| 1 | #!/bin/bash
|
|---|
| 2 | #
|
|---|
| 3 | # This a script, which launches other scripts (all scripts, that are run
|
|---|
| 4 | # on primary basis)
|
|---|
| 5 | #
|
|---|
| 6 |
|
|---|
| 7 | source `dirname $0`/../Sourcefile.sh
|
|---|
| 8 | printprocesslog "INFO starting $0"
|
|---|
| 9 |
|
|---|
| 10 | set -C
|
|---|
| 11 | shopt -s expand_aliases
|
|---|
| 12 |
|
|---|
| 13 | # function to continue in loop and go to next script
|
|---|
| 14 | function sleepawhile()
|
|---|
| 15 | {
|
|---|
| 16 | usedsleeptime=$sleeptime
|
|---|
| 17 | case $1 in
|
|---|
| 18 | "error") if ! [ "$errorsleeptime" = "" ]
|
|---|
| 19 | then
|
|---|
| 20 | if [ $errorsleeptime -lt $sleeptimelimit ]
|
|---|
| 21 | then
|
|---|
| 22 | errorsleeptime=`echo " $errorsleeptime + $errorsleeptimedefault " | bc`
|
|---|
| 23 | fi
|
|---|
| 24 | usedsleeptime=$errorsleeptime
|
|---|
| 25 | fi
|
|---|
| 26 | ;;
|
|---|
| 27 | "ok") errorsleeptime=$errorsleeptimedefault
|
|---|
| 28 | ;;
|
|---|
| 29 | esac
|
|---|
| 30 | echo `date +%F\ %T`" sleeping "$usedsleeptime" seconds... (status: "$1")" >> $jmscriptlog 2>&1
|
|---|
| 31 | echo "" >> $jmscriptlog 2>&1
|
|---|
| 32 | sleep $usedsleeptime
|
|---|
| 33 | continue
|
|---|
| 34 | }
|
|---|
| 35 |
|
|---|
| 36 | echo "" >> $jmscriptlog 2>&1
|
|---|
| 37 | echo "" >> $jmscriptlog 2>&1
|
|---|
| 38 | echo -n `date +%F\ %T`" starting jobmanager for setup "$AUTOMATIONSETUP >> $jmscriptlog 2>&1
|
|---|
| 39 |
|
|---|
| 40 | user=`whoami`
|
|---|
| 41 |
|
|---|
| 42 | # choose commands according to queueing system (defined in setup)
|
|---|
| 43 | case $queuesys in
|
|---|
| 44 | sge) echo " on queuing system 'sun grid engine'" >> $jmscriptlog 2>&1
|
|---|
| 45 | # (-hard) -l hostname=compute-*
|
|---|
| 46 | # for qstat this returns the jobs running on that node + all jobs in the queue
|
|---|
| 47 | alias 'queuesubmit'='$sgepath/qsub -b y -q `echo ${queues[$i]}` -t 1-`echo $tosubmit` -v AUTOMATIONSETUP=$AUTOMATIONSETUP -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` -N `echo $step` `echo $scriptspath`/`echo ${scripts[$i]}` '
|
|---|
| 48 | #alias 'queuesubmit'='$sgepath/qsub -b y -q fact_long -t 1-`echo $tosubmit` -v AUTOMATIONSETUP=$AUTOMATIONSETUP -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` -N `echo $step` `echo $scriptspath`/`echo ${scripts[$i]}` '
|
|---|
| 49 | # alias 'queuesubmit'='$sgepath/qsub -b y -v AUTOMATIONSETUP=$AUTOMATIONSETUP -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $scriptspath`/`echo ${scripts[$i]}` '
|
|---|
| 50 | # alias 'queuesubmit'='$sgepath/qsub -sc runlogpath=`echo $runlogpath` -sc date=`echo $date` -sc scriptspath=`echo $scriptspath` -sc script=`echo ${scripts[$i]}` `echo $scriptspath`/job.sge '
|
|---|
| 51 | # FIXME: get complete scriptname (including command line option), needed for runstereo
|
|---|
| 52 | alias 'checkqueue'="$sgepath/qstat \`echo \$noderequirementstat\` | awk ' { print \"Owner\"\$4\" \" \$3\"Jobstatus\"\$5 } '"
|
|---|
| 53 | #break
|
|---|
| 54 | ;;
|
|---|
| 55 | # pbs) echo " on queuing system 'pbs'" >> $jmscriptlog 2>&1
|
|---|
| 56 | # alias 'queuesubmit'='$pbspath/qsub -t 1-`echo $tosubmit` -l walltime=`echo $walltime` -l pmem=`echo $pmem` -v AUTOMATIONSETUP=$AUTOMATIONSETUP,SOURCEFILEPATH=$SOURCEFILEPATH,SCRIPTNAME=`echo ${scripts[$i]}` -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` `echo $scriptspath`/`echo ${scripts[$i]}` '
|
|---|
| 57 | # # check queue (restricted to current user only)
|
|---|
| 58 | # alias 'checkqueue'="$pbspath/qstat -a -u $user | awk ' { print \"Owner\"\$2\" \" \$4\"Jobstatus\"\$10 } '"
|
|---|
| 59 | # #break
|
