Changeset 10038 for trunk/Mars/datacenter/scripts/jobmanager

Timestamp:

10/27/10 17:46:20 (14 years ago)

Author:

Daniela Dorner

Message:

new algorithm

File:

• trunk/Mars/datacenter/scripts/jobmanager (modified) (8 diffs)

trunk/Mars/datacenter/scripts/jobmanager

@@ -35 +35 @@
 
 # function to continue in loop and go to next script
-function nextscript()
+function sleepawhile()
 {
-   echo `date +%F\ %T`" sleeping \$$1 = $2 seconds... " >> $jmscriptlog 2>&1
-   sleep $2
+   usedsleeptime=$sleeptime
+   case $1 in 
+      "error") if ! [ "$errorsleeptime" = "" ]
+               then 
+                  if [ $errorsleeptime -lt $sleeptimelimit ]
+                  then 
+                     errorsleeptime=`echo " $errorsleeptime + $errorsleeptimedefault " | bc`
+                  fi
+                  usedsleeptime=$errorsleeptime
+               fi
+               ;;
+         "ok") errorsleeptime=$errorsleeptimedefault
+               ;;
+   esac
+   echo `date +%F\ %T`" sleeping "$usedsleeptime" seconds... (status: "$1")" >> $jmscriptlog 2>&1
    echo "" >> $jmscriptlog 2>&1
+   sleep $usedsleeptime
    continue
 }
@@ -54 +68 @@
             # (-hard) -l hostname=compute-*
             #   for qstat this returns the jobs running on that node + all jobs in the queue
-            alias 'queuesubmit'='/opt/gridengine/bin/lx26-amd64/qsub -b y -v AUTOMATIONSETUP=$AUTOMATIONSETUP -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` `echo $scriptspath`/`echo ${scripts[$i]}` '
+            alias 'queuesubmit'='/opt/gridengine/bin/lx26-amd64/qsub -b y -t 1-`echo $tosubmit` -v AUTOMATIONSETUP=$AUTOMATIONSETUP -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` `echo $scriptspath`/`echo ${scripts[$i]}` '
 #            alias 'queuesubmit'='/opt/gridengine/bin/lx26-amd64/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]}` '
 #            alias 'queuesubmit'='/opt/gridengine/bin/lx26-amd64/qsub -sc runlogpath=`echo $runlogpath` -sc date=`echo $date` -sc scriptspath=`echo $scriptspath` -sc script=`echo ${scripts[$i]}` `echo $scriptspath`/job.sge '
@@ -62 +76 @@
             ;;
       pbs)  echo " on queuing system 'pbs'" >> $jmscriptlog 2>&1
-            alias 'queuesubmit'='$pbspath/qsub -l walltime=$walltime -l pmem=$pmem -v AUTOMATIONSETUP=$AUTOMATIONSETUP,SOURCEFILEPATH=$SOURCEFILEPATH -e `echo $runlogpath`/error-`echo $date`.log -o `echo $runlogpath`/log-`echo $date`.log `echo $noderequirementsub` `echo $scriptspath`/`echo ${scripts[$i]}` '
+            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]}` '
             # check queue (restricted to current user only)
             alias 'checkqueue'="$pbspath/qstat -a -u $user | awk ' { print \"Owner\"\$2\" \" \$4\"Jobstatus\"\$10 } '"
@@ -68 +82 @@
             ;;
    condor)  echo " on queuing system 'condor'" >> $jmscriptlog 2>&1
-            alias 'queuesubmit'='$condorpath/condor_submit -a path=`echo $scriptspath` -a prog=`echo ${scripts[$i]}` -a date=`echo $date` -a dir=`echo $runlogpath` -a automationsetup=$AUTOMATIONSETUP `echo $scriptspath`/run.condor'
+            alias 'queuesubmit'='$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 $scriptspath`/run.condor'
             alias 'checkqueue'='$condorpath/condor_q -global -format "Owner%s " Owner -format "%s" CMD -format "Jobstatus%s\n" Jobstatus '
-            break
+            break 
             ;;
         *)  echo "" >> $jmscriptlog 2>&1
@@ -79 +93 @@
 echo "" >> $jmscriptlog 2>&1
 
