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;
102 $ENV{"TMPDIR"} ||= "/tmp";
103 unless (defined $ENV{"CRUNCH_TMP"}) {
104 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
105 if ($ENV{"USER"} ne "crunch" && $< != 0) {
106 # use a tmp dir unique for my uid
107 $ENV{"CRUNCH_TMP"} .= "-$<";
111 # Create the tmp directory if it does not exist
112 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
113 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
116 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
117 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
118 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
119 mkdir ($ENV{"JOB_WORK"});
128 my $docker_bin = "/usr/bin/docker.io";
129 GetOptions('force-unlock' => \$force_unlock,
130 'git-dir=s' => \$git_dir,
131 'job=s' => \$jobspec,
132 'job-api-token=s' => \$job_api_token,
133 'no-clear-tmp' => \$no_clear_tmp,
134 'resume-stash=s' => \$resume_stash,
135 'docker-bin=s' => \$docker_bin,
138 if (defined $job_api_token) {
139 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
142 my $have_slurm = exists $ENV{SLURM_JOB_ID} && exists $ENV{SLURM_NODELIST};
147 $main::ENV{CRUNCH_DEBUG} = 1;
151 $main::ENV{CRUNCH_DEBUG} = 0;
154 my $arv = Arvados->new('apiVersion' => 'v1');
163 if ($jobspec =~ /^[-a-z\d]+$/)
165 # $jobspec is an Arvados UUID, not a JSON job specification
166 $Job = api_call("jobs/get", uuid => $jobspec);
171 $Job = JSON::decode_json($jobspec);
176 # Make sure our workers (our slurm nodes, localhost, or whatever) are
177 # at least able to run basic commands: they aren't down or severely
180 if ($Job->{docker_image_locator}) {
181 $cmd = [$docker_bin, 'ps', '-q'];
183 Log(undef, "Sanity check is `@$cmd`");
184 srun(["srun", "--nodes=\Q$ENV{SLURM_NNODES}\E", "--ntasks-per-node=1"],
188 Log(undef, "Sanity check failed: ".exit_status_s($?));
191 Log(undef, "Sanity check OK");
194 my $User = api_call("users/current");
197 if (!$force_unlock) {
198 # Claim this job, and make sure nobody else does
199 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
201 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
210 map { croak ("No $_ specified") unless $Job->{$_} }
211 qw(script script_version script_parameters);
214 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
215 $Job->{'started_at'} = gmtime;
216 $Job->{'state'} = 'Running';
218 $Job = api_call("jobs/create", job => $Job);
220 $job_id = $Job->{'uuid'};
222 my $keep_logfile = $job_id . '.log.txt';
223 log_writer_start($keep_logfile);
225 $Job->{'runtime_constraints'} ||= {};
226 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
227 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
229 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
231 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
232 chomp($gem_versions);
233 chop($gem_versions); # Closing parentheses
238 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
240 Log (undef, "check slurm allocation");
243 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
247 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
248 push @sinfo, "$localcpus localhost";
250 if (exists $ENV{SLURM_NODELIST})
252 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
256 my ($ncpus, $slurm_nodelist) = split;
257 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
260 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
263 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
266 foreach (split (",", $ranges))
279 push @nodelist, map {
281 $n =~ s/\[[-,\d]+\]/$_/;
288 push @nodelist, $nodelist;
291 foreach my $nodename (@nodelist)
293 Log (undef, "node $nodename - $ncpus slots");
294 my $node = { name => $nodename,
298 foreach my $cpu (1..$ncpus)
300 push @slot, { node => $node,
304 push @node, @nodelist;
309 # Ensure that we get one jobstep running on each allocated node before
310 # we start overloading nodes with concurrent steps
312 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
315 $Job->update_attributes(
316 'tasks_summary' => { 'failed' => 0,
321 Log (undef, "start");
322 $SIG{'INT'} = sub { $main::please_freeze = 1; };
323 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
324 $SIG{'TERM'} = \&croak;
325 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
326 $SIG{'ALRM'} = sub { $main::please_info = 1; };
327 $SIG{'CONT'} = sub { $main::please_continue = 1; };
328 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
330 $main::please_freeze = 0;
331 $main::please_info = 0;
332 $main::please_continue = 0;
333 $main::please_refresh = 0;
334 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
336 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
337 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
338 $ENV{"JOB_UUID"} = $job_id;
341 my @jobstep_todo = ();
342 my @jobstep_done = ();
343 my @jobstep_tomerge = ();
344 my $jobstep_tomerge_level = 0;
345 my $squeue_checked = 0;
346 my $latest_refresh = scalar time;
350 if (defined $Job->{thawedfromkey})
352 thaw ($Job->{thawedfromkey});
356 my $first_task = api_call("job_tasks/create", job_task => {
357 'job_uuid' => $Job->{'uuid'},
362 push @jobstep, { 'level' => 0,
364 'arvados_task' => $first_task,
366 push @jobstep_todo, 0;
372 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
375 my $build_script = handle_readall(\*DATA);
376 my $nodelist = join(",", @node);
377 my $git_tar_count = 0;
379 if (!defined $no_clear_tmp) {
380 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
381 Log (undef, "Clean work dirs");
383 my $cleanpid = fork();
386 # Find FUSE mounts that look like Keep mounts (the mount path has the
387 # word "keep") and unmount them. Then clean up work directories.
388 # TODO: When #5036 is done and widely deployed, we can get rid of the
389 # regular expression and just unmount everything with type fuse.keep.
390 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
391 ['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']);
396 last if $cleanpid == waitpid (-1, WNOHANG);
397 freeze_if_want_freeze ($cleanpid);
398 select (undef, undef, undef, 0.1);
400 Log (undef, "Cleanup command exited ".exit_status_s($?));
403 # If this job requires a Docker image, install that.
404 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem);
405 if ($docker_locator = $Job->{docker_image_locator}) {
406 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
409 croak("No Docker image hash found from locator $docker_locator");
411 $docker_stream =~ s/^\.//;
412 my $docker_install_script = qq{
413 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
414 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
417 my $docker_pid = fork();
418 if ($docker_pid == 0)
420 srun (["srun", "--nodelist=" . join(',', @node)],
421 ["/bin/sh", "-ec", $docker_install_script]);
426 last if $docker_pid == waitpid (-1, WNOHANG);
427 freeze_if_want_freeze ($docker_pid);
428 select (undef, undef, undef, 0.1);
432 croak("Installing Docker image from $docker_locator exited "
436 # Determine whether this version of Docker supports memory+swap limits.