|---|
| 60 | # ;;
|
|---|
| 61 | # condor) echo " on queuing system 'condor'" >> $jmscriptlog 2>&1
|
|---|
| 62 | # alias 'queuesubmit'='$condorpath/condor_submit -a path=`echo $scriptspath` -a prog=`echo ${scripts[$i]}` -a date=`echo $date` -a dir=`echo $runlogpath` -a num=`echo $tosubmit` -a automationsetup=$AUTOMATIONSETUP `echo $scriptspath`/run.condor'
|
|---|
| 63 | # alias 'checkqueue'='$condorpath/condor_q -global -format "Owner%s " Owner -format "%s" CMD -format "Jobstatus%s\n" Jobstatus '
|
|---|
| 64 | # #break
|
|---|
| 65 | # ;;
|
|---|
| 66 | *) echo "" >> $jmscriptlog 2>&1
|
|---|
| 67 | finish >> $jmscriptlog 2>&1
|
|---|
| 68 | ;;
|
|---|
| 69 | esac
|
|---|
| 70 |
|
|---|
| 71 | echo "" >> $jmscriptlog 2>&1
|
|---|
| 72 |
|
|---|
| 73 | # for processing with local storage on different nodes
|
|---|
| 74 | currentnode=$minnode
|
|---|
| 75 | numevaluated=0
|
|---|
| 76 |
|
|---|
| 77 | # endless loop
|
|---|
| 78 | notcount=0
|
|---|
| 79 | errorsleeptime=$errorsleeptimedefault
|
|---|
| 80 | while (( $notcount < 100 ))
|
|---|
| 81 | do
|
|---|
| 82 | # get and set some information for the processing
|
|---|
| 83 | source `dirname $0`/../Sourcefile.sh
|
|---|
| 84 | # reset some values
|
|---|
| 85 | tosubmit=0
|
|---|
| 86 | idleratio=0
|
|---|
| 87 | addtoscript=
|
|---|
| 88 |
|
|---|
| 89 | # get processes in queue
|
|---|
| 90 | q=(`checkqueue 2>&1 `)
|
|---|
| 91 | if echo $q | egrep \(Error\|failed\)
|
|---|
| 92 | then
|
|---|
| 93 | echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmscriptlog 2>&1
|
|---|
| 94 | printprocesslog "WARN checking query ($queuesys) failed"
|
|---|
| 95 | echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmerrorlog
|
|---|
| 96 | sleepawhile "error"
|
|---|
| 97 | fi
|
|---|
| 98 |
|
|---|
| 99 | # general check whether one should submit something depending on chosen algorithm
|
|---|
| 100 | # algorithm 1:
|
|---|
| 101 | # submit new jobs in case there are less than $limitidle idle jobs in the queue
|
|---|
| 102 | # algorithm 2:
|
|---|
| 103 | # submit new jobs in case the total number of jobs in the queue has fallen below $totalpno
|
|---|
| 104 | case $algorithm in
|
|---|
| 105 | 1) # algorithm 1
|
|---|
| 106 | # get number of idle jobs in the queue
|
|---|
| 107 | q5=( `echo ${q[@]} | egrep -o \(Jobstatus1\|Jobstatusq\|JobstatusQ\)` )
|
|---|
| 108 | idle=${#q5[@]}
|
|---|
| 109 | if [ $idle -gt $limitidle ]
|
|---|
| 110 | then
|
|---|
| 111 | echo `date +%F\ %T`" more than "$limitidle" jobs waiting ("$idle")" >> $jmscriptlog 2>&1
|
|---|
| 112 | sleepawhile "ok"
|
|---|
| 113 | fi
|
|---|
| 114 | ;;
|
|---|
| 115 | 2) # algorithm 2
|
|---|
| 116 | # get processes of user in queue
|
|---|
| 117 | q1=( `echo ${q[@]} | egrep -o "Owner$user"`)
|
|---|
| 118 | queued=${#q1[@]}
|
|---|
| 119 | hour=`date +%k`
|
|---|
| 120 | # choose array of total number of jobs to be done
|
|---|
| 121 | # according to the day of the week
|
|---|
| 122 | dayofweek=`date +%u`
|
|---|
| 123 | case $dayofweek in
|
|---|
| 124 | 0 | 6) totalpno=${pnototal[$hour]} ;;
|
|---|
| 125 | *) totalpno=${pnototalwe[$hour]} ;;
|
|---|
| 126 | esac
|
|---|
| 127 | # get total number of jobs to be submitted
|
|---|
| 128 | if [ $queued -gt $totalpno ]
|
|---|
| 129 | then
|
|---|
| 130 | echo `date +%F\ %T`" more than "$totalpno" jobs waiting ("$queued")" >> $jmscriptlog 2>&1
|
|---|
| 131 | sleepawhile "ok"
|
|---|
| 132 | else
|
|---|
| 133 | tosubmittotal=`echo "$totalpno - $queued" | bc -l`
|
|---|
| 134 | fi
|
|---|
| 135 | ;;
|
|---|
| 136 | *) echo "Please give an algorithm to calculate the number of allowed jobs."
