2 # -*- mode: perl; perl-indent-level: 2; -*-
6 crunch-job: Execute job steps, save snapshots as requested, collate output.
10 Obtain job details from Arvados, run tasks on compute nodes (typically
11 invoked by scheduler on controller):
13 crunch-job --job x-y-z
15 Obtain job details from command line, run tasks on local machine
16 (typically invoked by application or developer on VM):
18 crunch-job --job '{"script_version":"/path/to/tree","script":"scriptname",...}'
26 If the job is already locked, steal the lock and run it anyway.
30 Path to .git directory where the specified commit is found.
34 Arvados API authorization token to use during the course of the job.
38 =head1 RUNNING JOBS LOCALLY
40 crunch-job's log messages appear on stderr along with the job tasks'
41 stderr streams. The log is saved in Keep at each checkpoint and when
44 If the job succeeds, the job's output locator is printed on stdout.
46 While the job is running, the following signals are accepted:
50 =item control-C, SIGINT, SIGQUIT
52 Save a checkpoint, terminate any job tasks that are running, and stop.
56 Save a checkpoint and continue.
60 Refresh node allocation (i.e., check whether any nodes have been added
61 or unallocated). Currently this is a no-op.
69 use POSIX ':sys_wait_h';
70 use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
74 use Warehouse::Stream;
75 use IPC::System::Simple qw(capturex);
77 $ENV{"TMPDIR"} ||= "/tmp";
78 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
79 if ($ENV{"USER"} ne "crunch" && $< != 0) {
80 # use a tmp dir unique for my uid
81 $ENV{"CRUNCH_TMP"} .= "-$<";
83 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
84 $ENV{"CRUNCH_TMP"} = $ENV{"JOB_WORK"}; # deprecated
85 mkdir ($ENV{"JOB_WORK"});
92 GetOptions('force-unlock' => \$force_unlock,
93 'git-dir=s' => \$git_dir,
95 'job-api-token=s' => \$job_api_token,
96 'resume-stash=s' => \$resume_stash,
99 if (defined $job_api_token) {
100 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
103 my $have_slurm = exists $ENV{SLURM_JOBID} && exists $ENV{SLURM_NODELIST};
104 my $job_has_uuid = $jobspec =~ /^[-a-z\d]+$/;
105 my $local_job = !$job_has_uuid;
114 $main::ENV{CRUNCH_DEBUG} = 1;
118 $main::ENV{CRUNCH_DEBUG} = 0;
123 my $arv = Arvados->new;
124 my $metastream = Warehouse::Stream->new(whc => new Warehouse);
126 $metastream->write_start('log.txt');
128 my $User = $arv->{'users'}->{'current'}->execute;
136 $Job = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
137 if (!$force_unlock) {
138 if ($Job->{'is_locked_by_uuid'}) {
139 croak("Job is locked: " . $Job->{'is_locked_by_uuid'});
141 if ($Job->{'success'} ne undef) {
142 croak("Job 'success' flag (" . $Job->{'success'} . ") is not null");
144 if ($Job->{'running'}) {
145 croak("Job 'running' flag is already set");
147 if ($Job->{'started_at'}) {
148 croak("Job 'started_at' time is already set (" . $Job->{'started_at'} . ")");
154 $Job = JSON::decode_json($jobspec);
158 map { croak ("No $_ specified") unless $Job->{$_} }
159 qw(script script_version script_parameters);
162 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
163 $Job->{'started_at'} = gmtime;
165 $Job = $arv->{'jobs'}->{'create'}->execute('job' => $Job);
169 $job_id = $Job->{'uuid'};
173 $Job->{'resource_limits'} ||= {};
174 $Job->{'resource_limits'}->{'max_tasks_per_node'} ||= 0;
175 my $max_ncpus = $Job->{'resource_limits'}->{'max_tasks_per_node'};
178 Log (undef, "check slurm allocation");
181 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
185 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
186 push @sinfo, "$localcpus localhost";
188 if (exists $ENV{SLURM_NODELIST})
190 push @sinfo, `sinfo -h --format='%c %N' --nodes='$ENV{SLURM_NODELIST}'`;
194 my ($ncpus, $slurm_nodelist) = split;
195 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
198 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
201 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
204 foreach (split (",", $ranges))
217 push @nodelist, map {
219 $n =~ s/\[[-,\d]+\]/$_/;