437 srun(["srun", "--nodelist=" . $node[0]],
438 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
440 $docker_limitmem = ($? == 0);
442 if ($Job->{arvados_sdk_version}) {
443 # The job also specifies an Arvados SDK version. Add the SDKs to the
444 # tar file for the build script to install.
445 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
446 $Job->{arvados_sdk_version}));
447 add_git_archive("git", "--git-dir=$git_dir", "archive",
448 "--prefix=.arvados.sdk/",
449 $Job->{arvados_sdk_version}, "sdk");
453 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
454 # If script_version looks like an absolute path, *and* the --git-dir
455 # argument was not given -- which implies we were not invoked by
456 # crunch-dispatch -- we will use the given path as a working
457 # directory instead of resolving script_version to a git commit (or
458 # doing anything else with git).
459 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
460 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
463 # Resolve the given script_version to a git commit sha1. Also, if
464 # the repository is remote, clone it into our local filesystem: this
465 # ensures "git archive" will work, and is necessary to reliably
466 # resolve a symbolic script_version like "master^".
467 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
469 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
471 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
473 # If we're running under crunch-dispatch, it will have already
474 # pulled the appropriate source tree into its own repository, and
475 # given us that repo's path as $git_dir.
477 # If we're running a "local" job, we might have to fetch content
478 # from a remote repository.
480 # (Currently crunch-dispatch gives a local path with --git-dir, but
481 # we might as well accept URLs there too in case it changes its
483 my $repo = $git_dir || $Job->{'repository'};
485 # Repository can be remote or local. If remote, we'll need to fetch it
486 # to a local dir before doing `git log` et al.
489 if ($repo =~ m{://|^[^/]*:}) {
490 # $repo is a git url we can clone, like git:// or https:// or
491 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
492 # not recognized here because distinguishing that from a local
493 # path is too fragile. If you really need something strange here,
494 # use the ssh:// form.
495 $repo_location = 'remote';
496 } elsif ($repo =~ m{^\.*/}) {
497 # $repo is a local path to a git index. We'll also resolve ../foo
498 # to ../foo/.git if the latter is a directory. To help
499 # disambiguate local paths from named hosted repositories, this
500 # form must be given as ./ or ../ if it's a relative path.
501 if (-d "$repo/.git") {
502 $repo = "$repo/.git";
504 $repo_location = 'local';
506 # $repo is none of the above. It must be the name of a hosted
508 my $arv_repo_list = api_call("repositories/list",
509 'filters' => [['name','=',$repo]]);
510 my @repos_found = @{$arv_repo_list->{'items'}};
511 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
513 Log(undef, "Repository '$repo' -> "
514 . join(", ", map { $_->{'uuid'} } @repos_found));
517 croak("Error: Found $n_found repositories with name '$repo'.");
519 $repo = $repos_found[0]->{'fetch_url'};
520 $repo_location = 'remote';
522 Log(undef, "Using $repo_location repository '$repo'");
523 $ENV{"CRUNCH_SRC_URL"} = $repo;
525 # Resolve given script_version (we'll call that $treeish here) to a
526 # commit sha1 ($commit).
527 my $treeish = $Job->{'script_version'};
529 if ($repo_location eq 'remote') {
530 # We minimize excess object-fetching by re-using the same bare
531 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
532 # just keep adding remotes to it as needed.
533 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
534 my $gitcmd = "git --git-dir=\Q$local_repo\E";
536 # Set up our local repo for caching remote objects, making
538 if (!-d $local_repo) {
539 make_path($local_repo) or croak("Error: could not create $local_repo");
541 # This works (exits 0 and doesn't delete fetched objects) even
542 # if $local_repo is already initialized:
543 `$gitcmd init --bare`;
545 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
548 # If $treeish looks like a hash (or abbrev hash) we look it up in
549 # our local cache first, since that's cheaper. (We don't want to
550 # do that with tags/branches though -- those change over time, so
551 # they should always be resolved by the remote repo.)
552 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
553 # Hide stderr because it's normal for this to fail:
554 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
556 # Careful not to resolve a branch named abcdeff to commit 1234567:
557 $sha1 =~ /^$treeish/ &&
558 $sha1 =~ /^([0-9a-f]{40})$/s) {
560 Log(undef, "Commit $commit already present in $local_repo");
564 if (!defined $commit) {
565 # If $treeish isn't just a hash or abbrev hash, or isn't here
566 # yet, we need to fetch the remote to resolve it correctly.
568 # First, remove all local heads. This prevents a name that does
569 # not exist on the remote from resolving to (or colliding with)
570 # a previously fetched branch or tag (possibly from a different
572 remove_tree("$local_repo/refs/heads", {keep_root => 1});
574 Log(undef, "Fetching objects from $repo to $local_repo");
575 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
577 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
581 # Now that the data is all here, we will use our local repo for
582 # the rest of our git activities.
586 my $gitcmd = "git --git-dir=\Q$repo\E";
587 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
588 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
589 croak("`$gitcmd rev-list` exited "
591 .", '$treeish' not found. Giving up.");
594 Log(undef, "Version $treeish is commit $commit");
596 if ($commit ne $Job->{'script_version'}) {
597 # Record the real commit id in the database, frozentokey, logs,
598 # etc. -- instead of an abbreviation or a branch name which can
599 # become ambiguous or point to a different commit in the future.