|
|---|
| 137 | exit
|
|---|
| 138 | ;;
|
|---|
| 139 | esac
|
|---|
| 140 | echo `date +%F\ %T`" Total number of jobs to be submitted: "$tosubmittotal >> $jmscriptlog 2>&1
|
|---|
| 141 |
|
|---|
| 142 | # first loop to determine
|
|---|
| 143 | # a) how many jobs of this script have to be done
|
|---|
| 144 | # b) how many jobs of this script are running or queued
|
|---|
| 145 | todo=()
|
|---|
| 146 | tododb=()
|
|---|
| 147 | for (( i=0 ; i < ${#scripts[@]} ; i++ ))
|
|---|
| 148 | do
|
|---|
| 149 | # set the step to be evaluated
|
|---|
| 150 | step=${scriptscolname[$i]}
|
|---|
| 151 | getstepinfo
|
|---|
| 152 |
|
|---|
| 153 | # check if the script is restricted to one node
|
|---|
| 154 | # (i.e. where output of previous step(s) is stored)
|
|---|
| 155 | # this information is taken from the steps.rc file
|
|---|
| 156 | # currently this is implemented for sge only
|
|---|
| 157 | # then get number of jobs to be done
|
|---|
| 158 | if [ "$noderestricted" = "yes" ]
|
|---|
| 159 | then
|
|---|
| 160 | # get number of next node
|
|---|
| 161 | if [ $numevaluated -ge $numrestrictedscripts ]
|
|---|
| 162 | then
|
|---|
| 163 | currentnode=`echo " $currentnode + 1 " | bc -l`
|
|---|
| 164 | numevaluated=1
|
|---|
| 165 | else
|
|---|
| 166 | numevaluated=`echo " $numevaluated + 1 " | bc -l`
|
|---|
| 167 | fi
|
|---|
| 168 | if [ $currentnode -gt $maxnode ]
|
|---|
| 169 | then
|
|---|
| 170 | currentnode=$minnode
|
|---|
| 171 | fi
|
|---|
| 172 | # check if node is excluded
|
|---|
| 173 | for excludednode in ${excludednodes[@]}
|
|---|
| 174 | do
|
|---|
| 175 | if [ $currentnode -eq $excludednode ]
|
|---|
| 176 | then
|
|---|
| 177 | echo `date +%F\ %T`" Node compute-0-$currentnode is currently excluded." >> $jmscriptlog 2>&1
|
|---|
| 178 | continue 2
|
|---|
| 179 | fi
|
|---|
| 180 | done
|
|---|
| 181 | # define requirement for submission
|
|---|
| 182 | # FIXME: currently only for sge at isdc
|
|---|
| 183 | echo `date +%F\ %T`" Checking for node $currentnode. " >> $jmscriptlog 2>&1
|
|---|
| 184 | noderequirementsub=" -hard -l hostname=compute-0-${currentnode}"
|
|---|
| 185 | noderequirementstat=" -l hostname=compute-0-${currentnode}"
|
|---|
| 186 | # get number of jobs to be done from the DB
|
|---|
| 187 | getstatus $currentnode >> $jmscriptlog 2>&1
|
|---|
| 188 | else
|
|---|
| 189 | noderequirementsub=""
|
|---|
| 190 | noderequirementstat=""
|
|---|
| 191 | # get number of jobs to be done from the DB
|
|---|
| 192 | getstatus >> $jmscriptlog 2>&1
|
|---|
| 193 | fi
|
|---|
| 194 | # store the number of processes to be done for this script
|
|---|
| 195 | todo[$i]=$numproc
|
|---|
| 196 | tododb[$i]=$numproc
|
|---|
| 197 |
|
|---|
| 198 | # FIXME: sge cuts scriptname to 8 digits in qstat
|
|---|
| 199 | # number of idle jobs, i.e. jobs waiting in the queue to run
|
|---|
| 200 | # condor: 1
|
|---|
| 201 | # sge: q
|
