2 # -*- mode: perl; perl-indent-level: 2; indent-tabs-mode: nil; -*-
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 Do not clear per-job/task temporary directories during initial job
39 setup. This can speed up development and debugging when running jobs
44 =head1 RUNNING JOBS LOCALLY
46 crunch-job's log messages appear on stderr along with the job tasks'
47 stderr streams. The log is saved in Keep at each checkpoint and when
50 If the job succeeds, the job's output locator is printed on stdout.
52 While the job is running, the following signals are accepted:
56 =item control-C, SIGINT, SIGQUIT
58 Save a checkpoint, terminate any job tasks that are running, and stop.
62 Save a checkpoint and continue.
66 Refresh node allocation (i.e., check whether any nodes have been added
67 or unallocated) and attributes of the Job record that should affect
68 behavior (e.g., cancel job if cancelled_at becomes non-nil).
76 use POSIX ':sys_wait_h';
77 use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
85 $ENV{"TMPDIR"} ||= "/tmp";
86 unless (defined $ENV{"CRUNCH_TMP"}) {
87 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
88 if ($ENV{"USER"} ne "crunch" && $< != 0) {
89 # use a tmp dir unique for my uid
90 $ENV{"CRUNCH_TMP"} .= "-$<";
93 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
94 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
95 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
96 mkdir ($ENV{"JOB_WORK"});
100 if (defined $ENV{"ARV_CLI"}) {
101 $arv_cli = $ENV{"ARV_CLI"};
113 GetOptions('force-unlock' => \$force_unlock,
114 'git-dir=s' => \$git_dir,
115 'job=s' => \$jobspec,
116 'job-api-token=s' => \$job_api_token,
117 'no-clear-tmp' => \$no_clear_tmp,
118 'resume-stash=s' => \$resume_stash,
121 if (defined $job_api_token) {
122 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
125 my $have_slurm = exists $ENV{SLURM_JOBID} && exists $ENV{SLURM_NODELIST};
126 my $job_has_uuid = $jobspec =~ /^[-a-z\d]+$/;
127 my $local_job = !$job_has_uuid;
132 $main::ENV{CRUNCH_DEBUG} = 1;
136 $main::ENV{CRUNCH_DEBUG} = 0;
141 my $arv = Arvados->new('apiVersion' => 'v1');
144 my $User = $arv->{'users'}->{'current'}->execute;
152 $Job = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
153 if (!$force_unlock) {
154 if ($Job->{'is_locked_by_uuid'}) {
155 croak("Job is locked: " . $Job->{'is_locked_by_uuid'});
157 if ($Job->{'success'} ne undef) {
158 croak("Job 'success' flag (" . $Job->{'success'} . ") is not null");
160 if ($Job->{'running'}) {
161 croak("Job 'running' flag is already set");
163 if ($Job->{'started_at'}) {
164 croak("Job 'started_at' time is already set (" . $Job->{'started_at'} . ")");
170 $Job = JSON::decode_json($jobspec);
174 map { croak ("No $_ specified") unless $Job->{$_} }
175 qw(script script_version script_parameters);
178 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
179 $Job->{'started_at'} = gmtime;
181 $Job = $arv->{'jobs'}->{'create'}->execute('job' => $Job);
185 $job_id = $Job->{'uuid'};
187 my $keep_logfile = $job_id . '.log.txt';
188 my $local_logfile = File::Temp->new();
190 $Job->{'runtime_constraints'} ||= {};
191 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
192 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
195 Log (undef, "check slurm allocation");
198 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
202 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
203 push @sinfo, "$localcpus localhost";
205 if (exists $ENV{SLURM_NODELIST})
207 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
211 my ($ncpus, $slurm_nodelist) = split;
212 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
215 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
218 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
221 foreach (split (",", $ranges))
234 push @nodelist, map {
236 $n =~ s/\[[-,\d]+\]/$_/;
243 push @nodelist, $nodelist;
246 foreach my $nodename (@nodelist)
