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 --git-dir /path/to/repo/.git
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/working/tree","script":"scriptname",...}'
20 crunch-job --job '{"repository":"https://github.com/curoverse/arvados.git","script_version":"master","script":"scriptname",...}'
28 If the job is already locked, steal the lock and run it anyway.
32 Path to a .git directory (or a git URL) where the commit given in the
33 job's C<script_version> attribute is to be found. If this is I<not>
34 given, the job's C<repository> attribute will be used.
38 Arvados API authorization token to use during the course of the job.
42 Do not clear per-job/task temporary directories during initial job
43 setup. This can speed up development and debugging when running jobs
48 UUID of the job to run, or a JSON-encoded job resource without a
49 UUID. If the latter is given, a new job object will be created.
53 =head1 RUNNING JOBS LOCALLY
55 crunch-job's log messages appear on stderr along with the job tasks'
56 stderr streams. The log is saved in Keep at each checkpoint and when
59 If the job succeeds, the job's output locator is printed on stdout.
61 While the job is running, the following signals are accepted:
65 =item control-C, SIGINT, SIGQUIT
67 Save a checkpoint, terminate any job tasks that are running, and stop.
71 Save a checkpoint and continue.
75 Refresh node allocation (i.e., check whether any nodes have been added
76 or unallocated) and attributes of the Job record that should affect
77 behavior (e.g., cancel job if cancelled_at becomes non-nil).
85 use POSIX ':sys_wait_h';
86 use POSIX qw(strftime);
87 use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
91 use Digest::MD5 qw(md5_hex);
97 use File::Path qw( make_path remove_tree );
99 use constant TASK_TEMPFAIL => 111;
100 use constant EX_TEMPFAIL => 75;
101 use constant EX_RETRY_UNLOCKED => 93;
103 $ENV{"TMPDIR"} ||= "/tmp";
104 unless (defined $ENV{"CRUNCH_TMP"}) {
105 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
106 if ($ENV{"USER"} ne "crunch" && $< != 0) {
107 # use a tmp dir unique for my uid
108 $ENV{"CRUNCH_TMP"} .= "-$<";
112 # Create the tmp directory if it does not exist
113 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
114 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
117 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
118 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
119 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
120 mkdir ($ENV{"JOB_WORK"});
129 my $docker_bin = "/usr/bin/docker.io";
130 GetOptions('force-unlock' => \$force_unlock,
131 'git-dir=s' => \$git_dir,
132 'job=s' => \$jobspec,
133 'job-api-token=s' => \$job_api_token,
134 'no-clear-tmp' => \$no_clear_tmp,
135 'resume-stash=s' => \$resume_stash,
136 'docker-bin=s' => \$docker_bin,
139 if (defined $job_api_token) {
140 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
143 my $have_slurm = exists $ENV{SLURM_JOB_ID} && exists $ENV{SLURM_NODELIST};
148 $main::ENV{CRUNCH_DEBUG} = 1;
152 $main::ENV{CRUNCH_DEBUG} = 0;
155 my $arv = Arvados->new('apiVersion' => 'v1');
164 if ($jobspec =~ /^[-a-z\d]+$/)
166 # $jobspec is an Arvados UUID, not a JSON job specification
167 $Job = api_call("jobs/get", uuid => $jobspec);
172 $Job = JSON::decode_json($jobspec);
177 # Make sure our workers (our slurm nodes, localhost, or whatever) are
178 # at least able to run basic commands: they aren't down or severely
181 if ($Job->{docker_image_locator}) {
182 $cmd = [$docker_bin, 'ps', '-q'];
184 Log(undef, "Sanity check is `@$cmd`");
185 srun(["srun", "--nodes=\Q$ENV{SLURM_NNODES}\E", "--ntasks-per-node=1"],
189 Log(undef, "Sanity check failed: ".exit_status_s($?));
192 Log(undef, "Sanity check OK");
195 my $User = api_call("users/current");
198 if (!$force_unlock) {
199 # Claim this job, and make sure nobody else does
200 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
202 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
211 map { croak ("No $_ specified") unless $Job->{$_} }
212 qw(script script_version script_parameters);
215 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
216 $Job->{'started_at'} = gmtime;
217 $Job->{'state'} = 'Running';
219 $Job = api_call("jobs/create", job => $Job);
221 $job_id = $Job->{'uuid'};
223 my $keep_logfile = $job_id . '.log.txt';
224 log_writer_start($keep_logfile);
226 $Job->{'runtime_constraints'} ||= {};
227 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
228 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
230 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
232 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
233 chomp($gem_versions);
234 chop($gem_versions); # Closing parentheses
239 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
241 Log (undef, "check slurm allocation");
244 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
248 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
249 push @sinfo, "$localcpus localhost";
251 if (exists $ENV{SLURM_NODELIST})
253 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
257 my ($ncpus, $slurm_nodelist) = split;
258 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
261 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
264 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
267 foreach (split (",", $ranges))
280 push @nodelist, map {
282 $n =~ s/\[[-,\d]+\]/$_/;
289 push @nodelist, $nodelist;
292 foreach my $nodename (@nodelist)
294 Log (undef, "node $nodename - $ncpus slots");
295 my $node = { name => $nodename,
299 foreach my $cpu (1..$ncpus)
301 push @slot, { node => $node,
305 push @node, @nodelist;
310 # Ensure that we get one jobstep running on each allocated node before
311 # we start overloading nodes with concurrent steps
313 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
316 $Job->update_attributes(
317 'tasks_summary' => { 'failed' => 0,
322 Log (undef, "start");
323 $SIG{'INT'} = sub { $main::please_freeze = 1; };
324 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
325 $SIG{'TERM'} = \&croak;
326 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
327 $SIG{'ALRM'} = sub { $main::please_info = 1; };
328 $SIG{'CONT'} = sub { $main::please_continue = 1; };
329 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
331 $main::please_freeze = 0;
332 $main::please_info = 0;
333 $main::please_continue = 0;
334 $main::please_refresh = 0;
335 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
337 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
338 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
339 $ENV{"JOB_UUID"} = $job_id;
342 my @jobstep_todo = ();
343 my @jobstep_done = ();
344 my @jobstep_tomerge = ();
345 my $jobstep_tomerge_level = 0;
346 my $squeue_checked = 0;
347 my $latest_refresh = scalar time;
351 if (defined $Job->{thawedfromkey})
353 thaw ($Job->{thawedfromkey});
357 my $first_task = api_call("job_tasks/create", job_task => {
358 'job_uuid' => $Job->{'uuid'},
363 push @jobstep, { 'level' => 0,
365 'arvados_task' => $first_task,
367 push @jobstep_todo, 0;
373 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
376 my $build_script = handle_readall(\*DATA);
377 my $nodelist = join(",", @node);
378 my $git_tar_count = 0;
380 if (!defined $no_clear_tmp) {
381 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
382 Log (undef, "Clean work dirs");
384 my $cleanpid = fork();
387 # Find FUSE mounts that look like Keep mounts (the mount path has the
388 # word "keep") and unmount them. Then clean up work directories.