226 push @nodelist, $nodelist;
229 foreach my $nodename (@nodelist)
231 Log (undef, "node $nodename - $ncpus slots");
232 my $node = { name => $nodename,
236 foreach my $cpu (1..$ncpus)
238 push @slot, { node => $node,
242 push @node, @nodelist;
247 # Ensure that we get one jobstep running on each allocated node before
248 # we start overloading nodes with concurrent steps
250 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
257 # Claim this job, and make sure nobody else does
259 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
260 $Job->{'started_at'} = gmtime;
261 $Job->{'running'} = 1;
262 $Job->{'success'} = undef;
263 $Job->{'tasks_summary'} = { 'failed' => 0,
268 unless ($Job->save() && $Job->{'is_locked_by_uuid'} == $User->{'uuid'}) {
269 croak("Error while updating / locking job");
275 Log (undef, "start");
276 $SIG{'INT'} = sub { $main::please_freeze = 1; };
277 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
278 $SIG{'TERM'} = \&croak;
279 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
280 $SIG{'ALRM'} = sub { $main::please_info = 1; };
281 $SIG{'CONT'} = sub { $main::please_continue = 1; };
282 $main::please_freeze = 0;
283 $main::please_info = 0;
284 $main::please_continue = 0;
285 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
287 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
288 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
289 $ENV{"JOB_UUID"} = $job_id;
293 my @jobstep_todo = ();
294 my @jobstep_done = ();
295 my @jobstep_tomerge = ();
296 my $jobstep_tomerge_level = 0;
298 my $squeue_kill_checked;
299 my $output_in_keep = 0;
303 if (defined $Job->{thawedfromkey})
305 thaw ($Job->{thawedfromkey});
309 my $first_task = $arv->{'job_tasks'}->{'create'}->execute('job_task' => {
310 'job_uuid' => $Job->{'uuid'},
315 push @jobstep, { 'level' => 0,
317 'arvados_task' => $first_task,
319 push @jobstep_todo, 0;
326 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
328 my $skip_install = ($local_job && $Job->{script_version} =~ m{^/});
331 $ENV{"CRUNCH_SRC"} = $Job->{script_version};
337 $build_script = <DATA>;
339 Log (undef, "Install revision ".$Job->{script_version});
340 my $nodelist = join(",", @node);
342 # Clean out crunch_tmp/work and crunch_tmp/opt
344 my $cleanpid = fork();
347 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
348 ['bash', '-c', 'if mount | grep -q $JOB_WORK/; then sudo /bin/umount $JOB_WORK/* 2>/dev/null; fi; sleep 1; rm -rf $JOB_WORK $CRUNCH_TMP/opt']);
353 last if $cleanpid == waitpid (-1, WNOHANG);
354 freeze_if_want_freeze ($cleanpid);
355 select (undef, undef, undef, 0.1);
357 Log (undef, "Clean-work-dir exited $?");
359 # Install requested code version
362 my @srunargs = ("srun",
363 "--nodelist=$nodelist",
364 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
366 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
367 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
368 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
372 my $treeish = $Job->{'script_version'};
373 my $repo = $git_dir || $ENV{'CRUNCH_DEFAULT_GIT_DIR'};
374 # Todo: let script_version specify repository instead of expecting
375 # parent process to figure it out.
376 $ENV{"CRUNCH_SRC_URL"} = $repo;
378 # Create/update our clone of the remote git repo
380 if (!-d $ENV{"CRUNCH_SRC"}) {
381 system(qw(git clone), $repo, $ENV{"CRUNCH_SRC"}) == 0
382 or croak ("git clone $repo failed: exit ".($?>>8));
383 system("cd $ENV{CRUNCH_SRC} && git config clean.requireForce false");
385 `cd $ENV{CRUNCH_SRC} && git remote set-url origin \"\$CRUNCH_SRC_URL\" && git fetch -q origin`;
387 # If this looks like a subversion r#, look for it in git-svn commit messages
389 if ($treeish =~ m{^\d{1,4}$}) {
390 my $gitlog = `cd $ENV{CRUNCH_SRC} && git log --pretty="format:%H" --grep="git-svn-id:.*\@$treeish " origin/master`;
392 if ($gitlog =~ /^[a-f0-9]{40}$/) {
394 Log (undef, "Using commit $commit for script_version $treeish");
398 # If that didn't work, try asking git to look it up as a tree-ish.