248 Log (undef, "node $nodename - $ncpus slots");
249 my $node = { name => $nodename,
253 foreach my $cpu (1..$ncpus)
255 push @slot, { node => $node,
259 push @node, @nodelist;
264 # Ensure that we get one jobstep running on each allocated node before
265 # we start overloading nodes with concurrent steps
267 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
274 # Claim this job, and make sure nobody else does
275 unless ($Job->update_attributes('is_locked_by_uuid' => $User->{'uuid'}) &&
276 $Job->{'is_locked_by_uuid'} == $User->{'uuid'}) {
277 croak("Error while updating / locking job");
279 $Job->update_attributes('started_at' => scalar gmtime,
282 'tasks_summary' => { 'failed' => 0,
289 Log (undef, "start");
290 $SIG{'INT'} = sub { $main::please_freeze = 1; };
291 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
292 $SIG{'TERM'} = \&croak;
293 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
294 $SIG{'ALRM'} = sub { $main::please_info = 1; };
295 $SIG{'CONT'} = sub { $main::please_continue = 1; };
296 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
298 $main::please_freeze = 0;
299 $main::please_info = 0;
300 $main::please_continue = 0;
301 $main::please_refresh = 0;
302 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
304 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
305 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
306 $ENV{"JOB_UUID"} = $job_id;
310 my @jobstep_todo = ();
311 my @jobstep_done = ();
312 my @jobstep_tomerge = ();
313 my $jobstep_tomerge_level = 0;
315 my $squeue_kill_checked;
316 my $output_in_keep = 0;
317 my $latest_refresh = scalar time;
321 if (defined $Job->{thawedfromkey})
323 thaw ($Job->{thawedfromkey});
327 my $first_task = $arv->{'job_tasks'}->{'create'}->execute('job_task' => {
328 'job_uuid' => $Job->{'uuid'},
333 push @jobstep, { 'level' => 0,
335 'arvados_task' => $first_task,
337 push @jobstep_todo, 0;
343 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
350 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
352 my $skip_install = ($local_job && $Job->{script_version} =~ m{^/});
355 if (!defined $no_clear_tmp) {
356 my $clear_tmp_cmd = 'rm -rf $JOB_WORK $CRUNCH_TMP/opt $CRUNCH_TMP/src*';
357 system($clear_tmp_cmd) == 0
358 or croak ("`$clear_tmp_cmd` failed: ".($?>>8));
360 $ENV{"CRUNCH_SRC"} = $Job->{script_version};
361 for my $src_path ("$ENV{CRUNCH_SRC}/arvados/sdk/python") {
363 system("virtualenv", "$ENV{CRUNCH_TMP}/opt") == 0
364 or croak ("virtualenv $ENV{CRUNCH_TMP}/opt failed: exit ".($?>>8));
365 system ("cd $src_path && ./build.sh && \$CRUNCH_TMP/opt/bin/python setup.py install")
367 or croak ("setup.py in $src_path failed: exit ".($?>>8));
375 $build_script = <DATA>;
377 Log (undef, "Install revision ".$Job->{script_version});
378 my $nodelist = join(",", @node);
380 if (!defined $no_clear_tmp) {
381 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
383 my $cleanpid = fork();
386 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
387 ['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 $CRUNCH_TMP/src*']);
392 last if $cleanpid == waitpid (-1, WNOHANG);
393 freeze_if_want_freeze ($cleanpid);
394 select (undef, undef, undef, 0.1);
396 Log (undef, "Clean-work-dir exited $?");
399 # Install requested code version
402 my @srunargs = ("srun",
403 "--nodelist=$nodelist",
404 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
406 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
407 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
411 my $treeish = $Job->{'script_version'};
413 # If we're running under crunch-dispatch, it will have pulled the
414 # appropriate source tree into its own repository, and given us that
415 # repo's path as $git_dir. If we're running a "local" job, and a
416 # script_version was specified, it's up to the user to provide the
417 # full path to a local repository in Job->{repository}.