-prev=$max
+# for processing with local storage on different nodes
 currentnode=$minnode
 numevaluated=0
+
 # endless loop
 notcount=0
-nothingtodocount=0
-nothingtodosleeptime=0
 errorsleeptime=$errorsleeptimedefault
 while (( $notcount < 100 ))
 do
+   # get and set some information for the processing
+   source `dirname $0`/sourcefile
+   # reset some values
+   tosubmit=0
+   idleratio=0
+   addtoscript=
+   
+   # get processes in queue
+   q=(`checkqueue 2>&1 `)
+   if echo $q | egrep \(Error\|failed\)
+   then 
+      echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmscriptlog 2>&1
+      printprocesslog "WARN checking query ($queuesys) failed"
+      echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmerrorlog
+      sleepawhile "error"
+   fi
+
+   # general check whether one should submit something depending on chosen algorithm
+   # algorithm 1: 
+   #   submit new jobs in case there are less than $limitidle idle jobs in the queue
+   # algorithm 2: 
+   #   submit new jobs in case the total number of jobs in the queue has fallen below $totalpno
+   case $algorithm in
+      1) # algorithm 1
+         # get number of idle jobs in the queue
+         q5=( `echo ${q[@]} | egrep -o \(Jobstatus1\|Jobstatusq\|JobstatusQ\)` )
+         idle=${#q5[@]}
+         if [ $idle -gt $limitidle ]
+         then 
+            echo `date +%F\ %T`" more than "$limitidle" jobs waiting ("$idle")" >> $jmscriptlog 2>&1
+            sleepawhile "ok"
+         fi
+         ;;
+      2) # algorithm 2
+         # get processes of user in queue
+         q1=( `echo ${q[@]} | egrep -o "Owner$user"`)
+         queued=${#q1[@]}
+         hour=`date +%k`
+         # choose array of total number of jobs to be done 
+         #   according to the day of the week
+         dayofweek=`date +%u`
+         case $dayofweek in
+            0 | 6)  totalpno=${pnototal[$hour]} ;;
+                *)  totalpno=${pnototalwe[$hour]} ;;
+         esac
+         # get total number of jobs to be submitted
+         if [ $queued -gt $totalpno ]
+         then 
+            echo `date +%F\ %T`" more than "$totalpno" jobs waiting ("$queued")" >> $jmscriptlog 2>&1
+            sleepawhile "ok"
+         else
+            tosubmittotal=`echo "$totalpno - $queued" | bc -l`
+         fi
+         ;;
+      *) echo "Please give an algorithm to calculate the number of allowed jobs."
+         exit
+         ;;
+   esac
+   echo `date +%F\ %T`" Total number of jobs to be submitted: "$tosubmittotal >> $jmscriptlog 2>&1
+
+
+   # first loop to determine 
+   # a) how many jobs of this script have to be done
+   # b) how many jobs of this script are running or queued
    for (( i=0 ; i < ${#scripts[@]} ; i++ ))
    do 
-      source `dirname $0`/sourcefile
-      echo `date +%F\ %T`" Evaluating processing status for script '${scripts[$i]}'" >> $jmscriptlog 2>&1
-
-      # check if there's something to do
+      # set the step to be evaluated
       step=${scriptscolname[$i]}
       getstepinfo
-      # check if walltime has to be set
-      if [ "$setwalltime" = "yes" ]
-      then
-         walltime=${walltimes[$i]}
-      fi
-      # check if memory has to be set
-      if [ "$setpmem" = "yes" ]
-      then
-         pmem=${pmems[$i]}
-      fi
+
       # check if the script is restricted to one node 
       #   (i.e. where output of previous step(s) is stored)
-      # this information is taken from the steps.rc file
+      #   this information is taken from the steps.rc file
+      #   currently this is implemented for sge only
+      # then get number of jobs to be done 
       if [ "$noderestricted" = "yes" ]
       then
@@ -115 +182 @@
          if [ $numevaluated -ge $numrestrictedscripts ]
          then 
-            currentnode=`echo $currentnode + 1 | bc -l`
+            currentnode=`echo " $currentnode + 1 " | bc -l`
             numevaluated=1
          else
-            numevaluated=`echo $numevaluated + 1 | bc -l`
+            numevaluated=`echo " $numevaluated + 1 " | bc -l`
          fi
          if [ $currentnode -gt $maxnode ]
@@ -138 +205 @@
          noderequirementsub=" -hard -l hostname=compute-0-${currentnode}"
          noderequirementstat=" -l hostname=compute-0-${currentnode}"
+         # get number of jobs to be done from the DB
          getstatus $currentnode >> $jmscriptlog 2>&1
       else
          noderequirementsub=""
          noderequirementstat=""
+         # get number of jobs to be done from the DB
          getstatus >> $jmscriptlog 2>&1
       fi
+      # store the number of processes to be done for this script
+      todo[$i]=$numproc
+      tododb[$i]=$numproc
       