400 if (!defined $commit) {
402 my $cooked_treeish = $treeish;
403 if ($treeish !~ m{^[0-9a-f]{5,}$}) {
404 # Looks like a git branch name -- make sure git knows it's
405 # relative to the remote repo
406 $cooked_treeish = "origin/$treeish";
409 my $found = `cd $ENV{CRUNCH_SRC} && git rev-list -1 $cooked_treeish`;
411 if ($found =~ /^[0-9a-f]{40}$/s) {
413 if ($commit ne $treeish) {
414 # Make sure we record the real commit id in the database,
415 # frozentokey, logs, etc. -- instead of an abbreviation or a
416 # branch name which can become ambiguous or point to a
417 # different commit in the future.
418 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
419 Log (undef, "Using commit $commit for tree-ish $treeish");
420 if ($commit ne $treeish) {
421 $Job->{'script_version'} = $commit;
422 !$job_has_uuid or $Job->save() or croak("Error while updating job");
428 if (defined $commit) {
429 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
430 @execargs = ("sh", "-c",
431 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
432 $git_archive = `cd $ENV{CRUNCH_SRC} && git archive $commit`;
435 croak ("could not figure out commit id for $treeish");
438 my $installpid = fork();
439 if ($installpid == 0)
441 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
446 last if $installpid == waitpid (-1, WNOHANG);
447 freeze_if_want_freeze ($installpid);
448 select (undef, undef, undef, 0.1);
450 Log (undef, "Install exited $?");
455 foreach (qw (script script_version script_parameters resource_limits))
459 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
461 foreach (split (/\n/, $Job->{knobs}))
463 Log (undef, "knob " . $_);
474 my $thisround_succeeded = 0;
475 my $thisround_failed = 0;
476 my $thisround_failed_multiple = 0;
478 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
479 or $a <=> $b } @jobstep_todo;
480 my $level = $jobstep[$jobstep_todo[0]]->{level};
481 Log (undef, "start level $level");
486 my @freeslot = (0..$#slot);
489 my $progress_is_dirty = 1;
490 my $progress_stats_updated = 0;
492 update_progress_stats();
497 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
499 $main::please_continue = 0;
501 my $id = $jobstep_todo[$todo_ptr];
502 my $Jobstep = $jobstep[$id];
503 if ($Jobstep->{level} != $level)
507 if ($Jobstep->{attempts} > 9)
509 Log ($id, "jobstep $id failed $$Jobstep{attempts} times -- giving up");
514 pipe $reader{$id}, "writer" or croak ($!);
515 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
516 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
518 my $childslot = $freeslot[0];
519 my $childnode = $slot[$childslot]->{node};
520 my $childslotname = join (".",
521 $slot[$childslot]->{node}->{name},
522 $slot[$childslot]->{cpu});
523 my $childpid = fork();
526 $SIG{'INT'} = 'DEFAULT';
527 $SIG{'QUIT'} = 'DEFAULT';
528 $SIG{'TERM'} = 'DEFAULT';
530 foreach (values (%reader))
534 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
535 open(STDOUT,">&writer");
536 open(STDERR,">&writer");
541 delete $ENV{"GNUPGHOME"};
542 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
543 $ENV{"TASK_QSEQUENCE"} = $id;
544 $ENV{"TASK_SEQUENCE"} = $level;
545 $ENV{"JOB_SCRIPT"} = $Job->{script};
546 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
547 $param =~ tr/a-z/A-Z/;
548 $ENV{"JOB_PARAMETER_$param"} = $value;
550 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
551 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
552 $ENV{"TASK_WORK"} = $ENV{"JOB_WORK"}."/".$slot[$childslot]->{cpu};
553 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
554 $ENV{"CRUNCH_NODE_SLOTS"} = $slot[$childslot]->{node}->{ncpus};
560 "--nodelist=".$childnode->{name},
561 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
562 "--job-name=$job_id.$id.$$",
564 my @execargs = qw(sh);