|---|
| 202 | # pbs: Q
|
|---|
| 203 | #q4=( `echo ${q[@]} | egrep -o \("${scripts[$i]}"Jobstatus1\|"${scripts[$i]}"Jobstatusq\|"${scripts[$i]}"JobstatusQ\)` )
|
|---|
| 204 | q4=( `echo ${q[@]} | egrep -o \("$step"Jobstatus1\|"$step"Jobstatusq\|"$step"JobstatusQ\)` )
|
|---|
| 205 | idlescript[$i]=${#q4[@]}
|
|---|
| 206 |
|
|---|
| 207 | #q2=( `echo ${q[@]} | egrep -o "${scripts[$i]}"`)
|
|---|
| 208 | q2=( `echo ${q[@]} | egrep -o "$step"`)
|
|---|
| 209 | queuedscript[$i]=${#q2[@]}
|
|---|
| 210 |
|
|---|
| 211 | stillfree[$i]=`echo "${maxjobs[$i]} - ${queuedscript[$i]} " | bc -l`
|
|---|
| 212 |
|
|---|
| 213 | if [ $numproc -eq 0 ] || [ ${todo[$i]} -le ${idlescript[$i]} ] || [ ${maxjobs[$i]} -le ${queuedscript[$i]} ]
|
|---|
| 214 | then
|
|---|
| 215 | # store the fraction of cpus to add it to another process
|
|---|
| 216 | idleratio=`echo " ${ratio[$i]} + $idleratio " | bc -l`
|
|---|
| 217 | ratio[$i]=0
|
|---|
| 218 | todo[$i]=0
|
|---|
| 219 | idlenum=$i
|
|---|
| 220 | continue
|
|---|
| 221 | fi
|
|---|
| 222 | done
|
|---|
| 223 | echo `date +%F\ %T`" Evaluated scripts: "${scripts[@]} >> $jmscriptlog 2>&1
|
|---|
| 224 | echo `date +%F\ %T`" Running scripts: "${queuedscript[@]}" (max: "${maxjobs[@]}")" >> $jmscriptlog 2>&1
|
|---|
| 225 | echo `date +%F\ %T`" Number of jobs to be done (from DB): "${tododb[@]} >> $jmscriptlog 2>&1
|
|---|
| 226 | echo `date +%F\ %T`" Number of jobs to be done (updated): "${todo[@]} >> $jmscriptlog 2>&1
|
|---|
| 227 | echo `date +%F\ %T`" Ratio: "${ratio[@]}" (idle: "$idleratio")" >> $jmscriptlog 2>&1
|
|---|
| 228 |
|
|---|
| 229 | # loop to update the ratio taking into account the ratio of
|
|---|
| 230 | # a) steps where nothing has to done
|
|---|
| 231 | # b) steps where already enough jobs are in the queue
|
|---|
| 232 | # sum up this idle ratio
|
|---|
| 233 | # determine for which step still most jobs have to be done
|
|---|
| 234 | if ! [ "$idleratio" = "0" ]
|
|---|
| 235 | then
|
|---|
| 236 | addtoscript=
|
|---|
| 237 | for (( i=0 ; i < ${#scripts[@]} ; i++ ))
|
|---|
| 238 | do
|
|---|
| 239 | if [ ${todo[$i]} -gt ${todo[$idlenum]} ] && [ ${todo[$i]} -gt 0 ]
|
|---|
| 240 | then
|
|---|
| 241 | if ! [ "$addtoscript" = "" ]
|
|---|
| 242 | then
|
|---|
| 243 | if [ ${todo[$i]} -lt ${todo[$addtoscript]} ]
|
|---|
| 244 | then
|
|---|
| 245 | continue
|
|---|
| 246 | fi
|
|---|
| 247 | fi
|
|---|
| 248 | addtoscript=$i
|
|---|
| 249 | fi
|
|---|
| 250 | done
|
|---|
| 251 |
|
|---|
| 252 | # continue in case nothing has to be done for all steps
|
|---|
| 253 | # else: update the ratio for the step where most jobs have to be done
|
|---|
| 254 | # by adding the idle ratio
|
|---|
| 255 | if [ "$addtoscript" = "" ]
|
|---|
| 256 | then
|
|---|
| 257 | echo `date +%F\ %T`" No jobs to be done for any step." >> $jmscriptlog 2>&1
|
|---|
| 258 | sleepawhile "ok"
|
|---|
| 259 | else
|
|---|