-      # check number of processes to be done
-      echo `date +%F\ %T`" Database: $numproc ${scripts[$i]} still to be done (incl. idle jobs) [DB/step $db/$step]" >> $jmscriptlog 2>&1
-      if [ "$numproc" = "0" ]
+      # FIXME: sge cuts scriptname to 8 digits in qstat
+      # number of idle jobs, i.e. jobs waiting in the queue to run
+      #   condor: 1
+      #   sge: q
+      #   pbs: Q
+      q4=( `echo ${q[@]} | egrep -o \("${scripts[$i]}"Jobstatus1\|"${scripts[$i]}"Jobstatusq\|"${scripts[$i]}"JobstatusQ\)` )
+      idlescript[$i]=${#q4[@]}
+
+      q2=( `echo ${q[@]} | egrep -o "${scripts[$i]}"`)
+      queuedscript[$i]=${#q2[@]}
+
+      stillfree[$i]=`echo "${maxjobs[$i]} - ${queuedscript[$i]} " | bc -l`
+
+      if [ $numproc -eq 0 ] || [ ${todo[$i]} -le ${idlescript[$i]} ] || [ ${maxjobs[$i]} -le ${queuedscript[$i]} ]
       then 
-         prev=0
-         nothingtodocount=`expr $nothingtodocount + 1`
-         if [ $nothingtodocount -lt ${#scripts[@]} ]
-         then 
-            nextscript 0 0
-         else
-            if [ $nothingtodosleeptime -lt $sleeptimelimit ]
+         # store the fraction of cpus to add it to another process
+         idleratio=`echo " ${ratio[$i]} + $idleratio " | bc -l`
+         ratio[$i]=0
+         todo[$i]=0
+         idlenum=$i
+         continue
+      fi
+   done
+   echo `date +%F\ %T`" Evaluated scripts: "${scripts[@]} >> $jmscriptlog 2>&1
+   echo `date +%F\ %T`" Running scripts: "${queuedscript[@]}" (max: "${maxjobs[@]}")" >> $jmscriptlog 2>&1
+   echo `date +%F\ %T`" Number of jobs to be done (from DB): "${tododb[@]} >> $jmscriptlog 2>&1
+   echo `date +%F\ %T`" Number of jobs to be done (updated): "${todo[@]} >> $jmscriptlog 2>&1
+   echo `date +%F\ %T`" Ratio: "${ratio[@]}" (idle: "$idleratio")" >> $jmscriptlog 2>&1
+   
+   # loop to update the ratio taking into account the ratio of
+   #   a) steps where nothing has to done
+   #   b) steps where already enough jobs are in the queue
+   # sum up this idle ratio
+   # determine for which step still most jobs have to be done
+   if ! [ "$idleratio" = "0" ]
+   then 
+      addtoscript=
+      for (( i=0 ; i < ${#scripts[@]} ; i++ ))
+      do 
+         if [ ${todo[$i]} -gt ${todo[$idlenum]} ] && [ ${todo[$i]} -gt 0 ] 
+         then
+            if ! [ "$addtoscript" = "" ] 
             then 
-               nothingtodosleeptime=`echo " $nothingtodocount * $sleeptime " | bc`
+               if [ ${todo[$i]} -lt ${todo[$addtoscript]} ]
+               then
+                  continue
+               fi
             fi
-            nextscript nothingtodosleeptime $nothingtodosleeptime
-         fi
+            addtoscript=$i
+         fi
+      done