565 my $build_script_to_send = "";
567 "mkdir -p $ENV{JOB_WORK} $ENV{CRUNCH_TMP} $ENV{TASK_WORK} "
568 ."&& cd $ENV{CRUNCH_TMP} ";
571 $build_script_to_send = $build_script;
575 $ENV{"PYTHONPATH"} =~ s{^}{:} if $ENV{"PYTHONPATH"};
576 $ENV{"PYTHONPATH"} =~ s{^}{$ENV{CRUNCH_SRC}/sdk/python}; # xxx hack
578 "&& exec $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
579 my @execargs = ('bash', '-c', $command);
580 srun (\@srunargs, \@execargs, undef, $build_script_to_send);
584 if (!defined $childpid)
591 $proc{$childpid} = { jobstep => $id,
594 jobstepname => "$job_id.$id.$childpid",
596 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
597 $slot[$childslot]->{pid} = $childpid;
599 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
600 Log ($id, "child $childpid started on $childslotname");
601 $Jobstep->{attempts} ++;
602 $Jobstep->{starttime} = time;
603 $Jobstep->{node} = $childnode->{name};
604 $Jobstep->{slotindex} = $childslot;
605 delete $Jobstep->{stderr};
606 delete $Jobstep->{finishtime};
608 splice @jobstep_todo, $todo_ptr, 1;
611 $progress_is_dirty = 1;
615 (@slot > @freeslot && $todo_ptr+1 > $#jobstep_todo))
617 last THISROUND if $main::please_freeze;
618 if ($main::please_info)
620 $main::please_info = 0;
624 update_progress_stats();
632 update_progress_stats();
633 select (undef, undef, undef, 0.1);
635 elsif (time - $progress_stats_updated >= 30)
637 update_progress_stats();
639 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
640 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
642 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
643 .($thisround_failed+$thisround_succeeded)
644 .") -- giving up on this round";
645 Log (undef, $message);
649 # move slots from freeslot to holdslot (or back to freeslot) if necessary
650 for (my $i=$#freeslot; $i>=0; $i--) {
651 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
652 push @holdslot, (splice @freeslot, $i, 1);
655 for (my $i=$#holdslot; $i>=0; $i--) {
656 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
657 push @freeslot, (splice @holdslot, $i, 1);
661 # give up if no nodes are succeeding
662 if (!grep { $_->{node}->{losing_streak} == 0 &&
663 $_->{node}->{hold_count} < 4 } @slot) {
664 my $message = "Every node has failed -- giving up on this round";
665 Log (undef, $message);
672 push @freeslot, splice @holdslot;
673 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
676 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
679 goto THISROUND if $main::please_continue;
680 $main::please_info = 0, freeze(), collate_output(), save_meta(1) if $main::please_info;
685 update_progress_stats();
686 select (undef, undef, undef, 0.1);
687 killem (keys %proc) if $main::please_freeze;
691 update_progress_stats();
692 freeze_if_want_freeze();
695 if (!defined $success)
698 $thisround_succeeded == 0 &&
699 ($thisround_failed == 0 || $thisround_failed > 4))
701 my $message = "stop because $thisround_failed tasks failed and none succeeded";
702 Log (undef, $message);
711 goto ONELEVEL if !defined $success;
714 release_allocation();
717 $Job->{'output'} = &collate_output();
718 $Job->{'running'} = 0;
719 $Job->{'success'} = $Job->{'output'} && $success;
720 $Job->{'finished_at'} = gmtime;
721 $Job->save if $job_has_uuid;
723 if ($Job->{'output'})
726 my $manifest_text = capturex("whget", $Job->{'output'});
727 $arv->{'collections'}->{'create'}->execute('collection' => {
728 'uuid' => $Job->{'output'},
729 'manifest_text' => $manifest_text,
733 Log (undef, "Failed to register output manifest: $@");
737 Log (undef, "finish");
744 sub update_progress_stats
746 $progress_stats_updated = time;
747 return if !$progress_is_dirty;
748 my ($todo, $done, $running) = (scalar @jobstep_todo,
749 scalar @jobstep_done,