600 if (!$Job->update_attributes('script_version' => $commit)) {
601 croak("Error: failed to update job's script_version attribute");
605 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
606 add_git_archive("$gitcmd archive ''\Q$commit\E");
609 my $git_archive = combined_git_archive();
610 if (!defined $git_archive) {
611 Log(undef, "Skip install phase (no git archive)");
613 Log(undef, "Warning: This probably means workers have no source tree!");
618 my $install_script_tries_left = 3;
619 for (my $attempts = 0; $attempts < 3; $attempts++) {
620 Log(undef, "Run install script on all workers");
622 my @srunargs = ("srun",
623 "--nodelist=$nodelist",
624 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
625 my @execargs = ("sh", "-c",
626 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
628 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
629 my ($install_stderr_r, $install_stderr_w);
630 pipe $install_stderr_r, $install_stderr_w or croak("pipe() failed: $!");
631 set_nonblocking($install_stderr_r);
632 my $installpid = fork();
633 if ($installpid == 0)
635 close($install_stderr_r);
636 fcntl($install_stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
637 open(STDOUT, ">&", $install_stderr_w);
638 open(STDERR, ">&", $install_stderr_w);
639 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
642 close($install_stderr_w);
643 # Tell freeze_if_want_freeze how to kill the child, otherwise the
644 # "waitpid(installpid)" loop won't get interrupted by a freeze:
645 $proc{$installpid} = {};
647 # Track whether anything appears on stderr other than slurm errors
648 # ("srun: ...") and the "starting: ..." message printed by the
649 # srun subroutine itself:
650 my $stderr_anything_from_script = 0;
651 my $match_our_own_errors = '^(srun: error: |starting: \[)';
652 while ($installpid != waitpid(-1, WNOHANG)) {
653 freeze_if_want_freeze ($installpid);
654 # Wait up to 0.1 seconds for something to appear on stderr, then
655 # do a non-blocking read.
656 my $bits = fhbits($install_stderr_r);
657 select ($bits, undef, $bits, 0.1);
658 if (0 < sysread ($install_stderr_r, $stderr_buf, 8192, length($stderr_buf)))
660 while ($stderr_buf =~ /^(.*?)\n/) {
662 substr $stderr_buf, 0, 1+length($line), "";
663 Log(undef, "stderr $line");
664 if ($line !~ /$match_our_own_errors/) {
665 $stderr_anything_from_script = 1;
670 delete $proc{$installpid};
671 $install_exited = $?;
672 close($install_stderr_r);
673 if (length($stderr_buf) > 0) {
674 if ($stderr_buf !~ /$match_our_own_errors/) {
675 $stderr_anything_from_script = 1;
677 Log(undef, "stderr $stderr_buf")
680 Log (undef, "Install script exited ".exit_status_s($install_exited));
681 last if $install_exited == 0 || $main::please_freeze;
682 # If the install script fails but doesn't print an error message,
683 # the next thing anyone is likely to do is just run it again in
684 # case it was a transient problem like "slurm communication fails
685 # because the network isn't reliable enough". So we'll just do
686 # that ourselves (up to 3 attempts in total). OTOH, if there is an
687 # error message, the problem is more likely to have a real fix and
688 # we should fail the job so the fixing process can start, instead
689 # of doing 2 more attempts.
690 last if $stderr_anything_from_script;
693 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
694 unlink($tar_filename);
697 if ($install_exited != 0) {
702 foreach (qw (script script_version script_parameters runtime_constraints))
706 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
708 foreach (split (/\n/, $Job->{knobs}))
710 Log (undef, "knob " . $_);
715 $main::success = undef;
721 my $thisround_succeeded = 0;
722 my $thisround_failed = 0;
723 my $thisround_failed_multiple = 0;
725 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
726 or $a <=> $b } @jobstep_todo;
727 my $level = $jobstep[$jobstep_todo[0]]->{level};
729 my $initial_tasks_this_level = 0;
730 foreach my $id (@jobstep_todo) {
731 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
734 # If the number of tasks scheduled at this level #T is smaller than the number
735 # of slots available #S, only use the first #T slots, or the first slot on
736 # each node, whichever number is greater.
738 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
739 # based on these numbers. Using fewer slots makes more resources available
740 # to each individual task, which should normally be a better strategy when
741 # there are fewer of them running with less parallelism.
743 # Note that this calculation is not redone if the initial tasks at
744 # this level queue more tasks at the same level. This may harm
745 # overall task throughput for that level.
747 if ($initial_tasks_this_level < @node) {
748 @freeslot = (0..$#node);
749 } elsif ($initial_tasks_this_level < @slot) {
750 @freeslot = (0..$initial_tasks_this_level - 1);
752 @freeslot = (0..$#slot);
754 my $round_num_freeslots = scalar(@freeslot);
756 my %round_max_slots = ();
757 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
758 my $this_slot = $slot[$freeslot[$ii]];
759 my $node_name = $this_slot->{node}->{name};
760 $round_max_slots{$node_name} ||= $this_slot->{cpu};
761 last if (scalar(keys(%round_max_slots)) >= @node);
764 Log(undef, "start level $level with $round_num_freeslots slots");
767 my $progress_is_dirty = 1;
768 my $progress_stats_updated = 0;
770 update_progress_stats();
774 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
776 my $id = $jobstep_todo[$todo_ptr];
777 my $Jobstep = $jobstep[$id];
778 if ($Jobstep->{level} != $level)
783 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
784 set_nonblocking($reader{$id});
786 my $childslot = $freeslot[0];
787 my $childnode = $slot[$childslot]->{node};
788 my $childslotname = join (".",
789 $slot[$childslot]->{node}->{name},
790 $slot[$childslot]->{cpu});
792 my $childpid = fork();
795 $SIG{'INT'} = 'DEFAULT';
796 $SIG{'QUIT'} = 'DEFAULT';
797 $SIG{'TERM'} = 'DEFAULT';
799 foreach (values (%reader))
803 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
804 open(STDOUT,">&writer");
805 open(STDERR,">&writer");
810 delete $ENV{"GNUPGHOME"};
811 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
812 $ENV{"TASK_QSEQUENCE"} = $id;
813 $ENV{"TASK_SEQUENCE"} = $level;
814 $ENV{"JOB_SCRIPT"} = $Job->{script};
815 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
816 $param =~ tr/a-z/A-Z/;
817 $ENV{"JOB_PARAMETER_$param"} = $value;
819 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
820 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
821 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
822 $ENV{"HOME"} = $ENV{"TASK_WORK"};
823 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
824 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
825 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
826 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
832 "--nodelist=".$childnode->{name},
833 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
834 "--job-name=$job_id.$id.$$",
837 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
838 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
839 ."&& cd $ENV{CRUNCH_TMP} "
840 # These environment variables get used explicitly later in
841 # $command. No tool is expected to read these values directly.
842 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
843 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
844 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
845 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
846 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
849 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
850 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
851 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
852 # We only set memory limits if Docker lets us limit both memory and swap.