419 # TODO: Accept URLs too, not just local paths. Use git-ls-remote and
420 # git-archive --remote where appropriate.
422 # TODO: Accept a locally-hosted Arvados repository by name or
423 # UUID. Use arvados.v1.repositories.list or .get to figure out the
424 # appropriate fetch-url.
425 my $repo = $git_dir || $ENV{'CRUNCH_DEFAULT_GIT_DIR'} || $Job->{'repository'};
427 $ENV{"CRUNCH_SRC_URL"} = $repo;
429 if (-d "$repo/.git") {
430 # We were given a working directory, but we are only interested in
432 $repo = "$repo/.git";
435 # If this looks like a subversion r#, look for it in git-svn commit messages
437 if ($treeish =~ m{^\d{1,4}$}) {
438 my $gitlog = `git --git-dir=\Q$repo\E log --pretty="format:%H" --grep="git-svn-id:.*\@"\Q$treeish\E" " master`;
440 if ($gitlog =~ /^[a-f0-9]{40}$/) {
442 Log (undef, "Using commit $commit for script_version $treeish");
446 # If that didn't work, try asking git to look it up as a tree-ish.
448 if (!defined $commit) {
449 my $found = `git --git-dir=\Q$repo\E rev-list -1 ''\Q$treeish\E`;
451 if ($found =~ /^[0-9a-f]{40}$/s) {
453 if ($commit ne $treeish) {
454 # Make sure we record the real commit id in the database,
455 # frozentokey, logs, etc. -- instead of an abbreviation or a
456 # branch name which can become ambiguous or point to a
457 # different commit in the future.
458 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
459 Log (undef, "Using commit $commit for tree-ish $treeish");
460 if ($commit ne $treeish) {
461 $Job->{'script_version'} = $commit;
463 $Job->update_attributes('script_version' => $commit) or
464 croak("Error while updating job");
470 if (defined $commit) {
471 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
472 @execargs = ("sh", "-c",
473 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
474 $git_archive = `git --git-dir=\Q$repo\E archive ''\Q$commit\E`;
477 croak ("could not figure out commit id for $treeish");
480 my $installpid = fork();
481 if ($installpid == 0)
483 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
488 last if $installpid == waitpid (-1, WNOHANG);
489 freeze_if_want_freeze ($installpid);
490 select (undef, undef, undef, 0.1);
492 Log (undef, "Install exited $?");
497 # Grab our lock again (we might have deleted and re-created CRUNCH_TMP above)
498 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
501 # If this job requires a Docker image, install that.
502 my $docker_bin = "/usr/bin/docker.io";
503 my $docker_image = $Job->{runtime_constraints}->{docker_image} || "";
505 my $docker_pid = fork();
506 if ($docker_pid == 0)
508 srun (["srun", "--nodelist=" . join(' ', @node)],
509 [$docker_bin, 'pull', $docker_image]);
514 last if $docker_pid == waitpid (-1, WNOHANG);
515 freeze_if_want_freeze ($docker_pid);
516 select (undef, undef, undef, 0.1);
518 # If the Docker image was specified as a hash, pull will fail.
519 # Ignore that error. We'll see what happens when we try to run later.