750 scalar @slot - scalar @freeslot - scalar @holdslot);
751 $Job->{'tasks_summary'} ||= {};
752 $Job->{'tasks_summary'}->{'todo'} = $todo;
753 $Job->{'tasks_summary'}->{'done'} = $done;
754 $Job->{'tasks_summary'}->{'running'} = $running;
755 $Job->save if $job_has_uuid;
756 Log (undef, "status: $done done, $running running, $todo todo");
757 $progress_is_dirty = 0;
764 my $pid = waitpid (-1, WNOHANG);
765 return 0 if $pid <= 0;
767 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
769 . $slot[$proc{$pid}->{slot}]->{cpu});
770 my $jobstepid = $proc{$pid}->{jobstep};
771 my $elapsed = time - $proc{$pid}->{time};
772 my $Jobstep = $jobstep[$jobstepid];
775 my $exitinfo = "exit $exitcode";
776 $Jobstep->{'arvados_task'}->reload;
777 my $success = $Jobstep->{'arvados_task'}->{success};
779 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$success");
781 if (!defined $success) {
782 # task did not indicate one way or the other --> fail
783 $Jobstep->{'arvados_task'}->{success} = 0;
784 $Jobstep->{'arvados_task'}->save;
790 my $no_incr_attempts;
791 $no_incr_attempts = 1 if $Jobstep->{node_fail};
794 ++$thisround_failed_multiple if $Jobstep->{attempts} > 1;
796 # Check for signs of a failed or misconfigured node
797 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
798 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
799 # Don't count this against jobstep failure thresholds if this
800 # node is already suspected faulty and srun exited quickly
801 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
803 $Jobstep->{attempts} > 1) {
804 Log ($jobstepid, "blaming failure on suspect node " . $slot[$proc{$pid}->{slot}]->{node}->{name} . " instead of incrementing jobstep attempts");
805 $no_incr_attempts = 1;
806 --$Jobstep->{attempts};
808 ban_node_by_slot($proc{$pid}->{slot});
811 push @jobstep_todo, $jobstepid;
812 Log ($jobstepid, "failure in $elapsed seconds");
814 --$Jobstep->{attempts} if $no_incr_attempts;
815 $Job->{'tasks_summary'}->{'failed'}++;
819 ++$thisround_succeeded;
820 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
821 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
822 push @jobstep_done, $jobstepid;
823 Log ($jobstepid, "success in $elapsed seconds");
825 $Jobstep->{exitcode} = $exitcode;
826 $Jobstep->{finishtime} = time;
827 process_stderr ($jobstepid, $success);
828 Log ($jobstepid, "output " . $Jobstep->{'arvados_task'}->{output});
830 close $reader{$jobstepid};
831 delete $reader{$jobstepid};
832 delete $slot[$proc{$pid}->{slot}]->{pid};
833 push @freeslot, $proc{$pid}->{slot};
837 my $newtask_list = $arv->{'job_tasks'}->{'list'}->execute(
839 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
841 'order' => 'qsequence'
843 foreach my $arvados_task (@{$newtask_list->{'items'}}) {
845 'level' => $arvados_task->{'sequence'},
847 'arvados_task' => $arvados_task
849 push @jobstep, $jobstep;
850 push @jobstep_todo, $#jobstep;
853 $progress_is_dirty = 1;
860 # return if the kill list was checked <4 seconds ago
861 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
865 $squeue_kill_checked = time;
867 # use killem() on procs whose killtime is reached
870 if (exists $proc{$_}->{killtime}
871 && $proc{$_}->{killtime} <= time)
877 # return if the squeue was checked <60 seconds ago
878 if (defined $squeue_checked && $squeue_checked > time - 60)
882 $squeue_checked = time;
886 # here is an opportunity to check for mysterious problems with local procs
890 # get a list of steps still running
891 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
893 if ($squeue[-1] ne "ok")
899 # which of my jobsteps are running, according to squeue?
903 if (/^(\d+)\.(\d+) (\S+)/)
905 if ($1 eq $ENV{SLURM_JOBID})
912 # which of my active child procs (>60s old) were not mentioned by squeue?