853 # Memory limits alone have been supported longer, but subprocesses tend
854 # to get SIGKILL if they exceed that without any swap limit set.
855 # See #5642 for additional background.
856 if ($docker_limitmem) {
857 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
860 # Dynamically configure the container to use the host system as its
861 # DNS server. Get the host's global addresses from the ip command,
862 # and turn them into docker --dns options using gawk.
864 q{$(ip -o address show scope global |
865 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
867 # The source tree and $destdir directory (which we have
868 # installed on the worker host) are available in the container,
869 # under the same path.
870 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
871 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
873 # Currently, we make arv-mount's mount point appear at /keep
874 # inside the container (instead of using the same path as the
875 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
876 # crunch scripts and utilities must not rely on this. They must
877 # use $TASK_KEEPMOUNT.
878 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
879 $ENV{TASK_KEEPMOUNT} = "/keep";
881 # TASK_WORK is almost exactly like a docker data volume: it
882 # starts out empty, is writable, and persists until no
883 # containers use it any more. We don't use --volumes-from to
884 # share it with other containers: it is only accessible to this
885 # task, and it goes away when this task stops.
887 # However, a docker data volume is writable only by root unless
888 # the mount point already happens to exist in the container with
889 # different permissions. Therefore, we [1] assume /tmp already
890 # exists in the image and is writable by the crunch user; [2]
891 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
892 # writable if they are created by docker while setting up the
893 # other --volumes); and [3] create $TASK_WORK inside the
894 # container using $build_script.
895 $command .= "--volume=/tmp ";
896 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
897 $ENV{"HOME"} = $ENV{"TASK_WORK"};
898 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
900 # TODO: Share a single JOB_WORK volume across all task
901 # containers on a given worker node, and delete it when the job
902 # ends (and, in case that doesn't work, when the next job
905 # For now, use the same approach as TASK_WORK above.
906 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
908 while (my ($env_key, $env_val) = each %ENV)
910 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
911 $command .= "--env=\Q$env_key=$env_val\E ";
914 $command .= "--env=\QHOME=$ENV{HOME}\E ";
915 $command .= "\Q$docker_hash\E ";
916 $command .= "stdbuf --output=0 --error=0 ";
917 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
920 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
921 $command .= "stdbuf --output=0 --error=0 ";
922 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
925 my @execargs = ('bash', '-c', $command);
926 srun (\@srunargs, \@execargs, undef, $build_script);
927 # exec() failed, we assume nothing happened.
928 die "srun() failed on build script\n";
931 if (!defined $childpid)
938 $proc{$childpid} = { jobstep => $id,
941 jobstepname => "$job_id.$id.$childpid",
943 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
944 $slot[$childslot]->{pid} = $childpid;
946 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
947 Log ($id, "child $childpid started on $childslotname");
948 $Jobstep->{starttime} = time;
949 $Jobstep->{node} = $childnode->{name};
950 $Jobstep->{slotindex} = $childslot;
951 delete $Jobstep->{stderr};
952 delete $Jobstep->{finishtime};
954 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
955 $Jobstep->{'arvados_task'}->save;
957 splice @jobstep_todo, $todo_ptr, 1;
960 $progress_is_dirty = 1;
964 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
966 last THISROUND if $main::please_freeze || defined($main::success);
967 if ($main::please_info)
969 $main::please_info = 0;
971 create_output_collection();
973 update_progress_stats();
980 check_refresh_wanted();
982 update_progress_stats();
983 select (undef, undef, undef, 0.1);
985 elsif (time - $progress_stats_updated >= 30)
987 update_progress_stats();
989 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
990 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
992 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
993 .($thisround_failed+$thisround_succeeded)
994 .") -- giving up on this round";
995 Log (undef, $message);
999 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1000 for (my $i=$#freeslot; $i>=0; $i--) {
1001 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1002 push @holdslot, (splice @freeslot, $i, 1);
1005 for (my $i=$#holdslot; $i>=0; $i--) {
1006 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1007 push @freeslot, (splice @holdslot, $i, 1);
1011 # give up if no nodes are succeeding
1012 if (!grep { $_->{node}->{losing_streak} == 0 &&
1013 $_->{node}->{hold_count} < 4 } @slot) {
1014 my $message = "Every node has failed -- giving up on this round";
1015 Log (undef, $message);
1022 push @freeslot, splice @holdslot;
1023 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1026 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1029 if ($main::please_continue) {
1030 $main::please_continue = 0;
1033 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1035 if (!reapchildren())
1037 check_refresh_wanted();
1039 update_progress_stats();
1040 select (undef, undef, undef, 0.1);
1041 killem (keys %proc) if $main::please_freeze;
1045 update_progress_stats();
1046 freeze_if_want_freeze();
1049 if (!defined $main::success)
1051 if (@jobstep_todo &&
1052 $thisround_succeeded == 0 &&
1053 ($thisround_failed == 0 || $thisround_failed > 4))
1055 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1056 Log (undef, $message);
1065 goto ONELEVEL if !defined $main::success;
1068 release_allocation();
1070 my $collated_output = &create_output_collection();
1072 if (!$collated_output) {
1073 Log (undef, "Failed to write output collection");
1076 Log(undef, "job output $collated_output");
1077 $Job->update_attributes('output' => $collated_output);
1080 Log (undef, "finish");
1085 if ($collated_output && $main::success) {
1086 $final_state = 'Complete';
1088 $final_state = 'Failed';
1090 $Job->update_attributes('state' => $final_state);
1092 exit (($final_state eq 'Complete') ? 0 : 1);
1096 sub update_progress_stats
1098 $progress_stats_updated = time;
1099 return if !$progress_is_dirty;
1100 my ($todo, $done, $running) = (scalar @jobstep_todo,
1101 scalar @jobstep_done,
1102 scalar @slot - scalar @freeslot - scalar @holdslot);
1103 $Job->{'tasks_summary'} ||= {};
1104 $Job->{'tasks_summary'}->{'todo'} = $todo;
1105 $Job->{'tasks_summary'}->{'done'} = $done;
1106 $Job->{'tasks_summary'}->{'running'} = $running;
1107 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1108 Log (undef, "status: $done done, $running running, $todo todo");
1109 $progress_is_dirty = 0;
1116 my $pid = waitpid (-1, WNOHANG);
1117 return 0 if $pid <= 0;
1119 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1121 . $slot[$proc{$pid}->{slot}]->{cpu});
1122 my $jobstepid = $proc{$pid}->{jobstep};
1123 my $elapsed = time - $proc{$pid}->{time};
1124 my $Jobstep = $jobstep[$jobstepid];
1126 my $childstatus = $?;
1127 my $exitvalue = $childstatus >> 8;
1128 my $exitinfo = "exit ".exit_status_s($childstatus);
1129 $Jobstep->{'arvados_task'}->reload;
1130 my $task_success = $Jobstep->{'arvados_task'}->{success};
1132 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1134 if (!defined $task_success) {
1135 # task did not indicate one way or the other --> fail
1136 $Jobstep->{'arvados_task'}->{success} = 0;
1137 $Jobstep->{'arvados_task'}->save;
1144 $temporary_fail ||= $Jobstep->{node_fail};
1145 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1147 ++$thisround_failed;
1148 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1150 # Check for signs of a failed or misconfigured node
1151 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1152 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1153 # Don't count this against jobstep failure thresholds if this
1154 # node is already suspected faulty and srun exited quickly
1155 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1157 Log ($jobstepid, "blaming failure on suspect node " .
