USER | grep sleep" | awk '{print $2;}'); do + echo "$h slots=$threadCount" >> hostfile +done + +# mpirun matOptimize on first host in hostfile +headNode=$(head -1 hostfile | awk '{print $1;}') +radius=8 +ssh $headNode "cd $ottoDir/$today && \ + $scriptDir/kuRunMatOptimize.sh -T $threadCount -r $radius \ + -S move_log.cluster -i merged.pb -o gisaidAndPublic.$today.masked.preTrim.pb \ + >& matOptimize.cluster.log" + +# Release the ku nodes by stopping the parasol batch +ssh ku "cd $ottoDir/$today/reserveKuNodes && para stop"