| 260 | ratio[$addtoscript]=`echo " ${ratio[$addtoscript]} + $idleratio " | bc -l`
|
|---|
| 261 | fi
|
|---|
| 262 | fi
|
|---|
| 263 | echo `date +%F\ %T`" Updated ratio: "${ratio[@]} >> $jmscriptlog 2>&1
|
|---|
| 264 |
|
|---|
| 265 |
|
|---|
| 266 | # loop to submit jobs to queueing system
|
|---|
| 267 | for (( i=0 ; i < ${#scripts[@]} ; i++ ))
|
|---|
| 268 | do
|
|---|
| 269 | # calculate number of jobs to be submitted
|
|---|
| 270 | tosubmit=`echo "scale=0; $tosubmittotal * ${ratio[$i]} / 1 " | bc -l`
|
|---|
| 271 | if [ ${todo[$i]} -lt $tosubmit ]
|
|---|
| 272 | then
|
|---|
| 273 | echo `date +%F\ %T`" Updating tosubmit for "${scripts[$i]}" from "$tosubmit" to "${todo[$i]} >> $jmscriptlog 2>&1
|
|---|
| 274 | tosubmit=${todo[$i]}
|
|---|
| 275 | fi
|
|---|
| 276 | if [ $tosubmit -eq 0 ]
|
|---|
| 277 | then
|
|---|
| 278 | echo `date +%F\ %T`" No jobs to be submitted for script '"${scripts[$i]}"'" >> $jmscriptlog 2>&1
|
|---|
| 279 | continue
|
|---|
| 280 | fi
|
|---|
| 281 | if [ $tosubmit -gt ${stillfree[$i]} ]
|
|---|
| 282 | then
|
|---|
| 283 | echo `date +%F\ %T`" Updating tosubmit for "${scripts[$i]}" from "$tosubmit" to "${stillfree[$i]} >> $jmscriptlog 2>&1
|
|---|
| 284 | tosubmit=${stillfree[$i]}
|
|---|
| 285 | fi
|
|---|
| 286 |
|
|---|
| 287 | # set the step to be evaluated
|
|---|
| 288 | step=${scriptscolname[$i]}
|
|---|
| 289 | # check if walltime has to be set
|
|---|
| 290 | if [ "$setwalltime" = "yes" ]
|
|---|
| 291 | then
|
|---|
| 292 | walltime=${walltimes[$i]}
|
|---|
| 293 | fi
|
|---|
| 294 | # check if memory has to be set
|
|---|
| 295 | if [ "$setpmem" = "yes" ]
|
|---|
| 296 | then
|
|---|
| 297 | pmem=${pmems[$i]}
|
|---|
| 298 | fi
|
|---|
| 299 |
|
|---|
| 300 | # set $tosubmit to 1 if something is 'qw', because in this case
|
|---|
| 301 | # only one line for several jobs might be shown
|
|---|
| 302 | if [ ${idlescript[$i]} -gt 0 ] # && [ "$step" == "Callisto" ]
|
|---|
| 303 | then
|
|---|
| 304 | tosubmit=1
|
|---|
| 305 | fi
|
|---|
| 306 | echo "tosubmit: "$tosubmit
|
|---|
| 307 | # submit $tosubmit scripts to queuing system
|
|---|
| 308 | #tosubmit=1 #workaround for test on fact cluster
|
|---|
| 309 | echo `date +%F\ %T`" Submitting "$tosubmit" jobs for script '"${scripts[$i]}"' to "$queuesys >> $jmscriptlog 2>&1
|
|---|
| 310 | date=`date +%Y-%m-%d`
|
|---|
| 311 | if ! queuesubmit 2>> $jmerrorlog
|
|---|
| 312 | then
|
|---|
| 313 | echo `date`" WARN submitting job ($queuesys) failed" >> $jmerrorlog
|
|---|
| 314 | echo `date +%F\ %T`" WARN $queuesys is not working -> sleeping $errorsleeptime [\$errorsleeptime]" >> $jmscriptlog 2>&1
|
|---|
| 315 | printprocesslog "WARN submitting ${scripts[$i]} ($queuesys) failed"
|
|---|
| 316 | sleepawhile "error"
|
|---|
| 317 | fi
|
|---|
| 318 | echo ""
|
|---|
| 319 | done
|
|---|
| 320 | sleepawhile "ok"
|
|---|
| 321 | done
|
|---|
| 322 |
|
|---|