1158 $slot[$proc{$pid}->{slot}]->{node}->{name});
1159 $temporary_fail ||= 1;
1161 ban_node_by_slot($proc{$pid}->{slot});
1164 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1165 ++$Jobstep->{'failures'},
1166 $temporary_fail ? 'temporary' : 'permanent',
1169 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1170 # Give up on this task, and the whole job
1173 # Put this task back on the todo queue
1174 push @jobstep_todo, $jobstepid;
1175 $Job->{'tasks_summary'}->{'failed'}++;
1179 ++$thisround_succeeded;
1180 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1181 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1182 push @jobstep_done, $jobstepid;
1183 Log ($jobstepid, "success in $elapsed seconds");
1185 $Jobstep->{exitcode} = $childstatus;
1186 $Jobstep->{finishtime} = time;
1187 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1188 $Jobstep->{'arvados_task'}->save;
1189 process_stderr ($jobstepid, $task_success);
1190 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1191 length($Jobstep->{'arvados_task'}->{output}),
1192 $Jobstep->{'arvados_task'}->{output}));
1194 close $reader{$jobstepid};
1195 delete $reader{$jobstepid};
1196 delete $slot[$proc{$pid}->{slot}]->{pid};
1197 push @freeslot, $proc{$pid}->{slot};
1200 if ($task_success) {
1202 my $newtask_list = [];
1203 my $newtask_results;
1205 $newtask_results = api_call(
1208 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1210 'order' => 'qsequence',
1211 'offset' => scalar(@$newtask_list),
1213 push(@$newtask_list, @{$newtask_results->{items}});
1214 } while (@{$newtask_results->{items}});
1215 foreach my $arvados_task (@$newtask_list) {
1217 'level' => $arvados_task->{'sequence'},
1219 'arvados_task' => $arvados_task
1221 push @jobstep, $jobstep;
1222 push @jobstep_todo, $#jobstep;
1226 $progress_is_dirty = 1;
1230 sub check_refresh_wanted
1232 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1233 if (@stat && $stat[9] > $latest_refresh) {
1234 $latest_refresh = scalar time;
1235 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1236 for my $attr ('cancelled_at',
1237 'cancelled_by_user_uuid',
1238 'cancelled_by_client_uuid',
1240 $Job->{$attr} = $Job2->{$attr};
1242 if ($Job->{'state'} ne "Running") {
1243 if ($Job->{'state'} eq "Cancelled") {
1244 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1246 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1249 $main::please_freeze = 1;
1256 my $last_squeue_check = $squeue_checked;
1258 # Do not call `squeue` or check the kill list more than once every
1260 return if $last_squeue_check > time - 15;
1261 $squeue_checked = time;
1263 # Look for children from which we haven't received stderr data since
1264 # the last squeue check. If no such children exist, all procs are
1265 # alive and there's no need to even look at squeue.
1267 # As long as the crunchstat poll interval (10s) is shorter than the
1268 # squeue check interval (15s) this should make the squeue check an
1270 my $silent_procs = 0;
1271 for my $jobstep (values %proc)
1273 if ($jobstep->{stderr_at} < $last_squeue_check)
1278 return if $silent_procs == 0;
1280 # use killem() on procs whose killtime is reached
1281 while (my ($pid, $jobstep) = each %proc)
1283 if (exists $jobstep->{killtime}
1284 && $jobstep->{killtime} <= time
1285 && $jobstep->{stderr_at} < $last_squeue_check)
1287 Log($jobstep->{jobstep}, "killing orphaned srun process $pid (task disappeared from slurm queue)");
1294 # here is an opportunity to check for mysterious problems with local procs
1298 # get a list of steps still running
1299 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%i %j' --noheader`;
1302 Log(undef, "warning: squeue exit status $? ($!)");
1307 # which of my jobsteps are running, according to squeue?
1311 if (/^(\d+)\.(\d+) (\S+)/)
1313 if ($1 eq $ENV{SLURM_JOB_ID})
1320 # Check for child procs >60s old and not mentioned by squeue.
1321 while (my ($pid, $jobstep) = each %proc)
1323 if ($jobstep->{time} < time - 60
1324 && $jobstep->{jobstepname}
1325 && !exists $ok{$jobstep->{jobstepname}}
1326 && !exists $jobstep->{killtime})
1328 # According to slurm, this task has ended (successfully or not)
1329 # -- but our srun child hasn't exited. First we must wait (30
1330 # seconds) in case this is just a race between communication
1331 # channels. Then, if our srun child process still hasn't
1332 # terminated, we'll conclude some slurm communication
1333 # error/delay has caused the task to die without notifying srun,
1334 # and we'll kill srun ourselves.