389 # TODO: When #5036 is done and widely deployed, we can get rid of the
390 # regular expression and just unmount everything with type fuse.keep.
391 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
392 ['bash', '-ec', 'mount -t fuse,fuse.keep | awk \'($3 ~ /\ykeep\y/){print $3}\' | xargs -r -n 1 fusermount -u -z; sleep 1; rm -rf $JOB_WORK $CRUNCH_INSTALL $CRUNCH_TMP/task $CRUNCH_TMP/src* $CRUNCH_TMP/*.cid']);
397 last if $cleanpid == waitpid (-1, WNOHANG);
398 freeze_if_want_freeze ($cleanpid);
399 select (undef, undef, undef, 0.1);
401 Log (undef, "Cleanup command exited ".exit_status_s($?));
404 # If this job requires a Docker image, install that.
405 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem);
406 if ($docker_locator = $Job->{docker_image_locator}) {
407 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
410 croak("No Docker image hash found from locator $docker_locator");
412 $docker_stream =~ s/^\.//;
413 my $docker_install_script = qq{
414 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
415 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
418 my $docker_pid = fork();
419 if ($docker_pid == 0)
421 srun (["srun", "--nodelist=" . join(',', @node)],
422 ["/bin/sh", "-ec", $docker_install_script]);
427 last if $docker_pid == waitpid (-1, WNOHANG);
428 freeze_if_want_freeze ($docker_pid);
429 select (undef, undef, undef, 0.1);
433 croak("Installing Docker image from $docker_locator exited "
437 # Determine whether this version of Docker supports memory+swap limits.
438 srun(["srun", "--nodelist=" . $node[0]],
439 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
441 $docker_limitmem = ($? == 0);
443 if ($Job->{arvados_sdk_version}) {
444 # The job also specifies an Arvados SDK version. Add the SDKs to the
445 # tar file for the build script to install.
446 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
447 $Job->{arvados_sdk_version}));
448 add_git_archive("git", "--git-dir=$git_dir", "archive",
449 "--prefix=.arvados.sdk/",
450 $Job->{arvados_sdk_version}, "sdk");
454 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
455 # If script_version looks like an absolute path, *and* the --git-dir
456 # argument was not given -- which implies we were not invoked by
457 # crunch-dispatch -- we will use the given path as a working
458 # directory instead of resolving script_version to a git commit (or
459 # doing anything else with git).
460 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
461 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
464 # Resolve the given script_version to a git commit sha1. Also, if
465 # the repository is remote, clone it into our local filesystem: this
466 # ensures "git archive" will work, and is necessary to reliably
467 # resolve a symbolic script_version like "master^".
468 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
470 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
472 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
474 # If we're running under crunch-dispatch, it will have already
475 # pulled the appropriate source tree into its own repository, and
476 # given us that repo's path as $git_dir.
478 # If we're running a "local" job, we might have to fetch content
479 # from a remote repository.
481 # (Currently crunch-dispatch gives a local path with --git-dir, but
482 # we might as well accept URLs there too in case it changes its
484 my $repo = $git_dir || $Job->{'repository'};
486 # Repository can be remote or local. If remote, we'll need to fetch it
487 # to a local dir before doing `git log` et al.
490 if ($repo =~ m{://|^[^/]*:}) {
491 # $repo is a git url we can clone, like git:// or https:// or
492 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
493 # not recognized here because distinguishing that from a local
494 # path is too fragile. If you really need something strange here,
495 # use the ssh:// form.
496 $repo_location = 'remote';
497 } elsif ($repo =~ m{^\.*/}) {
498 # $repo is a local path to a git index. We'll also resolve ../foo
499 # to ../foo/.git if the latter is a directory. To help
500 # disambiguate local paths from named hosted repositories, this
501 # form must be given as ./ or ../ if it's a relative path.
502 if (-d "$repo/.git") {
503 $repo = "$repo/.git";
505 $repo_location = 'local';
507 # $repo is none of the above. It must be the name of a hosted
509 my $arv_repo_list = api_call("repositories/list",
510 'filters' => [['name','=',$repo]]);
511 my @repos_found = @{$arv_repo_list->{'items'}};
512 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
514 Log(undef, "Repository '$repo' -> "
515 . join(", ", map { $_->{'uuid'} } @repos_found));
518 croak("Error: Found $n_found repositories with name '$repo'.");
520 $repo = $repos_found[0]->{'fetch_url'};
521 $repo_location = 'remote';
523 Log(undef, "Using $repo_location repository '$repo'");
524 $ENV{"CRUNCH_SRC_URL"} = $repo;
526 # Resolve given script_version (we'll call that $treeish here) to a
527 # commit sha1 ($commit).
528 my $treeish = $Job->{'script_version'};
530 if ($repo_location eq 'remote') {
531 # We minimize excess object-fetching by re-using the same bare
532 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
533 # just keep adding remotes to it as needed.
534 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
535 my $gitcmd = "git --git-dir=\Q$local_repo\E";
537 # Set up our local repo for caching remote objects, making
539 if (!-d $local_repo) {
540 make_path($local_repo) or croak("Error: could not create $local_repo");
542 # This works (exits 0 and doesn't delete fetched objects) even
543 # if $local_repo is already initialized:
544 `$gitcmd init --bare`;
546 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
549 # If $treeish looks like a hash (or abbrev hash) we look it up in
550 # our local cache first, since that's cheaper. (We don't want to
551 # do that with tags/branches though -- those change over time, so
552 # they should always be resolved by the remote repo.)
553 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
554 # Hide stderr because it's normal for this to fail:
555 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
557 # Careful not to resolve a branch named abcdeff to commit 1234567:
558 $sha1 =~ /^$treeish/ &&
559 $sha1 =~ /^([0-9a-f]{40})$/s) {
561 Log(undef, "Commit $commit already present in $local_repo");
565 if (!defined $commit) {
566 # If $treeish isn't just a hash or abbrev hash, or isn't here
567 # yet, we need to fetch the remote to resolve it correctly.
569 # First, remove all local heads. This prevents a name that does
570 # not exist on the remote from resolving to (or colliding with)
571 # a previously fetched branch or tag (possibly from a different
573 remove_tree("$local_repo/refs/heads", {keep_root => 1});
575 Log(undef, "Fetching objects from $repo to $local_repo");
576 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
578 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
582 # Now that the data is all here, we will use our local repo for
583 # the rest of our git activities.