915 if ($proc{$_}->{time} < time - 60
916 && !exists $ok{$proc{$_}->{jobstepname}}
917 && !exists $proc{$_}->{killtime})
919 # kill this proc if it hasn't exited in 30 seconds
920 $proc{$_}->{killtime} = time + 30;
926 sub release_allocation
930 Log (undef, "release job allocation");
931 system "scancel $ENV{SLURM_JOBID}";
939 foreach my $job (keys %reader)
942 while (0 < sysread ($reader{$job}, $buf, 8192))
944 print STDERR $buf if $ENV{CRUNCH_DEBUG};
945 $jobstep[$job]->{stderr} .= $buf;
946 preprocess_stderr ($job);
947 if (length ($jobstep[$job]->{stderr}) > 16384)
949 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
958 sub preprocess_stderr
962 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
964 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
965 Log ($job, "stderr $line");
966 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOBID} has expired|Unable to confirm allocation for job) /) {
968 $main::please_freeze = 1;
970 elsif ($line =~ /srun: error: (Node failure on|Unable to create job step) /) {
971 $jobstep[$job]->{node_fail} = 1;
972 ban_node_by_slot($jobstep[$job]->{slotindex});
982 preprocess_stderr ($job);
985 Log ($job, "stderr $_");
986 } split ("\n", $jobstep[$job]->{stderr});
992 my $whc = Warehouse->new;
993 Log (undef, "collate");
994 $whc->write_start (1);
998 next if (!exists $_->{'arvados_task'}->{output} ||
999 !$_->{'arvados_task'}->{'success'} ||
1000 $_->{'exitcode'} != 0);
1001 my $output = $_->{'arvados_task'}->{output};
1002 if ($output !~ /^[0-9a-f]{32}(\+\S+)*$/)
1004 $output_in_keep ||= $output =~ / [0-9a-f]{32}\S*\+K/;
1005 $whc->write_data ($output);
1007 elsif (@jobstep == 1)
1009 $joboutput = $output;
1012 elsif (defined (my $outblock = $whc->fetch_block ($output)))
1014 $output_in_keep ||= $outblock =~ / [0-9a-f]{32}\S*\+K/;
1015 $whc->write_data ($outblock);
1019 my $errstr = $whc->errstr;
1020 $whc->write_data ("XXX fetch_block($output) failed: $errstr XXX\n");
1024 $joboutput = $whc->write_finish if !defined $joboutput;
1027 Log (undef, "output $joboutput");
1028 $Job->{'output'} = $joboutput;
1029 $Job->save if $job_has_uuid;
1033 Log (undef, "output undef");
1043 my $sig = 2; # SIGINT first
1044 if (exists $proc{$_}->{"sent_$sig"} &&
1045 time - $proc{$_}->{"sent_$sig"} > 4)
1047 $sig = 15; # SIGTERM if SIGINT doesn't work
1049 if (exists $proc{$_}->{"sent_$sig"} &&
1050 time - $proc{$_}->{"sent_$sig"} > 4)
1052 $sig = 9; # SIGKILL if SIGTERM doesn't work
1054 if (!exists $proc{$_}->{"sent_$sig"})
1056 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1058 select (undef, undef, undef, 0.1);
1061 kill $sig, $_; # srun wants two SIGINT to really interrupt
1063 $proc{$_}->{"sent_$sig"} = time;
1064 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1074 vec($bits,fileno($_),1) = 1;
1080 sub Log # ($jobstep_id, $logmessage)
1082 if ($_[1] =~ /\n/) {
1083 for my $line (split (/\n/, $_[1])) {
1088 my $fh = select STDERR; $|=1; select $fh;
1089 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1090 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1093 if ($metastream || -t STDERR) {
1094 my @gmtime = gmtime;
1095 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1096 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1098 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1100 return if !$metastream;
1101 $metastream->write_data ($datetime . " " . $message);
1107 my ($package, $file, $line) = caller;
1108 my $message = "@_ at $file line $line\n";
1109 Log (undef, $message);
1110 freeze() if @jobstep_todo;
1111 collate_output() if @jobstep_todo;
1113 save_meta() if $metastream;
1120 return if !$job_has_uuid;
1122 $Job->{'running'} = 0;