520 if (($? != 0) && ($docker_image !~ /^[0-9a-fA-F]{5,64}$/))
522 croak("Installing Docker image $docker_image returned exit code $?");
526 foreach (qw (script script_version script_parameters runtime_constraints))
530 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
532 foreach (split (/\n/, $Job->{knobs}))
534 Log (undef, "knob " . $_);
539 $main::success = undef;
545 my $thisround_succeeded = 0;
546 my $thisround_failed = 0;
547 my $thisround_failed_multiple = 0;
549 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
550 or $a <=> $b } @jobstep_todo;
551 my $level = $jobstep[$jobstep_todo[0]]->{level};
552 Log (undef, "start level $level");
557 my @freeslot = (0..$#slot);
560 my $progress_is_dirty = 1;
561 my $progress_stats_updated = 0;
563 update_progress_stats();
568 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
570 my $id = $jobstep_todo[$todo_ptr];
571 my $Jobstep = $jobstep[$id];
572 if ($Jobstep->{level} != $level)
577 pipe $reader{$id}, "writer" or croak ($!);
578 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
579 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
581 my $childslot = $freeslot[0];
582 my $childnode = $slot[$childslot]->{node};
583 my $childslotname = join (".",
584 $slot[$childslot]->{node}->{name},
585 $slot[$childslot]->{cpu});
586 my $childpid = fork();
589 $SIG{'INT'} = 'DEFAULT';
590 $SIG{'QUIT'} = 'DEFAULT';
591 $SIG{'TERM'} = 'DEFAULT';
593 foreach (values (%reader))
597 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
598 open(STDOUT,">&writer");
599 open(STDERR,">&writer");
604 delete $ENV{"GNUPGHOME"};
605 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
606 $ENV{"TASK_QSEQUENCE"} = $id;
607 $ENV{"TASK_SEQUENCE"} = $level;
608 $ENV{"JOB_SCRIPT"} = $Job->{script};
609 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
610 $param =~ tr/a-z/A-Z/;
611 $ENV{"JOB_PARAMETER_$param"} = $value;
613 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
614 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
615 $ENV{"TASK_WORK"} = $ENV{"JOB_WORK"}."/$id.$$";
616 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
617 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
618 $ENV{"CRUNCH_NODE_SLOTS"} = $slot[$childslot]->{node}->{ncpus};
619 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
625 "--nodelist=".$childnode->{name},
626 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
627 "--job-name=$job_id.$id.$$",
629 my $build_script_to_send = "";
631 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
632 ."mkdir -p $ENV{JOB_WORK} $ENV{CRUNCH_TMP} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
633 ."&& cd $ENV{CRUNCH_TMP} ";
636 $build_script_to_send = $build_script;
640 $command .= "&& exec arv-mount --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
643 $command .= "$docker_bin run -i -a stdin -a stdout -a stderr ";
644 # Dynamically configure the container to use the host system as its
645 # DNS server. Get the host's global addresses from the ip command,
646 # and turn them into docker --dns options using gawk.
648 q{$(ip -o address show scope global |
649 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
650 foreach my $env_key (qw(CRUNCH_SRC CRUNCH_TMP TASK_KEEPMOUNT))
652 $command .= "-v \Q$ENV{$env_key}:$ENV{$env_key}:rw\E ";
654 while (my ($env_key, $env_val) = each %ENV)
656 $command .= "-e \Q$env_key=$env_val\E ";
658 $command .= "$docker_image ";
660 $command .= "$ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
661 my @execargs = ('bash', '-c', $command);
662 srun (\@srunargs, \@execargs, undef, $build_script_to_send);
666 if (!defined $childpid)
673 $proc{$childpid} = { jobstep => $id,
676 jobstepname => "$job_id.$id.$childpid",
678 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
679 $slot[$childslot]->{pid} = $childpid;
681 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
682 Log ($id, "child $childpid started on $childslotname");
683 $Jobstep->{starttime} = time;
684 $Jobstep->{node} = $childnode->{name};
685 $Jobstep->{slotindex} = $childslot;
686 delete $Jobstep->{stderr};
687 delete $Jobstep->{finishtime};
689 splice @jobstep_todo, $todo_ptr, 1;
692 $progress_is_dirty = 1;
696 (@slot > @freeslot && $todo_ptr+1 > $#jobstep_todo))