1335 $jobstep->{killtime} = time + 30;
1336 Log($jobstep->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
1342 sub release_allocation
1346 Log (undef, "release job allocation");
1347 system "scancel $ENV{SLURM_JOB_ID}";
1355 foreach my $job (keys %reader)
1358 while (0 < sysread ($reader{$job}, $buf, 8192))
1360 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1361 $jobstep[$job]->{stderr_at} = time;
1362 $jobstep[$job]->{stderr} .= $buf;
1363 preprocess_stderr ($job);
1364 if (length ($jobstep[$job]->{stderr}) > 16384)
1366 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1375 sub preprocess_stderr
1379 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1381 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1382 Log ($job, "stderr $line");
1383 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1385 $main::please_freeze = 1;
1387 elsif ($line =~ /(srun: error: (Node failure on|Unable to create job step|.*: Communication connection failure))|arvados.errors.Keep/) {
1388 $jobstep[$job]->{node_fail} = 1;
1389 ban_node_by_slot($jobstep[$job]->{slotindex});
1398 my $task_success = shift;
1399 preprocess_stderr ($job);
1402 Log ($job, "stderr $_");
1403 } split ("\n", $jobstep[$job]->{stderr});
1410 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1411 Log(undef, "fetch_block run error from arv-get $hash: $!");
1414 my $output_block = "";
1417 my $bytes = sysread($keep, $buf, 1024 * 1024);
1418 if (!defined $bytes) {
1419 Log(undef, "fetch_block read error from arv-get: $!");
1420 $output_block = undef;
1422 } elsif ($bytes == 0) {
1423 # sysread returns 0 at the end of the pipe.
1426 # some bytes were read into buf.
1427 $output_block .= $buf;
1432 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1433 $output_block = undef;
1435 return $output_block;
1438 # Create a collection by concatenating the output of all tasks (each
1439 # task's output is either a manifest fragment, a locator for a
1440 # manifest fragment stored in Keep, or nothing at all). Return the
1441 # portable_data_hash of the new collection.
1442 sub create_output_collection
1444 Log (undef, "collate");
1446 my ($child_out, $child_in);
1447 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1450 print (arvados.api("v1").collections().
1451 create(body={"manifest_text": sys.stdin.read()}).
1452 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1456 my $manifest_size = 0;
1460 my $output = $_->{'arvados_task'}->{output};
1461 next if (!defined($output));
1463 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1464 $next_write = fetch_block($output);
1466 $next_write = $output;
1468 if (defined($next_write)) {
1469 if (!defined(syswrite($child_in, $next_write))) {
1470 # There's been an error writing. Stop the loop.
1471 # We'll log details about the exit code later.
1474 $manifest_size += length($next_write);
1477 my $uuid = $_->{'arvados_task'}->{'uuid'};
1478 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1483 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1486 my $s = IO::Select->new($child_out);
1487 if ($s->can_read(120)) {
1488 sysread($child_out, $joboutput, 1024 * 1024);
1491 Log(undef, "output collection creation exited " . exit_status_s($?));
1497 Log (undef, "timed out while creating output collection");
1498 foreach my $signal (2, 2, 2, 15, 15, 9) {
1499 kill($signal, $pid);
1500 last if waitpid($pid, WNOHANG) == -1;
1514 my $sig = 2; # SIGINT first
1515 if (exists $proc{$_}->{"sent_$sig"} &&
1516 time - $proc{$_}->{"sent_$sig"} > 4)
1518 $sig = 15; # SIGTERM if SIGINT doesn't work
1520 if (exists $proc{$_}->{"sent_$sig"} &&
1521 time - $proc{$_}->{"sent_$sig"} > 4)
1523 $sig = 9; # SIGKILL if SIGTERM doesn't work
1525 if (!exists $proc{$_}->{"sent_$sig"})
1527 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1529 select (undef, undef, undef, 0.1);
1532 kill $sig, $_; # srun wants two SIGINT to really interrupt
1534 $proc{$_}->{"sent_$sig"} = time;
1535 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1545 vec($bits,fileno($_),1) = 1;
1551 # Send log output to Keep via arv-put.
1553 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1554 # $log_pipe_pid is the pid of the arv-put subprocess.
1556 # The only functions that should access these variables directly are:
1558 # log_writer_start($logfilename)
1559 # Starts an arv-put pipe, reading data on stdin and writing it to
1560 # a $logfilename file in an output collection.
1562 # log_writer_send($txt)
1563 # Writes $txt to the output log collection.
1565 # log_writer_finish()
1566 # Closes the arv-put pipe and returns the output that it produces.
1568 # log_writer_is_active()
1569 # Returns a true value if there is currently a live arv-put
1570 # process, false otherwise.
1572 my ($log_pipe_in, $log_pipe_out, $log_pipe_pid);
1574 sub log_writer_start($)
1576 my $logfilename = shift;
1577 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1579 '--portable-data-hash',
1580 '--project-uuid', $Job->{owner_uuid},
1582 '--name', $logfilename,
1583 '--filename', $logfilename,
1587 sub log_writer_send($)
1590 print $log_pipe_in $txt;
1593 sub log_writer_finish()
1595 return unless $log_pipe_pid;
1597 close($log_pipe_in);
1600 my $s = IO::Select->new($log_pipe_out);
1601 if ($s->can_read(120)) {
1602 sysread($log_pipe_out, $arv_put_output, 1024);
1603 chomp($arv_put_output);
1605 Log (undef, "timed out reading from 'arv-put'");
1608 waitpid($log_pipe_pid, 0);
1609 $log_pipe_pid = $log_pipe_in = $log_pipe_out = undef;