1123 $Job->{'success'} = 0;
1124 $Job->{'finished_at'} = gmtime;
1131 my $justcheckpoint = shift; # false if this will be the last meta saved
1132 my $m = $metastream;
1133 $m = $m->copy if $justcheckpoint;
1135 my $loglocator = $m->as_key;
1136 undef $metastream if !$justcheckpoint; # otherwise Log() will try to use it
1137 Log (undef, "meta key is $loglocator");
1138 $Job->{'log'} = $loglocator;
1139 $Job->save if $job_has_uuid;
1143 sub freeze_if_want_freeze
1145 if ($main::please_freeze)
1147 release_allocation();
1150 # kill some srun procs before freeze+stop
1151 map { $proc{$_} = {} } @_;
1154 killem (keys %proc);
1155 select (undef, undef, undef, 0.1);
1157 while (($died = waitpid (-1, WNOHANG)) > 0)
1159 delete $proc{$died};
1174 Log (undef, "Freeze not implemented");
1181 croak ("Thaw not implemented");
1185 Log (undef, "thaw from $key");
1190 @jobstep_tomerge = ();
1191 $jobstep_tomerge_level = 0;
1194 my $stream = new Warehouse::Stream ( whc => $whc,
1195 hash => [split (",", $key)] );
1197 while (my $dataref = $stream->read_until (undef, "\n\n"))
1199 if ($$dataref =~ /^job /)
1201 foreach (split ("\n", $$dataref))
1203 my ($k, $v) = split ("=", $_, 2);
1204 $frozenjob->{$k} = freezeunquote ($v);
1209 if ($$dataref =~ /^merge (\d+) (.*)/)
1211 $jobstep_tomerge_level = $1;
1213 = map { freezeunquote ($_) } split ("\n", freezeunquote($2));
1218 foreach (split ("\n", $$dataref))
1220 my ($k, $v) = split ("=", $_, 2);
1221 $Jobstep->{$k} = freezeunquote ($v) if $k;
1223 $Jobstep->{attempts} = 0;
1224 push @jobstep, $Jobstep;
1226 if ($Jobstep->{exitcode} eq "0")
1228 push @jobstep_done, $#jobstep;
1232 push @jobstep_todo, $#jobstep;
1236 foreach (qw (script script_version script_parameters))
1238 $Job->{$_} = $frozenjob->{$_};
1240 $Job->save if $job_has_uuid;
1256 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1263 my $srunargs = shift;
1264 my $execargs = shift;
1265 my $opts = shift || {};
1267 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1268 print STDERR (join (" ",
1269 map { / / ? "'$_'" : $_ }
1272 if $ENV{CRUNCH_DEBUG};
1274 if (defined $stdin) {
1275 my $child = open STDIN, "-|";
1276 defined $child or die "no fork: $!";
1278 print $stdin or die $!;
1279 close STDOUT or die $!;
1284 return system (@$args) if $opts->{fork};
1287 warn "ENV size is ".length(join(" ",%ENV));
1288 die "exec failed: $!: @$args";
1292 sub ban_node_by_slot {
1293 # Don't start any new jobsteps on this node for 60 seconds
1295 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1296 $slot[$slotid]->{node}->{hold_count}++;
1297 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1303 # checkout-and-build
1307 my $destdir = $ENV{"CRUNCH_SRC"};
1308 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1309 my $repo = $ENV{"CRUNCH_SRC_URL"};
1311 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1313 if (readlink ("$destdir.commit") eq $commit) {
1317 unlink "$destdir.commit";
1318 open STDOUT, ">", "$destdir.log";
1319 open STDERR, ">&STDOUT";
1322 open TARX, "|-", "tar", "-C", $destdir, "-xf", "-";
1325 die "'tar -C $destdir -xf -' exited $?: $!";
1329 chomp ($pwd = `pwd`);
1330 my $install_dir = $ENV{"CRUNCH_INSTALL"} || "$pwd/opt";
1332 if (!-e "./install.sh" && -e "./tests/autotests.sh") {
1334 shell_or_die ("./tests/autotests.sh", $install_dir);
1335 } elsif (-e "./install.sh") {
1336 shell_or_die ("./install.sh", $install_dir);
1340 unlink "$destdir.commit.new";
1341 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
1342 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
1351 if ($ENV{"DEBUG"}) {
1352 print STDERR "@_\n";
1355 or die "@_ failed: $! exit 0x".sprintf("%x",$?);