698 last THISROUND if $main::please_freeze;
699 if ($main::please_info)
701 $main::please_info = 0;
705 update_progress_stats();
712 check_refresh_wanted();
714 update_progress_stats();
715 select (undef, undef, undef, 0.1);
717 elsif (time - $progress_stats_updated >= 30)
719 update_progress_stats();
721 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
722 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
724 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
725 .($thisround_failed+$thisround_succeeded)
726 .") -- giving up on this round";
727 Log (undef, $message);
731 # move slots from freeslot to holdslot (or back to freeslot) if necessary
732 for (my $i=$#freeslot; $i>=0; $i--) {
733 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
734 push @holdslot, (splice @freeslot, $i, 1);
737 for (my $i=$#holdslot; $i>=0; $i--) {
738 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
739 push @freeslot, (splice @holdslot, $i, 1);
743 # give up if no nodes are succeeding
744 if (!grep { $_->{node}->{losing_streak} == 0 &&
745 $_->{node}->{hold_count} < 4 } @slot) {
746 my $message = "Every node has failed -- giving up on this round";
747 Log (undef, $message);
754 push @freeslot, splice @holdslot;
755 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
758 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
761 if ($main::please_continue) {
762 $main::please_continue = 0;
765 $main::please_info = 0, freeze(), collate_output(), save_meta(1) if $main::please_info;
769 check_refresh_wanted();
771 update_progress_stats();
772 select (undef, undef, undef, 0.1);
773 killem (keys %proc) if $main::please_freeze;
777 update_progress_stats();
778 freeze_if_want_freeze();
781 if (!defined $main::success)
784 $thisround_succeeded == 0 &&
785 ($thisround_failed == 0 || $thisround_failed > 4))
787 my $message = "stop because $thisround_failed tasks failed and none succeeded";
788 Log (undef, $message);
797 goto ONELEVEL if !defined $main::success;
800 release_allocation();
803 $Job->update_attributes('output' => &collate_output(),
805 'success' => $Job->{'output'} && $main::success,
806 'finished_at' => scalar gmtime)
809 if ($Job->{'output'})
812 my $manifest_text = `arv keep get ''\Q$Job->{'output'}\E`;
813 $arv->{'collections'}->{'create'}->execute('collection' => {
814 'uuid' => $Job->{'output'},
815 'manifest_text' => $manifest_text,
817 if ($Job->{'output_is_persistent'}) {
818 $arv->{'links'}->{'create'}->execute('link' => {
819 'tail_kind' => 'arvados#user',
820 'tail_uuid' => $User->{'uuid'},
821 'head_kind' => 'arvados#collection',
822 'head_uuid' => $Job->{'output'},
823 'link_class' => 'resources',
829 Log (undef, "Failed to register output manifest: $@");
833 Log (undef, "finish");
840 sub update_progress_stats
842 $progress_stats_updated = time;
843 return if !$progress_is_dirty;
844 my ($todo, $done, $running) = (scalar @jobstep_todo,
845 scalar @jobstep_done,
846 scalar @slot - scalar @freeslot - scalar @holdslot);
847 $Job->{'tasks_summary'} ||= {};
848 $Job->{'tasks_summary'}->{'todo'} = $todo;
849 $Job->{'tasks_summary'}->{'done'} = $done;
850 $Job->{'tasks_summary'}->{'running'} = $running;
852 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
854 Log (undef, "status: $done done, $running running, $todo todo");
855 $progress_is_dirty = 0;
862 my $pid = waitpid (-1, WNOHANG);
863 return 0 if $pid <= 0;
865 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
867 . $slot[$proc{$pid}->{slot}]->{cpu});
868 my $jobstepid = $proc{$pid}->{jobstep};
869 my $elapsed = time - $proc{$pid}->{time};
870 my $Jobstep = $jobstep[$jobstepid];
872 my $childstatus = $?;
873 my $exitvalue = $childstatus >> 8;
874 my $exitinfo = sprintf("exit %d signal %d%s",
877 ($childstatus & 128 ? ' core dump' : ''));
878 $Jobstep->{'arvados_task'}->reload;
879 my $task_success = $Jobstep->{'arvados_task'}->{success};
881 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
883 if (!defined $task_success) {
884 # task did not indicate one way or the other --> fail
885 $Jobstep->{'arvados_task'}->{success} = 0;
886 $Jobstep->{'arvados_task'}->save;
893 $temporary_fail ||= $Jobstep->{node_fail};
894 $temporary_fail ||= ($exitvalue == 111);