1611 Log("log_writer_finish: arv-put exited ".exit_status_s($?))
1614 return $arv_put_output;
1617 sub log_writer_is_active() {
1618 return $log_pipe_pid;
1621 sub Log # ($jobstep_id, $logmessage)
1623 if ($_[1] =~ /\n/) {
1624 for my $line (split (/\n/, $_[1])) {
1629 my $fh = select STDERR; $|=1; select $fh;
1630 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1631 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1634 if (log_writer_is_active() || -t STDERR) {
1635 my @gmtime = gmtime;
1636 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1637 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1639 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1641 if (log_writer_is_active()) {
1642 log_writer_send($datetime . " " . $message);
1649 my ($package, $file, $line) = caller;
1650 my $message = "@_ at $file line $line\n";
1651 Log (undef, $message);
1652 freeze() if @jobstep_todo;
1653 create_output_collection() if @jobstep_todo;
1663 if ($Job->{'state'} eq 'Cancelled') {
1664 $Job->update_attributes('finished_at' => scalar gmtime);
1666 $Job->update_attributes('state' => 'Failed');
1673 my $justcheckpoint = shift; # false if this will be the last meta saved
1674 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1675 return unless log_writer_is_active();
1677 my $loglocator = log_writer_finish();
1678 Log (undef, "log manifest is $loglocator");
1679 $Job->{'log'} = $loglocator;
1680 $Job->update_attributes('log', $loglocator);
1684 sub freeze_if_want_freeze
1686 if ($main::please_freeze)
1688 release_allocation();
1691 # kill some srun procs before freeze+stop
1692 map { $proc{$_} = {} } @_;
1695 killem (keys %proc);
1696 select (undef, undef, undef, 0.1);
1698 while (($died = waitpid (-1, WNOHANG)) > 0)
1700 delete $proc{$died};
1705 create_output_collection();
1715 Log (undef, "Freeze not implemented");
1722 croak ("Thaw not implemented");
1738 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1745 my $srunargs = shift;
1746 my $execargs = shift;
1747 my $opts = shift || {};
1749 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1751 $Data::Dumper::Terse = 1;
1752 $Data::Dumper::Indent = 0;
1753 my $show_cmd = Dumper($args);
1754 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1755 $show_cmd =~ s/\n/ /g;
1756 if ($opts->{fork}) {
1757 Log(undef, "starting: $show_cmd");
1759 # This is a child process: parent is in charge of reading our
1760 # stderr and copying it to Log() if needed.
1761 warn "starting: $show_cmd\n";
1764 if (defined $stdin) {
1765 my $child = open STDIN, "-|";
1766 defined $child or die "no fork: $!";
1768 print $stdin or die $!;
1769 close STDOUT or die $!;
1774 return system (@$args) if $opts->{fork};
1777 warn "ENV size is ".length(join(" ",%ENV));
1778 die "exec failed: $!: @$args";
1782 sub ban_node_by_slot {
1783 # Don't start any new jobsteps on this node for 60 seconds
1785 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1786 $slot[$slotid]->{node}->{hold_count}++;
1787 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1792 my ($lockfile, $error_message) = @_;
1793 open L, ">", $lockfile or croak("$lockfile: $!");
1794 if (!flock L, LOCK_EX|LOCK_NB) {
1795 croak("Can't lock $lockfile: $error_message\n");
1799 sub find_docker_image {
1800 # Given a Keep locator, check to see if it contains a Docker image.
1801 # If so, return its stream name and Docker hash.
1802 # If not, return undef for both values.
1803 my $locator = shift;
1804 my ($streamname, $filename);
1805 my $image = api_call("collections/get", uuid => $locator);
1807 foreach my $line (split(/\n/, $image->{manifest_text})) {
1808 my @tokens = split(/\s+/, $line);
1810 $streamname = shift(@tokens);
1811 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1812 if (defined($filename)) {
1813 return (undef, undef); # More than one file in the Collection.
1815 $filename = (split(/:/, $filedata, 3))[2];
1820 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1821 return ($streamname, $1);
1823 return (undef, undef);
1828 # Calculate the number of times an operation should be retried,
1829 # assuming exponential backoff, and that we're willing to retry as
1830 # long as tasks have been running. Enforce a minimum of 3 retries.
1831 my ($starttime, $endtime, $timediff, $retries);
1833 $starttime = $jobstep[0]->{starttime};
1834 $endtime = $jobstep[-1]->{finishtime};
1836 if (!defined($starttime)) {
1838 } elsif (!defined($endtime)) {
1839 $timediff = time - $starttime;
1841 $timediff = ($endtime - $starttime) - (time - $endtime);
1843 if ($timediff > 0) {
1844 $retries = int(log($timediff) / log(2));
1846 $retries = 1; # Use the minimum.
1848 return ($retries > 3) ? $retries : 3;
1852 # Pass in two function references.
1853 # This method will be called with the remaining arguments.
1854 # If it dies, retry it with exponential backoff until it succeeds,
1855 # or until the current retry_count is exhausted. After each failure
1856 # that can be retried, the second function will be called with
1857 # the current try count (0-based), next try time, and error message.
1858 my $operation = shift;
1859 my $retry_callback = shift;
1860 my $retries = retry_count();
1861 foreach my $try_count (0..$retries) {
1862 my $next_try = time + (2 ** $try_count);
1863 my $result = eval { $operation->(@_); };
1866 } elsif ($try_count < $retries) {
1867 $retry_callback->($try_count, $next_try, $@);
1868 my $sleep_time = $next_try - time;
1869 sleep($sleep_time) if ($sleep_time > 0);
1872 # Ensure the error message ends in a newline, so Perl doesn't add
1873 # retry_op's line number to it.
1879 # Pass in a /-separated API method name, and arguments for it.
1880 # This function will call that method, retrying as needed until
1881 # the current retry_count is exhausted, with a log on the first failure.
1882 my $method_name = shift;
1883 my $log_api_retry = sub {
1884 my ($try_count, $next_try_at, $errmsg) = @_;
1885 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1886 $errmsg =~ s/\s/ /g;
1887 $errmsg =~ s/\s+$//;
1889 if ($next_try_at < time) {
1890 $retry_msg = "Retrying.";
1892 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1893 $retry_msg = "Retrying at $next_try_fmt.";
1895 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1898 foreach my $key (split(/\//, $method_name)) {
1899 $method = $method->{$key};
1901 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1905 # Given a $?, return a human-readable exit code string like "0" or
1906 # "1" or "0 with signal 1" or "1 with signal 11".
1907 my $exitcode = shift;
1908 my $s = $exitcode >> 8;
1909 if ($exitcode & 0x7f) {
1910 $s .= " with signal " . ($exitcode & 0x7f);
1912 if ($exitcode & 0x80) {
1913 $s .= " with core dump";
1918 sub handle_readall {
1919 # Pass in a glob reference to a file handle.
1920 # Read all its contents and return them as a string.
1921 my $fh_glob_ref = shift;
1923 return <$fh_glob_ref>;
1926 sub tar_filename_n {
1928 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1931 sub add_git_archive {
1932 # Pass in a git archive command as a string or list, a la system().
1933 # This method will save its output to be included in the archive sent to the
1937 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1938 croak("Failed to save git archive: $!");
1940 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1942 waitpid($git_pid, 0);
1945 croak("Failed to save git archive: git exited " . exit_status_s($?));
1949 sub combined_git_archive {
1950 # Combine all saved tar archives into a single archive, then return its
1951 # contents in a string. Return undef if no archives have been saved.
