| 1 |
gutsche |
1.1 |
#!/usr/bin/env perl
|
| 2 |
|
|
|
| 3 |
|
|
##H This drop box agent collects the output from CRAB tasks after the
|
| 4 |
|
|
##H the task jobs have run to completion. Tasks with jobs in them
|
| 5 |
|
|
##H should not be passed into this agent until all jobs have completed.
|
| 6 |
|
|
##H
|
| 7 |
|
|
##H Usage:
|
| 8 |
|
|
##H TaskQuery
|
| 9 |
|
|
##H -state DIRECTORY [-next NEXT] [-wait SECS]
|
| 10 |
|
|
##H
|
| 11 |
|
|
##H -state agent state directory, including inbox
|
| 12 |
|
|
##H -next next agent to pass the drops to; can be given several times
|
| 13 |
|
|
##H -wait time to wait in seconds between work scans
|
| 14 |
gutsche |
1.2 |
##H -mode mode: EGEE or OSG, prevents kill of aborted jobs for EGEE
|
| 15 |
gutsche |
1.1 |
|
| 16 |
|
|
BEGIN {
|
| 17 |
|
|
use strict; use warnings; $^W=1;
|
| 18 |
|
|
our $me = $0; $me =~ s|.*/||;
|
| 19 |
|
|
our $home = $0; $home =~ s|/[^/]+$||; $home ||= "."; $home .= "/../PHEDEX/Toolkit/Common";
|
| 20 |
|
|
unshift(@INC, $home);
|
| 21 |
|
|
}
|
| 22 |
|
|
|
| 23 |
|
|
######################################################################
|
| 24 |
|
|
use UtilsHelp;
|
| 25 |
|
|
while (scalar @ARGV) {
|
| 26 |
|
|
if ($ARGV[0] eq '-state' && scalar @ARGV > 1) {
|
| 27 |
|
|
shift (@ARGV); $args{DROPDIR}= shift(@ARGV);
|
| 28 |
|
|
} elsif ($ARGV[0] eq '-next' && scalar @ARGV > 1) {
|
| 29 |
|
|
shift (@ARGV); push (@{$args{NEXTDIR}}, shift(@ARGV));
|
| 30 |
|
|
} elsif ($ARGV[0] eq '-wait' && scalar @ARGV > 1) {
|
| 31 |
|
|
shift (@ARGV); $args{WAITTIME} = shift(@ARGV);
|
| 32 |
gutsche |
1.2 |
} elsif ($ARGV[0] eq '-mode' && scalar @ARGV > 1) {
|
| 33 |
|
|
shift (@ARGV); $args{MODE} = shift(@ARGV);
|
| 34 |
gutsche |
1.1 |
} elsif ($ARGV[0] eq '-h') {
|
| 35 |
|
|
&usage();
|
| 36 |
|
|
} else {
|
| 37 |
|
|
last;
|
| 38 |
|
|
}
|
| 39 |
|
|
}
|
| 40 |
|
|
|
| 41 |
|
|
if (@ARGV || !$args{DROPDIR}) {
|
| 42 |
|
|
die "Insufficient parameters, use -h for help.\n";
|
| 43 |
|
|
}
|
| 44 |
|
|
|
| 45 |
|
|
(new TaskCollect (%args))->process();
|
| 46 |
|
|
|
| 47 |
|
|
######################################################################
|
| 48 |
|
|
# Routines specific to this agent.
|
| 49 |
|
|
package TaskCollect; use strict; use warnings; use base 'UtilsAgent';
|
| 50 |
|
|
use UtilsCommand;
|
| 51 |
|
|
use UtilsLogging;
|
| 52 |
|
|
use UtilsTiming;
|
| 53 |
|
|
|
| 54 |
|
|
sub getTimeSinceStart
|
| 55 |
|
|
{
|
| 56 |
|
|
# extract time since start of project in hours
|
| 57 |
|
|
my ($self, $taskdir) = @_;
|
| 58 |
|
|
my $time=time();
|
| 59 |
|
|
|
| 60 |
|
|
# read start time from file in taskdir
|
| 61 |
|
|
open( FILE, "< $taskdir/TASK_INIT.txt" ) or die "Can't open $taskdir/TASK_INIT.txt : $!";
|
| 62 |
|
|
my $starttime = 0.0;
|
| 63 |
|
|
while ( <FILE> ) {
|
| 64 |
|
|
chomp;
|
| 65 |
|
|
$starttime = scalar ($_);
|
| 66 |
|
|
}
|
| 67 |
|
|
close FILE;
|
| 68 |
|
|
|
| 69 |
|
|
my $diffsec = $time - $starttime;
|
| 70 |
|
|
my $diffhours = $diffsec/3600.;
|
| 71 |
|
|
return $diffhours;
|
| 72 |
|
|
}
|
| 73 |
|
|
|
| 74 |
|
|
sub new
|
| 75 |
|
|
{
|
| 76 |
|
|
my $proto = shift;
|
| 77 |
|
|
my $class = ref($proto) || $proto;
|
| 78 |
|
|
my $self = $class->SUPER::new(@_);