897 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
899 # Check for signs of a failed or misconfigured node
900 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
901 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
902 # Don't count this against jobstep failure thresholds if this
903 # node is already suspected faulty and srun exited quickly
904 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
906 Log ($jobstepid, "blaming failure on suspect node " .
907 $slot[$proc{$pid}->{slot}]->{node}->{name});
908 $temporary_fail ||= 1;
910 ban_node_by_slot($proc{$pid}->{slot});
913 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
914 ++$Jobstep->{'failures'},
915 $temporary_fail ? 'temporary ' : 'permanent',
918 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
919 # Give up on this task, and the whole job
921 $main::please_freeze = 1;
924 # Put this task back on the todo queue
925 push @jobstep_todo, $jobstepid;
927 $Job->{'tasks_summary'}->{'failed'}++;
931 ++$thisround_succeeded;
932 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
933 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
934 push @jobstep_done, $jobstepid;
935 Log ($jobstepid, "success in $elapsed seconds");
937 $Jobstep->{exitcode} = $childstatus;
938 $Jobstep->{finishtime} = time;
939 process_stderr ($jobstepid, $task_success);
940 Log ($jobstepid, "output " . $Jobstep->{'arvados_task'}->{output});
942 close $reader{$jobstepid};
943 delete $reader{$jobstepid};
944 delete $slot[$proc{$pid}->{slot}]->{pid};
945 push @freeslot, $proc{$pid}->{slot};
949 my $newtask_list = $arv->{'job_tasks'}->{'list'}->execute(
951 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
953 'order' => 'qsequence'
955 foreach my $arvados_task (@{$newtask_list->{'items'}}) {
957 'level' => $arvados_task->{'sequence'},
959 'arvados_task' => $arvados_task
961 push @jobstep, $jobstep;
962 push @jobstep_todo, $#jobstep;
965 $progress_is_dirty = 1;
969 sub check_refresh_wanted
971 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
972 if (@stat && $stat[9] > $latest_refresh) {
973 $latest_refresh = scalar time;
975 my $Job2 = $arv->{'jobs'}->{'get'}->execute('uuid' => $jobspec);
976 for my $attr ('cancelled_at',
977 'cancelled_by_user_uuid',
978 'cancelled_by_client_uuid') {
979 $Job->{$attr} = $Job2->{$attr};
981 if ($Job->{'cancelled_at'}) {
982 Log (undef, "Job cancelled at " . $Job->{cancelled_at} .
983 " by user " . $Job->{cancelled_by_user_uuid});
985 $main::please_freeze = 1;
993 # return if the kill list was checked <4 seconds ago
994 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
998 $squeue_kill_checked = time;
1000 # use killem() on procs whose killtime is reached
1003 if (exists $proc{$_}->{killtime}
1004 && $proc{$_}->{killtime} <= time)
1010 # return if the squeue was checked <60 seconds ago
1011 if (defined $squeue_checked && $squeue_checked > time - 60)
1015 $squeue_checked = time;
1019 # here is an opportunity to check for mysterious problems with local procs
1023 # get a list of steps still running
1024 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
1026 if ($squeue[-1] ne "ok")
1032 # which of my jobsteps are running, according to squeue?
1036 if (/^(\d+)\.(\d+) (\S+)/)
1038 if ($1 eq $ENV{SLURM_JOBID})
1045 # which of my active child procs (>60s old) were not mentioned by squeue?