1952 if ($git_tar_count < 1) {
1955 my $base_tar_name = tar_filename_n(1);
1956 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
1957 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
1958 if ($tar_exit != 0) {
1959 croak("Error preparing build archive: tar -A exited " .
1960 exit_status_s($tar_exit));
1963 if (!open(GIT_TAR, "<", $base_tar_name)) {
1964 croak("Could not open build archive: $!");
1966 my $tar_contents = handle_readall(\*GIT_TAR);
1968 return $tar_contents;
1971 sub set_nonblocking {
1973 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
1974 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
1980 # This is crunch-job's internal dispatch script. crunch-job running on the API
1981 # server invokes this script on individual compute nodes, or localhost if we're
1982 # running a job locally. It gets called in two modes:
1984 # * No arguments: Installation mode. Read a tar archive from the DATA
1985 # file handle; it includes the Crunch script's source code, and
1986 # maybe SDKs as well. Those should be installed in the proper
1987 # locations. This runs outside of any Docker container, so don't try to
1988 # introspect Crunch's runtime environment.
1990 # * With arguments: Crunch script run mode. This script should set up the
1991 # environment, then run the command specified in the arguments. This runs
1992 # inside any Docker container.
1995 use File::Path qw( make_path remove_tree );
1996 use POSIX qw(getcwd);
1998 use constant TASK_TEMPFAIL => 111;
2000 # Map SDK subdirectories to the path environments they belong to.
2001 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2003 my $destdir = $ENV{"CRUNCH_SRC"};
2004 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2005 my $repo = $ENV{"CRUNCH_SRC_URL"};
2006 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2007 my $job_work = $ENV{"JOB_WORK"};
2008 my $task_work = $ENV{"TASK_WORK"};
2010 open(STDOUT_ORIG, ">&", STDOUT);
2011 open(STDERR_ORIG, ">&", STDERR);
2013 for my $dir ($destdir, $job_work, $task_work) {
2016 -e $dir or die "Failed to create temporary directory ($dir): $!";
2021 remove_tree($task_work, {keep_root => 1});
2024 ### Crunch script run mode
2026 # We want to do routine logging during task 0 only. This gives the user
2027 # the information they need, but avoids repeating the information for every
2030 if ($ENV{TASK_SEQUENCE} eq "0") {
2033 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2039 my $python_src = "$install_dir/python";
2040 my $venv_dir = "$job_work/.arvados.venv";
2041 my $venv_built = -e "$venv_dir/bin/activate";
2042 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2043 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2044 "--python=python2.7", $venv_dir);
2045 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2047 $Log->("Built Python SDK virtualenv");
2050 my $pip_bin = "pip";
2052 $Log->("Running in Python SDK virtualenv");
2053 $pip_bin = "$venv_dir/bin/pip";
2054 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2055 @ARGV = ("/bin/sh", "-ec",
2056 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2057 } elsif (-d $python_src) {
2058 $Log->("Warning: virtualenv not found inside Docker container default " .
2059 "\$PATH. Can't install Python SDK.");
2062 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
2064 $Log->("Using Arvados SDK:");
2065 foreach my $line (split /\n/, $pkgs) {
2069 $Log->("Arvados SDK packages not found");
2072 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2073 my $sdk_path = "$install_dir/$sdk_dir";
2075 if ($ENV{$sdk_envkey}) {
2076 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2078 $ENV{$sdk_envkey} = $sdk_path;
2080 $Log->("Arvados SDK added to %s", $sdk_envkey);
2085 die "Cannot exec `@ARGV`: $!";
2088 ### Installation mode
2089 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2091 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2092 # This exact git archive (source + arvados sdk) is already installed
2093 # here, so there's no need to reinstall it.
2095 # We must consume our DATA section, though: otherwise the process
2096 # feeding it to us will get SIGPIPE.
2098 while (read(DATA, $buf, 65536)) { }
2103 unlink "$destdir.archive_hash";
2107 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2108 local $SIG{PIPE} = "IGNORE";
2109 warn "Extracting archive: $archive_hash\n";
2110 # --ignore-zeros is necessary sometimes: depending on how much NUL
2111 # padding tar -A put on our combined archive (which in turn depends
2112 # on the length of the component archives) tar without
2113 # --ignore-zeros will exit before consuming stdin and cause close()
2114 # to fail on the resulting SIGPIPE.
2115 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2116 die "Error launching 'tar -xC $destdir': $!";
2118 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2119 # get SIGPIPE. We must feed it data incrementally.
2121 while (read(DATA, $tar_input, 65536)) {
2122 print TARX $tar_input;
2125 die "'tar -xC $destdir' exited $?: $!";
2131 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2133 foreach my $sdk_lang (("python",
2134 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2135 if (-d "$sdk_root/$sdk_lang") {
2136 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2137 die "Failed to install $sdk_lang SDK: $!";
2143 my $python_dir = "$install_dir/python";
2144 if ((-d $python_dir) and can_run("python2.7") and
2145 (system("python2.7", "$python_dir/setup.py", "--quiet", "egg_info") != 0)) {
2146 # egg_info failed, probably when it asked git for a build tag.
2147 # Specify no build tag.
2148 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2149 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2153 # Hide messages from the install script (unless it fails: shell_or_die
2154 # will show $destdir.log in that case).
2155 open(STDOUT, ">>", "$destdir.log");
2156 open(STDERR, ">&", STDOUT);
2158 if (-e "$destdir/crunch_scripts/install") {
2159 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2160 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2162 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2163 } elsif (-e "./install.sh") {
2164 shell_or_die (undef, "./install.sh", $install_dir);
2167 if ($archive_hash) {
2168 unlink "$destdir.archive_hash.new";
2169 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2170 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2176 my $command_name = shift;
2177 open(my $which, "-|", "which", $command_name);
2178 while (<$which>) { }
2185 my $exitcode = shift;
2187 if ($ENV{"DEBUG"}) {
2188 print STDERR "@_\n";
2190 if (system (@_) != 0) {
2193 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2194 open STDERR, ">&STDERR_ORIG";
2195 system ("cat $destdir.log >&2");
2196 warn "@_ failed ($err): $exitstatus";
2197 if (defined($exitcode)) {
2201 exit (($code >> 8) || 1);