|
| 79 |
|
|
# OLI 060702
|
| 80 |
|
|
# deactivate resubmission.
|
| 81 |
gutsche |
1.2 |
my %params = (MODE => 'EGEE');
|
| 82 |
gutsche |
1.1 |
my %args = (@_);
|
| 83 |
|
|
map { $self->{$_} = $args{$_} || $params{$_} } keys %params;
|
| 84 |
|
|
bless $self, $class;
|
| 85 |
|
|
return $self;
|
| 86 |
|
|
}
|
| 87 |
|
|
|
| 88 |
|
|
# Actually process the drop.
|
| 89 |
|
|
sub processDrop {
|
| 90 |
|
|
my ($self, $drop) = @_;
|
| 91 |
|
|
|
| 92 |
|
|
# cmd variable
|
| 93 |
|
|
my $cmd = "";
|
| 94 |
|
|
|
| 95 |
|
|
# Sanity checking
|
| 96 |
|
|
return if (! $self->inspectDrop ($drop));
|
| 97 |
|
|
# waiting time for too many status checks
|
| 98 |
|
|
return if (($$self{NEXT_CHECK}{$drop} || 0) > time());
|
| 99 |
|
|
$$self{NEXT_CHECK}{$drop} = time() + 0;
|
| 100 |
|
|
delete $self->{BAD}{$drop};
|
| 101 |
|
|
&timeStart($self->{STARTTIME});
|
| 102 |
|
|
|
| 103 |
|
|
# Read CRAB confiugration
|
| 104 |
|
|
my $taskdir = &input ("$self->{WORKDIR}/$drop/task");
|
| 105 |
|
|
if (! $taskdir || ! -d $taskdir) {
|
| 106 |
|
|
&alert ("missing task directory in $drop");
|
| 107 |
|
|
$self->markBad ($drop);
|
| 108 |
|
|
return;
|
| 109 |
|
|
}
|
| 110 |
|
|
|
| 111 |
|
|
# Time to check job status on the task again. Let CRAB do the hard
|
| 112 |
|
|
# work, then check what it found.
|
| 113 |
|
|
$cmd = qq{
|
| 114 |
|
|
cd $taskdir || exit \$?;
|
| 115 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 116 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 117 |
asciaba |
1.5 |
time crab -continue -status; a=\$?; which UpdateOldCrabDb >/dev/null 2>&1 && UpdateOldCrabDb; exit \$a) >> JOB_STATUS_LOG.txt 2>&1};
|
| 118 |
gutsche |
1.1 |
if (my $rc = system($cmd)) {
|
| 119 |
|
|
&alert ("$drop $taskdir: failed to check job status: exit code @{[&runerror($rc)]}, move jobdb out of way");
|
| 120 |
|
|
# move crab jobs db out of the way to have the jobs not be counted in statistics for project creation (TaskSource)
|
| 121 |
|
|
$cmd = qq{
|
| 122 |
|
|
cd $taskdir || exit \$?;
|
| 123 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 124 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 125 |
|
|
mv -v crab_*/share/db/jobs crab.db.jobs.old) >> JOB_REMOVE_CRABDIR_FAILED_STATUS_CHECK_LOG.txt 2>&1};
|
| 126 |
|
|
if (my $rc3 = system($cmd)) {
|
| 127 |
|
|
&alert ("$drop $taskdir: failed to move crab jobs to crab.db.jobs.old");
|
| 128 |
|
|
}
|
| 129 |
|
|
# mark task as finished
|
| 130 |
|
|
&logmsg("stats: $drop $taskdir @{[&formatElapsedTime($self->{STARTTIME})]} failed checking status");
|
| 131 |
|
|
&output ("$taskdir/TASK_FINISHED.txt", &mytimeofday () . "\n");
|
| 132 |
|
|
$self->markBad ($drop);
|
| 133 |
|
|
return;
|
| 134 |
|
|
}
|
| 135 |
|
|
|
| 136 |
|
|
# parse jobdb
|
| 137 |
|
|
my $f = (<$taskdir/crab_*/share/db/jobs>)[0];
|
| 138 |
|
|
|
| 139 |
|
|
my @gridids = ();
|
| 140 |
|
|
my @crabids = ();
|
| 141 |