1046 foreach (keys %proc)
1048 if ($proc{$_}->{time} < time - 60
1049 && !exists $ok{$proc{$_}->{jobstepname}}
1050 && !exists $proc{$_}->{killtime})
1052 # kill this proc if it hasn't exited in 30 seconds
1053 $proc{$_}->{killtime} = time + 30;
1059 sub release_allocation
1063 Log (undef, "release job allocation");
1064 system "scancel $ENV{SLURM_JOBID}";
1072 foreach my $job (keys %reader)
1075 while (0 < sysread ($reader{$job}, $buf, 8192))
1077 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1078 $jobstep[$job]->{stderr} .= $buf;
1079 preprocess_stderr ($job);
1080 if (length ($jobstep[$job]->{stderr}) > 16384)
1082 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1091 sub preprocess_stderr
1095 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1097 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1098 Log ($job, "stderr $line");
1099 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1101 $main::please_freeze = 1;
1103 elsif ($line =~ /srun: error: (Node failure on|Unable to create job step) /) {
1104 $jobstep[$job]->{node_fail} = 1;
1105 ban_node_by_slot($jobstep[$job]->{slotindex});
1114 my $task_success = shift;
1115 preprocess_stderr ($job);
1118 Log ($job, "stderr $_");
1119 } split ("\n", $jobstep[$job]->{stderr});
1125 my ($keep, $child_out, $output_block);
1127 my $cmd = "$arv_cli keep get \Q$hash\E";
1128 open($keep, '-|', $cmd) or die "fetch_block: $cmd: $!";
1129 sysread($keep, $output_block, 64 * 1024 * 1024);
1131 return $output_block;
1136 Log (undef, "collate");
1138 my ($child_out, $child_in);
1139 my $pid = open2($child_out, $child_in, $arv_cli, 'keep', 'put', '--raw');
1143 next if (!exists $_->{'arvados_task'}->{output} ||
1144 !$_->{'arvados_task'}->{'success'} ||
1145 $_->{'exitcode'} != 0);
1146 my $output = $_->{'arvados_task'}->{output};
1147 if ($output !~ /^[0-9a-f]{32}(\+\S+)*$/)
1149 $output_in_keep ||= $output =~ / [0-9a-f]{32}\S*\+K/;
1150 print $child_in $output;
1152 elsif (@jobstep == 1)
1154 $joboutput = $output;
1157 elsif (defined (my $outblock = fetch_block ($output)))
1159 $output_in_keep ||= $outblock =~ / [0-9a-f]{32}\S*\+K/;
1160 print $child_in $outblock;
1164 Log (undef, "XXX fetch_block($output) failed XXX");
1170 if (!defined $joboutput) {
1171 my $s = IO::Select->new($child_out);
1172 if ($s->can_read(120)) {
1173 sysread($child_out, $joboutput, 64 * 1024 * 1024);
1176 Log (undef, "timed out reading from 'arv keep put'");
1183 Log (undef, "output $joboutput");
1184 $Job->update_attributes('output' => $joboutput) if $job_has_uuid;
1188 Log (undef, "output undef");
1198 my $sig = 2; # SIGINT first
1199 if (exists $proc{$_}->{"sent_$sig"} &&
1200 time - $proc{$_}->{"sent_$sig"} > 4)
1202 $sig = 15; # SIGTERM if SIGINT doesn't work
1204 if (exists $proc{$_}->{"sent_$sig"} &&
1205 time - $proc{$_}->{"sent_$sig"} > 4)
1207 $sig = 9; # SIGKILL if SIGTERM doesn't work
1209 if (!exists $proc{$_}->{"sent_$sig"})
1211 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1213 select (undef, undef, undef, 0.1);
1216 kill $sig, $_; # srun wants two SIGINT to really interrupt
1218 $proc{$_}->{"sent_$sig"} = time;
1219 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1229 vec($bits,fileno($_),1) = 1;
1235 sub Log # ($jobstep_id, $logmessage)
1237 if ($_[1] =~ /\n/) {
1238 for my $line (split (/\n/, $_[1])) {
1243 my $fh = select STDERR; $|=1; select $fh;
1244 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1245 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1248 if ($metastream || -t STDERR) {
1249 my @gmtime = gmtime;
1250 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1251 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1253 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1256 print $metastream $datetime . " " . $message;