587 my $gitcmd = "git --git-dir=\Q$repo\E";
588 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
589 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
590 croak("`$gitcmd rev-list` exited "
592 .", '$treeish' not found. Giving up.");
595 Log(undef, "Version $treeish is commit $commit");
597 if ($commit ne $Job->{'script_version'}) {
598 # Record the real commit id in the database, frozentokey, logs,
599 # etc. -- instead of an abbreviation or a branch name which can
600 # become ambiguous or point to a different commit in the future.
601 if (!$Job->update_attributes('script_version' => $commit)) {
602 croak("Error: failed to update job's script_version attribute");
606 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
607 add_git_archive("$gitcmd archive ''\Q$commit\E");
610 my $git_archive = combined_git_archive();
611 if (!defined $git_archive) {
612 Log(undef, "Skip install phase (no git archive)");
614 Log(undef, "Warning: This probably means workers have no source tree!");
619 my $install_script_tries_left = 3;
620 for (my $attempts = 0; $attempts < 3; $attempts++) {
621 Log(undef, "Run install script on all workers");
623 my @srunargs = ("srun",
624 "--nodelist=$nodelist",
625 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
626 my @execargs = ("sh", "-c",
627 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
629 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
630 my ($install_stderr_r, $install_stderr_w);
631 pipe $install_stderr_r, $install_stderr_w or croak("pipe() failed: $!");
632 set_nonblocking($install_stderr_r);
633 my $installpid = fork();
634 if ($installpid == 0)
636 close($install_stderr_r);
637 fcntl($install_stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
638 open(STDOUT, ">&", $install_stderr_w);
639 open(STDERR, ">&", $install_stderr_w);
640 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
643 close($install_stderr_w);
644 # Tell freeze_if_want_freeze how to kill the child, otherwise the
645 # "waitpid(installpid)" loop won't get interrupted by a freeze:
646 $proc{$installpid} = {};
648 # Track whether anything appears on stderr other than slurm errors
649 # ("srun: ...") and the "starting: ..." message printed by the
650 # srun subroutine itself:
651 my $stderr_anything_from_script = 0;
652 my $match_our_own_errors = '^(srun: error: |starting: \[)';
653 while ($installpid != waitpid(-1, WNOHANG)) {
654 freeze_if_want_freeze ($installpid);
655 # Wait up to 0.1 seconds for something to appear on stderr, then
656 # do a non-blocking read.
657 my $bits = fhbits($install_stderr_r);
658 select ($bits, undef, $bits, 0.1);
659 if (0 < sysread ($install_stderr_r, $stderr_buf, 8192, length($stderr_buf)))
661 while ($stderr_buf =~ /^(.*?)\n/) {
663 substr $stderr_buf, 0, 1+length($line), "";
664 Log(undef, "stderr $line");
665 if ($line !~ /$match_our_own_errors/) {
666 $stderr_anything_from_script = 1;
671 delete $proc{$installpid};
672 $install_exited = $?;
673 close($install_stderr_r);
674 if (length($stderr_buf) > 0) {
675 if ($stderr_buf !~ /$match_our_own_errors/) {
676 $stderr_anything_from_script = 1;
678 Log(undef, "stderr $stderr_buf")
681 Log (undef, "Install script exited ".exit_status_s($install_exited));
682 last if $install_exited == 0 || $main::please_freeze;
683 # If the install script fails but doesn't print an error message,
684 # the next thing anyone is likely to do is just run it again in
685 # case it was a transient problem like "slurm communication fails
686 # because the network isn't reliable enough". So we'll just do
687 # that ourselves (up to 3 attempts in total). OTOH, if there is an
688 # error message, the problem is more likely to have a real fix and
689 # we should fail the job so the fixing process can start, instead
690 # of doing 2 more attempts.
691 last if $stderr_anything_from_script;
694 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
695 unlink($tar_filename);
698 if ($install_exited != 0) {
703 foreach (qw (script script_version script_parameters runtime_constraints))
707 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
709 foreach (split (/\n/, $Job->{knobs}))
711 Log (undef, "knob " . $_);
716 $main::success = undef;
722 my $thisround_succeeded = 0;
723 my $thisround_failed = 0;
724 my $thisround_failed_multiple = 0;
725 my $working_slot_count = scalar(@slot);
727 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
728 or $a <=> $b } @jobstep_todo;
729 my $level = $jobstep[$jobstep_todo[0]]->{level};
731 my $initial_tasks_this_level = 0;
732 foreach my $id (@jobstep_todo) {
733 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
736 # If the number of tasks scheduled at this level #T is smaller than the number
737 # of slots available #S, only use the first #T slots, or the first slot on
738 # each node, whichever number is greater.
740 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
741 # based on these numbers. Using fewer slots makes more resources available
742 # to each individual task, which should normally be a better strategy when
743 # there are fewer of them running with less parallelism.
745 # Note that this calculation is not redone if the initial tasks at
746 # this level queue more tasks at the same level. This may harm
747 # overall task throughput for that level.
749 if ($initial_tasks_this_level < @node) {
750 @freeslot = (0..$#node);
751 } elsif ($initial_tasks_this_level < @slot) {
752 @freeslot = (0..$initial_tasks_this_level - 1);
754 @freeslot = (0..$#slot);
756 my $round_num_freeslots = scalar(@freeslot);
758 my %round_max_slots = ();
759 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
760 my $this_slot = $slot[$freeslot[$ii]];
761 my $node_name = $this_slot->{node}->{name};
762 $round_max_slots{$node_name} ||= $this_slot->{cpu};
763 last if (scalar(keys(%round_max_slots)) >= @node);
766 Log(undef, "start level $level with $round_num_freeslots slots");
769 my $progress_is_dirty = 1;
770 my $progress_stats_updated = 0;
772 update_progress_stats();
776 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
778 my $id = $jobstep_todo[$todo_ptr];
779 my $Jobstep = $jobstep[$id];
780 if ($Jobstep->{level} != $level)
785 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
786 set_nonblocking($reader{$id});
788 my $childslot = $freeslot[0];
789 my $childnode = $slot[$childslot]->{node};
790 my $childslotname = join (".",
791 $slot[$childslot]->{node}->{name},
792 $slot[$childslot]->{cpu});
794 my $childpid = fork();
797 $SIG{'INT'} = 'DEFAULT';
798 $SIG{'QUIT'} = 'DEFAULT';
799 $SIG{'TERM'} = 'DEFAULT';
801 foreach (values (%reader))
805 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
806 open(STDOUT,">&writer");
807 open(STDERR,">&writer");
812 delete $ENV{"GNUPGHOME"};
813 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
814 $ENV{"TASK_QSEQUENCE"} = $id;
815 $ENV{"TASK_SEQUENCE"} = $level;
816 $ENV{"JOB_SCRIPT"} = $Job->{script};
817 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
818 $param =~ tr/a-z/A-Z/;
819 $ENV{"JOB_PARAMETER_$param"} = $value;
821 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
822 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
823 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
824 $ENV{"HOME"} = $ENV{"TASK_WORK"};
825 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
826 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
827 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
828 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
834 "--nodelist=".$childnode->{name},
835 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
836 "--job-name=$job_id.$id.$$",
839 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
840 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
841 ."&& cd $ENV{CRUNCH_TMP} "
842 # These environment variables get used explicitly later in
843 # $command. No tool is expected to read these values directly.