|
|
my @jobstat = ();
|
| 142 |
|
|
foreach my $status (split(/\n/, &input($f) || '')) {
|
| 143 |
|
|
my @statusarray = split('\|', $status);
|
| 144 |
|
|
push(@crabids, $statusarray[4]);
|
| 145 |
|
|
push(@gridids, $statusarray[3]);
|
| 146 |
|
|
push(@jobstat, $statusarray[1]);
|
| 147 |
|
|
}
|
| 148 |
|
|
|
| 149 |
|
|
my $hold = 0;
|
| 150 |
|
|
|
| 151 |
|
|
# check time between start of project and now
|
| 152 |
|
|
my $lifetime = $self->getTimeSinceStart("$taskdir");
|
| 153 |
belforte |
1.6 |
# set time after which all jobs are cleaned and task marked bad
|
| 154 |
|
|
my $cleanuptime = 31.;
|
| 155 |
gutsche |
1.1 |
# set time after which all jobs (running, pending and others) are killed
|
| 156 |
belforte |
1.6 |
my $killtime = 30.;
|
| 157 |
gutsche |
1.1 |
# set time after which pending and other jobs are killed, running jobs are left running
|
| 158 |
|
|
my $stoptime = 24.;
|
| 159 |
|
|
|
| 160 |
|
|
for (my $i = 0; $i <= $#jobstat; ++$i) {
|
| 161 |
|
|
# Compute CRAB index and nice label for this job
|
| 162 |
|
|
my $crabi = $i+1;
|
| 163 |
|
|
my $joblabel = $gridids[$i];
|
| 164 |
|
|
$joblabel =~ s/([^-A-Za-z0-9_:.])/sprintf("%%%02x", ord($1))/ge;
|
| 165 |
|
|
|
| 166 |
|
|
# Determine job status.
|
| 167 |
|
|
my $status = $jobstat[$i];
|
| 168 |
|
|
if ($status !~ /^[YHSCARDKZ]$/) {
|
| 169 |
|
|
&alert ("$drop $taskdir: unexpected job status `$status' for `$crabi'");
|
| 170 |
|
|
# kill job
|
| 171 |
|
|
$cmd = qq{
|
| 172 |
|
|
cd $taskdir || exit \$?;
|
| 173 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 174 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 175 |
belforte |
1.4 |
time crab -continue -kill '$crabids[$i]'; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb) >> JOB_KILL_LOG.txt 2>&1};
|
| 176 |
gutsche |
1.1 |
system($cmd);
|
| 177 |
|
|
next;
|
| 178 |
|
|
}
|
| 179 |
|
|
|
| 180 |
|
|
# if job is created but not submitted, submit again, check afterwards, otherwise kill and continue
|
| 181 |
|
|
if ($status eq 'C') {
|
| 182 |
|
|
&alert ("$drop $taskdir: job $crabids[$i] not submitted yet, submit now ");
|
| 183 |
|
|
$cmd = qq{
|
| 184 |
|
|
cd $taskdir || exit \$?;
|
| 185 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 186 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 187 |
asciaba |
1.5 |
time crab -continue -submit '$crabids[$i]'; a=\$?; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb; exit \$a) >> JOB_SUBMIT_LOG.txt 2>&1};
|
| 188 |
gutsche |
1.1 |
if (my $rc = system($cmd)) {
|
| 189 |
|
|
&alert ("$drop $taskdir: failed to submit job $crabids[$i] with exit code @{[&runerror($rc)]}, kill job");
|
| 190 |
|
|
# kill job
|
| 191 |
|
|
$cmd = qq{
|
| 192 |
|
|
cd $taskdir || exit \$?;
|
| 193 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 194 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 195 |
belforte |
1.4 |