1263 my ($package, $file, $line) = caller;
1264 my $message = "@_ at $file line $line\n";
1265 Log (undef, $message);
1266 freeze() if @jobstep_todo;
1267 collate_output() if @jobstep_todo;
1269 save_meta() if $metastream;
1276 return if !$job_has_uuid;
1277 $Job->update_attributes('running' => 0,
1279 'finished_at' => scalar gmtime);
1285 my $justcheckpoint = shift; # false if this will be the last meta saved
1286 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1288 $local_logfile->flush;
1289 my $cmd = "$arv_cli keep put --filename ''\Q$keep_logfile\E "
1290 . quotemeta($local_logfile->filename);
1291 my $loglocator = `$cmd`;
1292 die "system $cmd failed: $?" if $?;
1294 $local_logfile = undef; # the temp file is automatically deleted
1295 Log (undef, "log manifest is $loglocator");
1296 $Job->{'log'} = $loglocator;
1297 $Job->update_attributes('log', $loglocator) if $job_has_uuid;
1301 sub freeze_if_want_freeze
1303 if ($main::please_freeze)
1305 release_allocation();
1308 # kill some srun procs before freeze+stop
1309 map { $proc{$_} = {} } @_;
1312 killem (keys %proc);
1313 select (undef, undef, undef, 0.1);
1315 while (($died = waitpid (-1, WNOHANG)) > 0)
1317 delete $proc{$died};
1332 Log (undef, "Freeze not implemented");
1339 croak ("Thaw not implemented");
1355 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1362 my $srunargs = shift;
1363 my $execargs = shift;
1364 my $opts = shift || {};
1366 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1367 print STDERR (join (" ",
1368 map { / / ? "'$_'" : $_ }
1371 if $ENV{CRUNCH_DEBUG};
1373 if (defined $stdin) {
1374 my $child = open STDIN, "-|";
1375 defined $child or die "no fork: $!";
1377 print $stdin or die $!;
1378 close STDOUT or die $!;
1383 return system (@$args) if $opts->{fork};
1386 warn "ENV size is ".length(join(" ",%ENV));
1387 die "exec failed: $!: @$args";
1391 sub ban_node_by_slot {
1392 # Don't start any new jobsteps on this node for 60 seconds
1394 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1395 $slot[$slotid]->{node}->{hold_count}++;
1396 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1401 my ($lockfile, $error_message) = @_;
1402 open L, ">", $lockfile or croak("$lockfile: $!");
1403 if (!flock L, LOCK_EX|LOCK_NB) {
1404 croak("Can't lock $lockfile: $error_message\n");
1411 # checkout-and-build
1415 my $destdir = $ENV{"CRUNCH_SRC"};
1416 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1417 my $repo = $ENV{"CRUNCH_SRC_URL"};
1419 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1421 if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
1425 unlink "$destdir.commit";
1426 open STDOUT, ">", "$destdir.log";
1427 open STDERR, ">&STDOUT";
1430 my @git_archive_data = <DATA>;
1431 if (@git_archive_data) {
1432 open TARX, "|-", "tar", "-C", $destdir, "-xf", "-";
1433 print TARX @git_archive_data;
1435 die "'tar -C $destdir -xf -' exited $?: $!";
1440 chomp ($pwd = `pwd`);
1441 my $install_dir = $ENV{"CRUNCH_INSTALL"} || "$pwd/opt";
1444 for my $src_path ("$destdir/arvados/sdk/python") {
1446 shell_or_die ("virtualenv", $install_dir);
1447 shell_or_die ("cd $src_path && ./build.sh && $install_dir/bin/python setup.py install");
1451 if (-e "$destdir/crunch_scripts/install") {
1452 shell_or_die ("$destdir/crunch_scripts/install", $install_dir);
1453 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
1455 shell_or_die ("./tests/autotests.sh", $install_dir);
1456 } elsif (-e "./install.sh") {
1457 shell_or_die ("./install.sh", $install_dir);
1461 unlink "$destdir.commit.new";
1462 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
1463 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
1472 if ($ENV{"DEBUG"}) {
1473 print STDERR "@_\n";
1476 or die "@_ failed: $! exit 0x".sprintf("%x",$?);