+      
+      # continue in case nothing has to be done for all steps
+      # else: update the ratio for the step where most jobs have to be done
+      #   by adding the idle ratio
+      if [ "$addtoscript" = "" ] 
+      then
+         echo `date +%F\ %T`" No jobs to be done for any step." >> $jmscriptlog 2>&1
+         sleepawhile "ok"
       else
-         nothingtodocount=0
-         nothingtodosleeptime=0
-      fi
-
-      # get processes in queue
-      q=(`checkqueue 2>&1 `)
-      if echo $q | egrep \(Error\|failed\)
-      then 
-         echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmscriptlog 2>&1
-         printprocesslog "WARN checking query ($queuesys) failed"
-         echo `date +%F\ %T`" WARN checking query ($queuesys) failed" >> $jmerrorlog
-         nextscript sleeptime $sleeptime
-      fi
-      # FIXME: sge cuts scriptname to 8 digits in qstat
-      # get processes of user in queue
-      q1=( `echo ${q[@]} | egrep -o "Owner$user"`)
-      queued=${#q1[@]}
-      # get number of scripts in queue
-      q2=( `echo ${q[@]} | egrep -o "${scripts[$i]}"`)
-      queuedscript=${#q2[@]}
-      # get running scripts
-      #   condor: 2
-      #   sge: r
-      #   pbs: R
-      q3=( `echo ${q[@]} | egrep -o \("${scripts[$i]}"Jobstatus2\|"${scripts[$i]}"Jobstatusr\|"${scripts[$i]}"JobstatusR\)` )
-      runningscript=${#q3[@]}
-      stillinqueue=`echo $queuedscript - $runningscript | bc `
-
-      #get total number of allowed process for current time
-      hour=`date +%k`
-      if [ ${pnototal[$hour]} -lt $totalmax ]
-      then
-         totalpno=${pnototal[$hour]}
-      else
-         totalpno=$totalmax
+         ratio[$addtoscript]=`echo " ${ratio[$addtoscript]} + $idleratio " | bc -l`
+      fi
+   fi
+   echo `date +%F\ %T`" Updated ratio: "${ratio[@]} >> $jmscriptlog 2>&1
+   
+   
+   # loop to submit jobs to queueing system
+   for (( i=0 ; i < ${#scripts[@]} ; i++ ))
+   do 
+      # calculate number of jobs to be submitted
+      tosubmit=`echo "scale=0; $tosubmittotal * ${ratio[$i]} / 1 " | bc -l`
+      if [ ${todo[$i]} -lt $tosubmit ]
+      then
+         echo `date +%F\ %T`" Updating tosubmit for "${scripts[$i]}" from "$tosubmit" to "${todo[$i]} >> $jmscriptlog 2>&1
+         tosubmit=${todo[$i]}
+      fi
+      if [ $tosubmit -eq 0 ]
+      then
+         echo `date +%F\ %T`" No jobs to be submitted for script '"${scripts[$i]}"'" >> $jmscriptlog 2>&1
+         continue
+      fi
+      if [ $tosubmit -gt ${stillfree[$i]} ]
+      then
+         echo `date +%F\ %T`" Updating tosubmit for "${scripts[$i]}" from "$tosubmit" to "${stillfree[$i]} >> $jmscriptlog 2>&1
+         tosubmit=${stillfree[$i]}
       fi
       