time crab -continue -kill '$crabids[$i]'; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb) >> JOB_KILL_LOG.txt 2>&1};
|
| 196 |
gutsche |
1.1 |
system($cmd);
|
| 197 |
|
|
}
|
| 198 |
|
|
next;
|
| 199 |
|
|
}
|
| 200 |
|
|
|
| 201 |
belforte |
1.6 |
if ( $lifetime >= $cleanuptime ) {
|
| 202 |
|
|
&alert ("$drop $taskdir: task kill did not work in time");
|
| 203 |
|
|
# move crab jobs db out of the way to have the jobs not be counted in statistics for project creation (TaskSource)
|
| 204 |
|
|
$cmd = qq{
|
| 205 |
|
|
cd $taskdir || exit \$?;
|
| 206 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 207 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 208 |
|
|
mv -v crab_*/share/db/jobs crab.db.jobs.old) >> JOB_REMOVE_CRABDIR_KILL_FAILED.txt 2>&1};
|
| 209 |
|
|
if (my $rc = system($cmd)) {
|
| 210 |
|
|
&alert ("$drop $taskdir: failed to move crab jobs to crab.db.jobs.old");
|
| 211 |
|
|
}
|
| 212 |
|
|
# mark task as finished
|
| 213 |
|
|
&output ("$taskdir/TASK_FINISHED.txt", &mytimeofday () . "\n");
|
| 214 |
|
|
$self->markBad ($drop);
|
| 215 |
|
|
return;
|
| 216 |
|
|
}
|
| 217 |
|
|
|
| 218 |
|
|
elsif ( $lifetime >= $killtime ) {
|
| 219 |
gutsche |
1.1 |
|
| 220 |
|
|
# That's all for cleared jobs.
|
| 221 |
|
|
next if ($status eq 'Y');
|
| 222 |
|
|
|
| 223 |
|
|
# That's all for killed jobs.
|
| 224 |
|
|
next if ($status eq 'K');
|
| 225 |
|
|
|
| 226 |
belforte |
1.3 |
# Retrieve job logging information if abort
|
| 227 |
|
|
if ($status eq 'A') {
|
| 228 |
|
|
my $outfile = "JOB_HISTORY.$crabi.$status.$joblabel.txt";
|
| 229 |
|
|
if (! -f "$taskdir/$outfile") {
|
| 230 |
|
|
$cmd = qq{
|
| 231 |
|
|
cd $taskdir || exit \$?;
|
| 232 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 233 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 234 |
|
|
time crab -continue -postMortem '$crabids[$i]') > $outfile 2>&1};
|
| 235 |
|
|
system($cmd);
|
| 236 |
|
|
}
|
| 237 |
gutsche |
1.1 |
}
|
| 238 |
|
|
|
| 239 |
|
|
# get output for job if done
|
| 240 |
|
|
if ($status eq 'D') {
|
| 241 |
|
|
next;
|
| 242 |
|
|
}
|
| 243 |
|
|
|
| 244 |
|
|
# kill the rest
|
| 245 |
|
|
$cmd = qq{
|
| 246 |
|
|
cd $taskdir || exit \$?;
|
| 247 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 248 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 249 |
belforte |
1.4 |
time crab -continue -kill '$crabids[$i]'; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb) >> JOB_KILL_LOG.txt 2>&1};
|
| 250 |
gutsche |
1.1 |
system($cmd);
|
| 251 |
belforte |
1.6 |
# come back to verify it was killed
|
| 252 |
|
|
$hold = 1;
|
| 253 |
gutsche |
1.1 |
next;
|
| 254 |
|
|
|
| 255 |
|
|
} elsif ( $lifetime >= $stoptime ) {
|
| 256 |
|
|
|
| 257 |
|
|
# keep the running
|
| 258 |
|
|
if ($status =~ /^[R]$/) {
|
| 259 |
|
|
$hold = 1;
|
| 260 |
|
|
next;
|
| 261 |
|
|
}
|
| 262 |
|
|
|
| 263 |
|
|
# That's all for cleared jobs.
|
| 264 |
|
|
next if ($status eq 'Y');