844 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
845 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
846 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
847 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
848 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
851 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
852 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
853 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
854 # We only set memory limits if Docker lets us limit both memory and swap.
855 # Memory limits alone have been supported longer, but subprocesses tend
856 # to get SIGKILL if they exceed that without any swap limit set.
857 # See #5642 for additional background.
858 if ($docker_limitmem) {
859 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
862 # Dynamically configure the container to use the host system as its
863 # DNS server. Get the host's global addresses from the ip command,
864 # and turn them into docker --dns options using gawk.
866 q{$(ip -o address show scope global |
867 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
869 # The source tree and $destdir directory (which we have
870 # installed on the worker host) are available in the container,
871 # under the same path.
872 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
873 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
875 # Currently, we make arv-mount's mount point appear at /keep
876 # inside the container (instead of using the same path as the
877 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
878 # crunch scripts and utilities must not rely on this. They must
879 # use $TASK_KEEPMOUNT.
880 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
881 $ENV{TASK_KEEPMOUNT} = "/keep";
883 # TASK_WORK is almost exactly like a docker data volume: it
884 # starts out empty, is writable, and persists until no
885 # containers use it any more. We don't use --volumes-from to
886 # share it with other containers: it is only accessible to this
887 # task, and it goes away when this task stops.
889 # However, a docker data volume is writable only by root unless
890 # the mount point already happens to exist in the container with
891 # different permissions. Therefore, we [1] assume /tmp already
892 # exists in the image and is writable by the crunch user; [2]
893 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
894 # writable if they are created by docker while setting up the
895 # other --volumes); and [3] create $TASK_WORK inside the
896 # container using $build_script.
897 $command .= "--volume=/tmp ";
898 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
899 $ENV{"HOME"} = $ENV{"TASK_WORK"};
900 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
902 # TODO: Share a single JOB_WORK volume across all task
903 # containers on a given worker node, and delete it when the job
904 # ends (and, in case that doesn't work, when the next job
907 # For now, use the same approach as TASK_WORK above.
908 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
910 while (my ($env_key, $env_val) = each %ENV)
912 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
913 $command .= "--env=\Q$env_key=$env_val\E ";
916 $command .= "--env=\QHOME=$ENV{HOME}\E ";
917 $command .= "\Q$docker_hash\E ";
918 $command .= "stdbuf --output=0 --error=0 ";
919 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
922 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
923 $command .= "stdbuf --output=0 --error=0 ";
924 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
927 my @execargs = ('bash', '-c', $command);
928 srun (\@srunargs, \@execargs, undef, $build_script);
929 # exec() failed, we assume nothing happened.
930 die "srun() failed on build script\n";
933 if (!defined $childpid)
940 $proc{$childpid} = { jobstep => $id,
943 jobstepname => "$job_id.$id.$childpid",
945 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
946 $slot[$childslot]->{pid} = $childpid;
948 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
949 Log ($id, "child $childpid started on $childslotname");
950 $Jobstep->{starttime} = time;
951 $Jobstep->{node} = $childnode->{name};
952 $Jobstep->{slotindex} = $childslot;
953 delete $Jobstep->{stderr};
954 delete $Jobstep->{finishtime};
956 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
957 $Jobstep->{'arvados_task'}->save;
959 splice @jobstep_todo, $todo_ptr, 1;
962 $progress_is_dirty = 1;
966 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
968 last THISROUND if $main::please_freeze || defined($main::success);
969 if ($main::please_info)
971 $main::please_info = 0;
973 create_output_collection();
975 update_progress_stats();
982 check_refresh_wanted();
984 update_progress_stats();
985 select (undef, undef, undef, 0.1);
987 elsif (time - $progress_stats_updated >= 30)
989 update_progress_stats();
991 $working_slot_count = scalar(grep { $_->{node}->{losing_streak} == 0 &&
992 $_->{node}->{hold_count} < 4 } @slot);
993 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
994 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
996 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
997 .($thisround_failed+$thisround_succeeded)
998 .") -- giving up on this round";
999 Log (undef, $message);
1003 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1004 for (my $i=$#freeslot; $i>=0; $i--) {
1005 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1006 push @holdslot, (splice @freeslot, $i, 1);
1009 for (my $i=$#holdslot; $i>=0; $i--) {
1010 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1011 push @freeslot, (splice @holdslot, $i, 1);
1015 # give up if no nodes are succeeding
1016 if ($working_slot_count < 1) {
1017 Log(undef, "Every node has failed -- giving up");
1024 push @freeslot, splice @holdslot;
1025 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1028 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1031 if ($main::please_continue) {
1032 $main::please_continue = 0;
1035 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1037 if (!reapchildren())
1039 check_refresh_wanted();
1041 update_progress_stats();
1042 select (undef, undef, undef, 0.1);
1043 killem (keys %proc) if $main::please_freeze;
1047 update_progress_stats();
1048 freeze_if_want_freeze();
1051 if (!defined $main::success)
1053 if (!@jobstep_todo) {
1055 } elsif ($working_slot_count < 1) {
1056 save_output_collection();
1058 exit(EX_RETRY_UNLOCKED);
1059 } elsif ($thisround_succeeded == 0 &&
1060 ($thisround_failed == 0 || $thisround_failed > 4)) {
1061 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1062 Log (undef, $message);
1067 goto ONELEVEL if !defined $main::success;
1070 release_allocation();
1072 my $collated_output = save_output_collection();
1073 Log (undef, "finish");
1078 if ($collated_output && $main::success) {
1079 $final_state = 'Complete';
1081 $final_state = 'Failed';
1083 $Job->update_attributes('state' => $final_state);
1085 exit (($final_state eq 'Complete') ? 0 : 1);
1089 sub update_progress_stats
1091 $progress_stats_updated = time;
1092 return if !$progress_is_dirty;
1093 my ($todo, $done, $running) = (scalar @jobstep_todo,
1094 scalar @jobstep_done,
1095 scalar @slot - scalar @freeslot - scalar @holdslot);
1096 $Job->{'tasks_summary'} ||= {};
1097 $Job->{'tasks_summary'}->{'todo'} = $todo;
1098 $Job->{'tasks_summary'}->{'done'} = $done;
1099 $Job->{'tasks_summary'}->{'running'} = $running;
1100 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1101 Log (undef, "status: $done done, $running running, $todo todo");
1102 $progress_is_dirty = 0;
1109 my $pid = waitpid (-1, WNOHANG);
1110 return 0 if $pid <= 0;
1112 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1114 . $slot[$proc{$pid}->{slot}]->{cpu});
1115 my $jobstepid = $proc{$pid}->{jobstep};
1116 my $elapsed = time - $proc{$pid}->{time};
1117 my $Jobstep = $jobstep[$jobstepid];
1119 my $childstatus = $?;
1120 my $exitvalue = $childstatus >> 8;
1121 my $exitinfo = "exit ".exit_status_s($childstatus);
1122 $Jobstep->{'arvados_task'}->reload;
1123 my $task_success = $Jobstep->{'arvados_task'}->{success};
1125 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1127 if (!defined $task_success) {
1128 # task did not indicate one way or the other --> fail
1129 $Jobstep->{'arvados_task'}->{success} = 0;
1130 $Jobstep->{'arvados_task'}->save;
1137 $temporary_fail ||= $Jobstep->{node_fail};
1138 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1140 ++$thisround_failed;
1141 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1143 # Check for signs of a failed or misconfigured node
1144 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1145 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1146 # Don't count this against jobstep failure thresholds if this
1147 # node is already suspected faulty and srun exited quickly
1148 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1150 Log ($jobstepid, "blaming failure on suspect node " .