-      #choose array according to the day of the week
-      dayofweek=`date +%u`
-      case $dayofweek in
-         0 | 6)  pnos=( ${pnoswe[@]} ) ;;
-             *)  pnos=( ${pnosweek[@]} ) ;;
-      esac
-      # get number of allowed scripts for current time
-      num=`echo "((( $i + 1 ) * 24 ) + ( $hour + 1 ) ) - 24 - 1 " | bc `
-      pnoscript=${pnos[$num]}
-      # if there was nothing to do for previous script, more scripts can be allowed
-      if [ $prev -eq 0 ]
-      then
-         echo `date +%F\ %T`" \$prev=0 => resetting \$pnoscript from $pnoscript to $max [\$max]" >> $jmscriptlog 2>&1
-         pnoscript=$max
-      fi
-      echo `date +%F\ %T`" queue for user '$user': total: $queued queued jobs [allowed \$totalpno = $totalpno]" >> $jmscriptlog 2>&1
-      echo `date +%F\ %T`" queue for user '$user': ${scripts[$i]}: $queuedscript queued, $runningscript running, $stillinqueue idle [allowed \$pnoscript = $pnoscript]" >> $jmscriptlog 2>&1
-      
-      # continue if there are already enough processes or scripts in the queue
-      if [ "$queued" -ge "$totalpno" ] || [ "$queuedscript" -ge "$pnoscript" ]
-      then
-         nextscript sleeptime $sleeptime
-      fi
-      # continue if the number of script is the queue is larger (or equal) than the number which still has to be done
-      if [ $numproc -le $stillinqueue ]
-      then 
-         echo `date +%F\ %T`" \$numproc ($numproc) <  \$stillinqueue ($stillinqueue) " >> $jmscriptlog 2>&1
-         nextscript sleeptime $sleeptime
-      fi
-      
-      # reset prev
-      prev=$max
-
-      # submit 1 script to queuing system
+      # set the step to be evaluated
+      step=${scriptscolname[$i]}
+      # check if walltime has to be set
+      if [ "$setwalltime" = "yes" ]
+      then
+         walltime=${walltimes[$i]}
+      fi
+      # check if memory has to be set
+      if [ "$setpmem" = "yes" ]
+      then
+         pmem=${pmems[$i]}
+      fi
+
+      # submit $tosubmit scripts to queuing system
+      echo `date +%F\ %T`" Submitting "$tosubmit" jobs for script '"${scripts[$i]}"' to "$queuesys >> $jmscriptlog 2>&1
       date=`date +%Y-%m-%d`
-      echo `date +%F\ %T`" committing 1 ${scripts[$i]} to $queuesys" >> $jmscriptlog 2>&1
       if ! queuesubmit 2>> $jmerrorlog
       then 
@@ -240 +325 @@
          echo `date +%F\ %T`" WARN $queuesys is not working -> sleeping $errorsleeptime [\$errorsleeptime]" >> $jmscriptlog 2>&1
          printprocesslog "WARN submitting ${scripts[$i]} ($queuesys) failed"
-         if [ $errorsleeptime -lt $sleeptimelimit ]
-         then 
-            errorsleeptime=`echo " $errorsleeptime + $errorsleeptimedefault " | bc`
-         fi
-         nextscript errorsleeptime $errorsleeptime
-      else
-         errorsleeptime=$errorsleeptimedefault
-      fi
-      nextscript sleeptime $sleeptime
+         sleepawhile "error"
+      fi
    done
+   sleepawhile "ok"
 done