|
| 265 |
|
|
|
| 266 |
|
|
# That's all for killed jobs.
|
| 267 |
|
|
next if ($status eq 'K');
|
| 268 |
|
|
|
| 269 |
belforte |
1.3 |
# Retrieve job logging information if abort
|
| 270 |
|
|
if ($status eq 'A') {
|
| 271 |
|
|
my $outfile = "JOB_HISTORY.$crabi.$status.$joblabel.txt";
|
| 272 |
|
|
if (! -f "$taskdir/$outfile") {
|
| 273 |
|
|
$cmd = qq{
|
| 274 |
|
|
cd $taskdir || exit \$?;
|
| 275 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 276 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 277 |
|
|
time crab -continue -postMortem '$crabids[$i]') > $outfile 2>&1};
|
| 278 |
|
|
system($cmd);
|
| 279 |
|
|
}
|
| 280 |
gutsche |
1.1 |
}
|
| 281 |
|
|
# get output for job if done
|
| 282 |
|
|
if ($status eq 'D') {
|
| 283 |
|
|
next;
|
| 284 |
|
|
}
|
| 285 |
|
|
|
| 286 |
|
|
# kill the rest
|
| 287 |
|
|
$cmd = qq{
|
| 288 |
|
|
cd $taskdir || exit \$?;
|
| 289 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 290 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 291 |
belforte |
1.4 |
time crab -continue -kill '$crabids[$i]'; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb) >> JOB_KILL_LOG.txt 2>&1};
|
| 292 |
gutsche |
1.1 |
system($cmd);
|
| 293 |
|
|
next;
|
| 294 |
|
|
|
| 295 |
|
|
} else {
|
| 296 |
|
|
|
| 297 |
|
|
# Process the easy cases.
|
| 298 |
|
|
if ($status =~ /^[SR]$/) {
|
| 299 |
|
|
$hold = 1;
|
| 300 |
|
|
next;
|
| 301 |
|
|
}
|
| 302 |
|
|
|
| 303 |
|
|
# That's all for cleared jobs.
|
| 304 |
|
|
next if ($status eq 'Y');
|
| 305 |
|
|
|
| 306 |
|
|
# That's all for killed jobs.
|
| 307 |
|
|
next if ($status eq 'K');
|
| 308 |
|
|
|
| 309 |
belforte |
1.3 |
# Retrieve job logging information if abort
|
| 310 |
|
|
if ($status eq 'A') {
|
| 311 |
|
|
my $outfile = "JOB_HISTORY.$crabi.$status.$joblabel.txt";
|
| 312 |
|
|
if (! -f "$taskdir/$outfile") {
|
| 313 |
|
|
$cmd = qq{
|
| 314 |
|
|
cd $taskdir || exit \$?;
|
| 315 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 316 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 317 |
|
|
time crab -continue -postMortem '$crabids[$i]') > $outfile 2>&1};
|
| 318 |
|
|
system($cmd);
|
| 319 |
|
|
}
|
| 320 |
gutsche |
1.1 |
}
|
| 321 |
|
|
|
| 322 |
|
|
# get output for job if done
|
| 323 |
|
|
if ($status eq 'D') {
|
| 324 |
|
|
next;
|
| 325 |
|
|
}
|
| 326 |
|
|
|
| 327 |
gutsche |
1.2 |
# kill the rest if mode OSG
|
| 328 |
|
|
if ( $self->{MODE} eq "OSG" ) {
|
| 329 |
|
|
$cmd = qq{
|
| 330 |
gutsche |
1.1 |
cd $taskdir || exit \$?;
|
| 331 |
|
|
TIMEFORMAT="Timing information in seconds: \%R real, \%U user, \%S system, \%P \%\%";
|
| 332 |
|
|
(date -u +"START \%Y\%m\%dZ\%H\%M\%S == \%s"; set -x;
|
| 333 |
belforte |
1.4 |
time crab -continue -kill '$crabids[$i]'; which CrabStatusAndUpdateOldCrabDb >/dev/null 2>&1 && CrabStatusAndUpdateOldCrabDb) >> JOB_KILL_LOG.txt 2>&1};
|
| 334 |
gutsche |
1.2 |
system($cmd);
|
| 335 |
|
|
next;
|
| 336 |
|
|
}
|
| 337 |
gutsche |
1.1 |
|
| 338 |
|
|
}
|
| 339 |
|
|
}
|
| 340 |
|
|
|
| 341 |
|
|
# Check if we are supposed to hold this task.
|
| 342 |
|
|
return if $hold;
|
| 343 |
|
|
|
| 344 |
|
|
&touch ("$self->{WORKDIR}/$drop/done");
|
| 345 |
|
|
delete $$self{NEXT_CHECK}{$drop};
|
| 346 |
|
|
$self->relayDrop ($drop);
|
| 347 |
|
|
&logmsg("stats: $drop $taskdir @{[&formatElapsedTime($self->{STARTTIME})]} success");
|
| 348 |
|
|
}
|