1151 $slot[$proc{$pid}->{slot}]->{node}->{name});
1152 $temporary_fail ||= 1;
1154 ban_node_by_slot($proc{$pid}->{slot});
1157 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1158 ++$Jobstep->{'failures'},
1159 $temporary_fail ? 'temporary' : 'permanent',
1162 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1163 # Give up on this task, and the whole job
1166 # Put this task back on the todo queue
1167 push @jobstep_todo, $jobstepid;
1168 $Job->{'tasks_summary'}->{'failed'}++;
1172 ++$thisround_succeeded;
1173 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1174 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1175 push @jobstep_done, $jobstepid;
1176 Log ($jobstepid, "success in $elapsed seconds");
1178 $Jobstep->{exitcode} = $childstatus;
1179 $Jobstep->{finishtime} = time;
1180 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1181 $Jobstep->{'arvados_task'}->save;
1182 process_stderr ($jobstepid, $task_success);
1183 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1184 length($Jobstep->{'arvados_task'}->{output}),
1185 $Jobstep->{'arvados_task'}->{output}));
1187 close $reader{$jobstepid};
1188 delete $reader{$jobstepid};
1189 delete $slot[$proc{$pid}->{slot}]->{pid};
1190 push @freeslot, $proc{$pid}->{slot};
1193 if ($task_success) {
1195 my $newtask_list = [];
1196 my $newtask_results;
1198 $newtask_results = api_call(
1201 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1203 'order' => 'qsequence',
1204 'offset' => scalar(@$newtask_list),
1206 push(@$newtask_list, @{$newtask_results->{items}});
1207 } while (@{$newtask_results->{items}});
1208 foreach my $arvados_task (@$newtask_list) {
1210 'level' => $arvados_task->{'sequence'},
1212 'arvados_task' => $arvados_task
1214 push @jobstep, $jobstep;
1215 push @jobstep_todo, $#jobstep;
1219 $progress_is_dirty = 1;
1223 sub check_refresh_wanted
1225 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1226 if (@stat && $stat[9] > $latest_refresh) {
1227 $latest_refresh = scalar time;
1228 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1229 for my $attr ('cancelled_at',
1230 'cancelled_by_user_uuid',
1231 'cancelled_by_client_uuid',
1233 $Job->{$attr} = $Job2->{$attr};
1235 if ($Job->{'state'} ne "Running") {
1236 if ($Job->{'state'} eq "Cancelled") {
1237 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1239 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1242 $main::please_freeze = 1;
1249 my $last_squeue_check = $squeue_checked;
1251 # Do not call `squeue` or check the kill list more than once every
1253 return if $last_squeue_check > time - 15;
1254 $squeue_checked = time;
1256 # Look for children from which we haven't received stderr data since
1257 # the last squeue check. If no such children exist, all procs are
1258 # alive and there's no need to even look at squeue.
1260 # As long as the crunchstat poll interval (10s) is shorter than the
1261 # squeue check interval (15s) this should make the squeue check an
1263 my $silent_procs = 0;
1264 for my $jobstep (values %proc)
1266 if ($jobstep->{stderr_at} < $last_squeue_check)
1271 return if $silent_procs == 0;
1273 # use killem() on procs whose killtime is reached
1274 while (my ($pid, $jobstep) = each %proc)
1276 if (exists $jobstep->{killtime}
1277 && $jobstep->{killtime} <= time
1278 && $jobstep->{stderr_at} < $last_squeue_check)
1281 if ($jobstep->{stderr_at}) {
1282 $sincewhen = " in last " . (time - $jobstep->{stderr_at}) . "s";
1284 Log($jobstep->{jobstep}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1291 # here is an opportunity to check for mysterious problems with local procs
1295 # Get a list of steps still running. Note: squeue(1) says --steps
1296 # selects a format (which we override anyway) and allows us to
1297 # specify which steps we're interested in (which we don't).
1298 # Importantly, it also changes the meaning of %j from "job name" to
1299 # "step name" and (although this isn't mentioned explicitly in the
1300 # docs) switches from "one line per job" mode to "one line per step"
1301 # mode. Without it, we'd just get a list of one job, instead of a
1303 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1306 Log(undef, "warning: squeue exit status $? ($!)");
1311 # which of my jobsteps are running, according to squeue?
1313 for my $jobstepname (@squeue)
1315 $ok{$jobstepname} = 1;
1318 # Check for child procs >60s old and not mentioned by squeue.
1319 while (my ($pid, $jobstep) = each %proc)
1321 if ($jobstep->{time} < time - 60
1322 && $jobstep->{jobstepname}
1323 && !exists $ok{$jobstep->{jobstepname}}
1324 && !exists $jobstep->{killtime})
1326 # According to slurm, this task has ended (successfully or not)
1327 # -- but our srun child hasn't exited. First we must wait (30
1328 # seconds) in case this is just a race between communication
1329 # channels. Then, if our srun child process still hasn't
1330 # terminated, we'll conclude some slurm communication
1331 # error/delay has caused the task to die without notifying srun,
1332 # and we'll kill srun ourselves.
1333 $jobstep->{killtime} = time + 30;
1334 Log($jobstep->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
1340 sub release_allocation
1344 Log (undef, "release job allocation");
1345 system "scancel $ENV{SLURM_JOB_ID}";
1353 foreach my $job (keys %reader)
1356 while (0 < sysread ($reader{$job}, $buf, 8192))
1358 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1359 $jobstep[$job]->{stderr_at} = time;
1360 $jobstep[$job]->{stderr} .= $buf;
1361 preprocess_stderr ($job);
1362 if (length ($jobstep[$job]->{stderr}) > 16384)
1364 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1373 sub preprocess_stderr
1377 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1379 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1380 Log ($job, "stderr $line");
1381 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1383 $main::please_freeze = 1;
1385 elsif ($line =~ /(srun: error: (Node failure on|Unable to create job step|.*: Communication connection failure))|arvados.errors.Keep/) {
1386 $jobstep[$job]->{node_fail} = 1;
1387 ban_node_by_slot($jobstep[$job]->{slotindex});
1396 my $task_success = shift;
1397 preprocess_stderr ($job);
1400 Log ($job, "stderr $_");
1401 } split ("\n", $jobstep[$job]->{stderr});
1408 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1409 Log(undef, "fetch_block run error from arv-get $hash: $!");
1412 my $output_block = "";
1415 my $bytes = sysread($keep, $buf, 1024 * 1024);
1416 if (!defined $bytes) {
1417 Log(undef, "fetch_block read error from arv-get: $!");
1418 $output_block = undef;
1420 } elsif ($bytes == 0) {
1421 # sysread returns 0 at the end of the pipe.
1424 # some bytes were read into buf.
1425 $output_block .= $buf;
1430 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1431 $output_block = undef;
1433 return $output_block;
1436 # Create a collection by concatenating the output of all tasks (each
1437 # task's output is either a manifest fragment, a locator for a
1438 # manifest fragment stored in Keep, or nothing at all). Return the
1439 # portable_data_hash of the new collection.
1440 sub create_output_collection
1442 Log (undef, "collate");
1444 my ($child_out, $child_in);
1445 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1448 print (arvados.api("v1").collections().
1449 create(body={"manifest_text": sys.stdin.read()}).
1450 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1454 my $manifest_size = 0;
1458 my $output = $_->{'arvados_task'}->{output};
1459 next if (!defined($output));
1461 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1462 $next_write = fetch_block($output);
1464 $next_write = $output;
1466 if (defined($next_write)) {
1467 if (!defined(syswrite($child_in, $next_write))) {
1468 # There's been an error writing. Stop the loop.
1469 # We'll log details about the exit code later.
1472 $manifest_size += length($next_write);
1475 my $uuid = $_->{'arvados_task'}->{'uuid'};
1476 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1481 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1484 my $s = IO::Select->new($child_out);
1485 if ($s->can_read(120)) {
1486 sysread($child_out, $joboutput, 1024 * 1024);
1489 Log(undef, "output collection creation exited " . exit_status_s($?));
1495 Log (undef, "timed out while creating output collection");
1496 foreach my $signal (2, 2, 2, 15, 15, 9) {
1497 kill($signal, $pid);
1498 last if waitpid($pid, WNOHANG) == -1;
1507 # Calls create_output_collection, logs the result, and returns it.
1508 # If that was successful, save that as the output in the job record.
1509 sub save_output_collection {
1510 my $collated_output = create_output_collection();
1512 if (!$collated_output) {
1513 Log(undef, "Failed to write output collection");
1516 Log(undef, "job output $collated_output");
1517 $Job->update_attributes('output' => $collated_output);
1519 return $collated_output;
1526 my $sig = 2; # SIGINT first
1527 if (exists $proc{$_}->{"sent_$sig"} &&
1528 time - $proc{$_}->{"sent_$sig"} > 4)
1530 $sig = 15; # SIGTERM if SIGINT doesn't work
1532 if (exists $proc{$_}->{"sent_$sig"} &&
1533 time - $proc{$_}->{"sent_$sig"} > 4)
1535 $sig = 9; # SIGKILL if SIGTERM doesn't work
1537 if (!exists $proc{$_}->{"sent_$sig"})
1539 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1541 select (undef, undef, undef, 0.1);
1544 kill $sig, $_; # srun wants two SIGINT to really interrupt
1546 $proc{$_}->{"sent_$sig"} = time;
1547 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1557 vec($bits,fileno($_),1) = 1;
1563 # Send log output to Keep via arv-put.
1565 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1566 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1567 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1568 # $log_pipe_pid is the pid of the arv-put subprocess.
1570 # The only functions that should access these variables directly are:
1572 # log_writer_start($logfilename)
1573 # Starts an arv-put pipe, reading data on stdin and writing it to
1574 # a $logfilename file in an output collection.
1576 # log_writer_read_output([$timeout])
1577 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1578 # Passes $timeout to the select() call, with a default of 0.01.
1579 # Returns the result of the last read() call on $log_pipe_out, or
1580 # -1 if read() wasn't called because select() timed out.
1581 # Only other log_writer_* functions should need to call this.
1583 # log_writer_send($txt)
1584 # Writes $txt to the output log collection.
1586 # log_writer_finish()
1587 # Closes the arv-put pipe and returns the output that it produces.
1589 # log_writer_is_active()
1590 # Returns a true value if there is currently a live arv-put
1591 # process, false otherwise.
1593 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1596 sub log_writer_start($)
1598 my $logfilename = shift;
1599 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1603 '--filename', $logfilename,
1605 $log_pipe_out_buf = "";
1606 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1609 sub log_writer_read_output {
1610 my $timeout = shift || 0.01;
1612 while ($read && $log_pipe_out_select->can_read($timeout)) {
1613 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1614 length($log_pipe_out_buf));
1616 if (!defined($read)) {
1617 Log(undef, "error reading log manifest from arv-put: $!");
1622 sub log_writer_send($)
1625 print $log_pipe_in $txt;
1626 log_writer_read_output();
1629 sub log_writer_finish()
1631 return unless $log_pipe_pid;
1633 close($log_pipe_in);
1635 my $read_result = log_writer_read_output(120);
1636 if ($read_result == -1) {
1637 Log (undef, "timed out reading from 'arv-put'");
1638 } elsif ($read_result != 0) {
1639 Log(undef, "failed to read arv-put log manifest to EOF");
1642 waitpid($log_pipe_pid, 0);
1644 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1647 close($log_pipe_out);
1648 my $arv_put_output = $log_pipe_out_buf;
1649 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1650 $log_pipe_out_select = undef;
1652 return $arv_put_output;
1655 sub log_writer_is_active() {
1656 return $log_pipe_pid;
1659 sub Log # ($jobstep_id, $logmessage)
1661 if ($_[1] =~ /\n/) {
1662 for my $line (split (/\n/, $_[1])) {
1667 my $fh = select STDERR; $|=1; select $fh;
1668 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1669 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1672 if (log_writer_is_active() || -t STDERR) {
1673 my @gmtime = gmtime;
1674 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1675 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1677 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1679 if (log_writer_is_active()) {
1680 log_writer_send($datetime . " " . $message);
1687 my ($package, $file, $line) = caller;
1688 my $message = "@_ at $file line $line\n";
1689 Log (undef, $message);
1690 freeze() if @jobstep_todo;
1691 create_output_collection() if @jobstep_todo;
1701 if ($Job->{'state'} eq 'Cancelled') {
1702 $Job->update_attributes('finished_at' => scalar gmtime);
1704 $Job->update_attributes('state' => 'Failed');
1711 my $justcheckpoint = shift; # false if this will be the last meta saved
1712 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1713 return unless log_writer_is_active();
1715 my $log_manifest = "";
1717 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1718 $log_manifest .= $prev_log_coll->{manifest_text};
1720 $log_manifest .= log_writer_finish();
1722 my $log_coll = api_call(
1723 "collections/create", ensure_unique_name => 1, collection => {
1724 manifest_text => $log_manifest,
1725 owner_uuid => $Job->{owner_uuid},
1726 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1728 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1729 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1733 sub freeze_if_want_freeze
1735 if ($main::please_freeze)
1737 release_allocation();
1740 # kill some srun procs before freeze+stop
1741 map { $proc{$_} = {} } @_;
1744 killem (keys %proc);
1745 select (undef, undef, undef, 0.1);
1747 while (($died = waitpid (-1, WNOHANG)) > 0)
1749 delete $proc{$died};
1754 create_output_collection();
1764 Log (undef, "Freeze not implemented");
1771 croak ("Thaw not implemented");
1787 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1794 my $srunargs = shift;
1795 my $execargs = shift;
1796 my $opts = shift || {};
1798 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1800 $Data::Dumper::Terse = 1;
1801 $Data::Dumper::Indent = 0;
1802 my $show_cmd = Dumper($args);
1803 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1804 $show_cmd =~ s/\n/ /g;
1805 if ($opts->{fork}) {
1806 Log(undef, "starting: $show_cmd");
1808 # This is a child process: parent is in charge of reading our
1809 # stderr and copying it to Log() if needed.
1810 warn "starting: $show_cmd\n";
1813 if (defined $stdin) {
1814 my $child = open STDIN, "-|";
1815 defined $child or die "no fork: $!";
1817 print $stdin or die $!;
1818 close STDOUT or die $!;
1823 return system (@$args) if $opts->{fork};
1826 warn "ENV size is ".length(join(" ",%ENV));
1827 die "exec failed: $!: @$args";
1831 sub ban_node_by_slot {
1832 # Don't start any new jobsteps on this node for 60 seconds
1834 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1835 $slot[$slotid]->{node}->{hold_count}++;
1836 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1841 my ($lockfile, $error_message) = @_;
1842 open L, ">", $lockfile or croak("$lockfile: $!");
1843 if (!flock L, LOCK_EX|LOCK_NB) {
1844 croak("Can't lock $lockfile: $error_message\n");
1848 sub find_docker_image {
1849 # Given a Keep locator, check to see if it contains a Docker image.
1850 # If so, return its stream name and Docker hash.
1851 # If not, return undef for both values.
1852 my $locator = shift;
1853 my ($streamname, $filename);
1854 my $image = api_call("collections/get", uuid => $locator);
1856 foreach my $line (split(/\n/, $image->{manifest_text})) {
1857 my @tokens = split(/\s+/, $line);
1859 $streamname = shift(@tokens);
1860 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1861 if (defined($filename)) {
1862 return (undef, undef); # More than one file in the Collection.
1864 $filename = (split(/:/, $filedata, 3))[2];
1869 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1870 return ($streamname, $1);
1872 return (undef, undef);
1877 # Calculate the number of times an operation should be retried,
1878 # assuming exponential backoff, and that we're willing to retry as
1879 # long as tasks have been running. Enforce a minimum of 3 retries.
1880 my ($starttime, $endtime, $timediff, $retries);
1882 $starttime = $jobstep[0]->{starttime};
1883 $endtime = $jobstep[-1]->{finishtime};
1885 if (!defined($starttime)) {
1887 } elsif (!defined($endtime)) {
1888 $timediff = time - $starttime;
1890 $timediff = ($endtime - $starttime) - (time - $endtime);
1892 if ($timediff > 0) {
1893 $retries = int(log($timediff) / log(2));
1895 $retries = 1; # Use the minimum.
1897 return ($retries > 3) ? $retries : 3;
1901 # Pass in two function references.
1902 # This method will be called with the remaining arguments.
1903 # If it dies, retry it with exponential backoff until it succeeds,
1904 # or until the current retry_count is exhausted. After each failure
1905 # that can be retried, the second function will be called with
1906 # the current try count (0-based), next try time, and error message.
1907 my $operation = shift;
1908 my $retry_callback = shift;
1909 my $retries = retry_count();
1910 foreach my $try_count (0..$retries) {
1911 my $next_try = time + (2 ** $try_count);
1912 my $result = eval { $operation->(@_); };
1915 } elsif ($try_count < $retries) {
1916 $retry_callback->($try_count, $next_try, $@);
1917 my $sleep_time = $next_try - time;
1918 sleep($sleep_time) if ($sleep_time > 0);
1921 # Ensure the error message ends in a newline, so Perl doesn't add
1922 # retry_op's line number to it.
1928 # Pass in a /-separated API method name, and arguments for it.
1929 # This function will call that method, retrying as needed until
1930 # the current retry_count is exhausted, with a log on the first failure.
1931 my $method_name = shift;
1932 my $log_api_retry = sub {
1933 my ($try_count, $next_try_at, $errmsg) = @_;
1934 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1935 $errmsg =~ s/\s/ /g;
1936 $errmsg =~ s/\s+$//;
1938 if ($next_try_at < time) {
1939 $retry_msg = "Retrying.";
1941 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1942 $retry_msg = "Retrying at $next_try_fmt.";
1944 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1947 foreach my $key (split(/\//, $method_name)) {
1948 $method = $method->{$key};
1950 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1954 # Given a $?, return a human-readable exit code string like "0" or
1955 # "1" or "0 with signal 1" or "1 with signal 11".
1956 my $exitcode = shift;
1957 my $s = $exitcode >> 8;
1958 if ($exitcode & 0x7f) {
1959 $s .= " with signal " . ($exitcode & 0x7f);
1961 if ($exitcode & 0x80) {
1962 $s .= " with core dump";
1967 sub handle_readall {
1968 # Pass in a glob reference to a file handle.
1969 # Read all its contents and return them as a string.
1970 my $fh_glob_ref = shift;
1972 return <$fh_glob_ref>;
1975 sub tar_filename_n {
1977 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1980 sub add_git_archive {
1981 # Pass in a git archive command as a string or list, a la system().
1982 # This method will save its output to be included in the archive sent to the
1986 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1987 croak("Failed to save git archive: $!");
1989 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1991 waitpid($git_pid, 0);
1994 croak("Failed to save git archive: git exited " . exit_status_s($?));
1998 sub combined_git_archive {
1999 # Combine all saved tar archives into a single archive, then return its
2000 # contents in a string. Return undef if no archives have been saved.
2001 if ($git_tar_count < 1) {
2004 my $base_tar_name = tar_filename_n(1);
2005 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2006 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2007 if ($tar_exit != 0) {
2008 croak("Error preparing build archive: tar -A exited " .
2009 exit_status_s($tar_exit));
2012 if (!open(GIT_TAR, "<", $base_tar_name)) {
2013 croak("Could not open build archive: $!");
2015 my $tar_contents = handle_readall(\*GIT_TAR);
2017 return $tar_contents;
2020 sub set_nonblocking {
2022 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2023 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2029 # This is crunch-job's internal dispatch script. crunch-job running on the API
2030 # server invokes this script on individual compute nodes, or localhost if we're
2031 # running a job locally. It gets called in two modes:
2033 # * No arguments: Installation mode. Read a tar archive from the DATA
2034 # file handle; it includes the Crunch script's source code, and
2035 # maybe SDKs as well. Those should be installed in the proper
2036 # locations. This runs outside of any Docker container, so don't try to
2037 # introspect Crunch's runtime environment.
2039 # * With arguments: Crunch script run mode. This script should set up the
2040 # environment, then run the command specified in the arguments. This runs
2041 # inside any Docker container.
2044 use File::Path qw( make_path remove_tree );
2045 use POSIX qw(getcwd);
2047 use constant TASK_TEMPFAIL => 111;
2049 # Map SDK subdirectories to the path environments they belong to.
2050 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2052 my $destdir = $ENV{"CRUNCH_SRC"};
2053 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2054 my $repo = $ENV{"CRUNCH_SRC_URL"};
2055 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2056 my $job_work = $ENV{"JOB_WORK"};
2057 my $task_work = $ENV{"TASK_WORK"};
2059 open(STDOUT_ORIG, ">&", STDOUT);
2060 open(STDERR_ORIG, ">&", STDERR);
2062 for my $dir ($destdir, $job_work, $task_work) {
2065 -e $dir or die "Failed to create temporary directory ($dir): $!";
2070 remove_tree($task_work, {keep_root => 1});
2073 ### Crunch script run mode
2075 # We want to do routine logging during task 0 only. This gives the user
2076 # the information they need, but avoids repeating the information for every
2079 if ($ENV{TASK_SEQUENCE} eq "0") {
2082 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2088 my $python_src = "$install_dir/python";
2089 my $venv_dir = "$job_work/.arvados.venv";
2090 my $venv_built = -e "$venv_dir/bin/activate";
2091 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2092 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2093 "--python=python2.7", $venv_dir);
2094 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2096 $Log->("Built Python SDK virtualenv");
2099 my $pip_bin = "pip";
2101 $Log->("Running in Python SDK virtualenv");
2102 $pip_bin = "$venv_dir/bin/pip";
2103 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2104 @ARGV = ("/bin/sh", "-ec",
2105 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2106 } elsif (-d $python_src) {
2107 $Log->("Warning: virtualenv not found inside Docker container default " .
2108 "\$PATH. Can't install Python SDK.");
2111 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
2113 $Log->("Using Arvados SDK:");
2114 foreach my $line (split /\n/, $pkgs) {
2118 $Log->("Arvados SDK packages not found");
2121 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2122 my $sdk_path = "$install_dir/$sdk_dir";
2124 if ($ENV{$sdk_envkey}) {
2125 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2127 $ENV{$sdk_envkey} = $sdk_path;
2129 $Log->("Arvados SDK added to %s", $sdk_envkey);
2134 die "Cannot exec `@ARGV`: $!";
2137 ### Installation mode
2138 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2140 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2141 # This exact git archive (source + arvados sdk) is already installed
2142 # here, so there's no need to reinstall it.
2144 # We must consume our DATA section, though: otherwise the process
2145 # feeding it to us will get SIGPIPE.
2147 while (read(DATA, $buf, 65536)) { }
2152 unlink "$destdir.archive_hash";
2156 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2157 local $SIG{PIPE} = "IGNORE";
2158 warn "Extracting archive: $archive_hash\n";
2159 # --ignore-zeros is necessary sometimes: depending on how much NUL
2160 # padding tar -A put on our combined archive (which in turn depends
2161 # on the length of the component archives) tar without
2162 # --ignore-zeros will exit before consuming stdin and cause close()
2163 # to fail on the resulting SIGPIPE.
2164 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2165 die "Error launching 'tar -xC $destdir': $!";
2167 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2168 # get SIGPIPE. We must feed it data incrementally.
2170 while (read(DATA, $tar_input, 65536)) {
2171 print TARX $tar_input;
2174 die "'tar -xC $destdir' exited $?: $!";
2180 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2182 foreach my $sdk_lang (("python",
2183 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2184 if (-d "$sdk_root/$sdk_lang") {
2185 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2186 die "Failed to install $sdk_lang SDK: $!";
2192 my $python_dir = "$install_dir/python";
2193 if ((-d $python_dir) and can_run("python2.7")) {
2194 open(my $egg_info_pipe, "-|",
2195 "python2.7 \Q$python_dir/setup.py\E --quiet egg_info 2>&1 >/dev/null");
2196 my @egg_info_errors = <$egg_info_pipe>;
2197 close($egg_info_pipe);
2199 if (@egg_info_errors and ($egg_info_errors[-1] =~ /\bgit\b/)) {
2200 # egg_info apparently failed because it couldn't ask git for a build tag.
2201 # Specify no build tag.
2202 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2203 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2206 my $egg_info_exit = $? >> 8;
2207 foreach my $errline (@egg_info_errors) {
2208 print STDERR_ORIG $errline;
2210 warn "python setup.py egg_info failed: exit $egg_info_exit";
2211 exit ($egg_info_exit || 1);
2216 # Hide messages from the install script (unless it fails: shell_or_die
2217 # will show $destdir.log in that case).
2218 open(STDOUT, ">>", "$destdir.log");
2219 open(STDERR, ">&", STDOUT);
2221 if (-e "$destdir/crunch_scripts/install") {
2222 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2223 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2225 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2226 } elsif (-e "./install.sh") {
2227 shell_or_die (undef, "./install.sh", $install_dir);
2230 if ($archive_hash) {
2231 unlink "$destdir.archive_hash.new";
2232 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2233 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2239 my $command_name = shift;
2240 open(my $which, "-|", "which", $command_name);
2241 while (<$which>) { }
2248 my $exitcode = shift;
2250 if ($ENV{"DEBUG"}) {
2251 print STDERR "@_\n";
2253 if (system (@_) != 0) {
2256 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2257 open STDERR, ">&STDERR_ORIG";
2258 system ("cat $destdir.log >&2");
2259 warn "@_ failed ($err): $exitstatus";
2260 if (defined($exitcode)) {
2264 exit (($code >> 8) || 1);