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 = "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 $local_job = JSON::decode_json($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 || $local_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 $local_job->{$_} }
211 qw(script script_version script_parameters);
214 $local_job->{'is_locked_by_uuid'} = $User->{'uuid'};
215 $local_job->{'started_at'} = gmtime;
216 $local_job->{'state'} = 'Running';
218 $Job = api_call("jobs/create", job => $local_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,
296 # The number of consecutive times a task has been dispatched
297 # to this node and failed.
299 # The number of consecutive times that SLURM has reported
300 # a node failure since the last successful task.
302 # Don't dispatch work to this node until this time
303 # (in seconds since the epoch) has passed.
305 foreach my $cpu (1..$ncpus)
307 push @slot, { node => $node,
311 push @node, @nodelist;
316 # Ensure that we get one jobstep running on each allocated node before
317 # we start overloading nodes with concurrent steps
319 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
322 $Job->update_attributes(
323 'tasks_summary' => { 'failed' => 0,
328 Log (undef, "start");
329 $SIG{'INT'} = sub { $main::please_freeze = 1; };
330 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
331 $SIG{'TERM'} = \&croak;
332 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
333 $SIG{'ALRM'} = sub { $main::please_info = 1; };
334 $SIG{'CONT'} = sub { $main::please_continue = 1; };
335 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
337 $main::please_freeze = 0;
338 $main::please_info = 0;
339 $main::please_continue = 0;
340 $main::please_refresh = 0;
341 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
343 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
344 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
345 $ENV{"JOB_UUID"} = $job_id;
348 my @jobstep_todo = ();
349 my @jobstep_done = ();
350 my @jobstep_tomerge = ();
351 my $jobstep_tomerge_level = 0;
352 my $squeue_checked = 0;
353 my $latest_refresh = scalar time;
357 if (defined $Job->{thawedfromkey})
359 thaw ($Job->{thawedfromkey});
363 my $first_task = api_call("job_tasks/create", job_task => {
364 'job_uuid' => $Job->{'uuid'},
369 push @jobstep, { 'level' => 0,
371 'arvados_task' => $first_task,
373 push @jobstep_todo, 0;
379 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
382 my $build_script = handle_readall(\*DATA);
383 my $nodelist = join(",", @node);
384 my $git_tar_count = 0;
386 if (!defined $no_clear_tmp) {
387 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
388 Log (undef, "Clean work dirs");
390 my $cleanpid = fork();
393 # Find FUSE mounts that look like Keep mounts (the mount path has the
394 # word "keep") and unmount them. Then clean up work directories.
395 # TODO: When #5036 is done and widely deployed, we can get rid of the
396 # regular expression and just unmount everything with type fuse.keep.
397 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
398 ['bash', '-ec', '-o', 'pipefail', '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']);
403 last if $cleanpid == waitpid (-1, WNOHANG);
404 freeze_if_want_freeze ($cleanpid);
405 select (undef, undef, undef, 0.1);
408 Log(undef, "Clean work dirs: exit ".exit_status_s($?));
409 exit(EX_RETRY_UNLOCKED);
413 # If this job requires a Docker image, install that.
414 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem);
415 if ($docker_locator = $Job->{docker_image_locator}) {
416 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
419 croak("No Docker image hash found from locator $docker_locator");
421 $docker_stream =~ s/^\.//;
422 my $docker_install_script = qq{
423 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
424 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
427 my $docker_pid = fork();
428 if ($docker_pid == 0)
430 srun (["srun", "--nodelist=" . join(',', @node)],
431 ["/bin/sh", "-ec", $docker_install_script]);
436 last if $docker_pid == waitpid (-1, WNOHANG);
437 freeze_if_want_freeze ($docker_pid);
438 select (undef, undef, undef, 0.1);
442 croak("Installing Docker image from $docker_locator exited "
446 # Determine whether this version of Docker supports memory+swap limits.
447 srun(["srun", "--nodelist=" . $node[0]],
448 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
450 $docker_limitmem = ($? == 0);
452 if ($Job->{arvados_sdk_version}) {
453 # The job also specifies an Arvados SDK version. Add the SDKs to the
454 # tar file for the build script to install.
455 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
456 $Job->{arvados_sdk_version}));
457 add_git_archive("git", "--git-dir=$git_dir", "archive",
458 "--prefix=.arvados.sdk/",
459 $Job->{arvados_sdk_version}, "sdk");
463 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
464 # If script_version looks like an absolute path, *and* the --git-dir
465 # argument was not given -- which implies we were not invoked by
466 # crunch-dispatch -- we will use the given path as a working
467 # directory instead of resolving script_version to a git commit (or
468 # doing anything else with git).
469 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
470 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
473 # Resolve the given script_version to a git commit sha1. Also, if
474 # the repository is remote, clone it into our local filesystem: this
475 # ensures "git archive" will work, and is necessary to reliably
476 # resolve a symbolic script_version like "master^".
477 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
479 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
481 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
483 # If we're running under crunch-dispatch, it will have already
484 # pulled the appropriate source tree into its own repository, and
485 # given us that repo's path as $git_dir.
487 # If we're running a "local" job, we might have to fetch content
488 # from a remote repository.
490 # (Currently crunch-dispatch gives a local path with --git-dir, but
491 # we might as well accept URLs there too in case it changes its
493 my $repo = $git_dir || $Job->{'repository'};
495 # Repository can be remote or local. If remote, we'll need to fetch it
496 # to a local dir before doing `git log` et al.
499 if ($repo =~ m{://|^[^/]*:}) {
500 # $repo is a git url we can clone, like git:// or https:// or
501 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
502 # not recognized here because distinguishing that from a local
503 # path is too fragile. If you really need something strange here,
504 # use the ssh:// form.
505 $repo_location = 'remote';
506 } elsif ($repo =~ m{^\.*/}) {
507 # $repo is a local path to a git index. We'll also resolve ../foo
508 # to ../foo/.git if the latter is a directory. To help
509 # disambiguate local paths from named hosted repositories, this
510 # form must be given as ./ or ../ if it's a relative path.
511 if (-d "$repo/.git") {
512 $repo = "$repo/.git";
514 $repo_location = 'local';
516 # $repo is none of the above. It must be the name of a hosted
518 my $arv_repo_list = api_call("repositories/list",
519 'filters' => [['name','=',$repo]]);
520 my @repos_found = @{$arv_repo_list->{'items'}};
521 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
523 Log(undef, "Repository '$repo' -> "
524 . join(", ", map { $_->{'uuid'} } @repos_found));
527 croak("Error: Found $n_found repositories with name '$repo'.");
529 $repo = $repos_found[0]->{'fetch_url'};
530 $repo_location = 'remote';
532 Log(undef, "Using $repo_location repository '$repo'");
533 $ENV{"CRUNCH_SRC_URL"} = $repo;
535 # Resolve given script_version (we'll call that $treeish here) to a
536 # commit sha1 ($commit).
537 my $treeish = $Job->{'script_version'};
539 if ($repo_location eq 'remote') {
540 # We minimize excess object-fetching by re-using the same bare
541 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
542 # just keep adding remotes to it as needed.
543 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
544 my $gitcmd = "git --git-dir=\Q$local_repo\E";
546 # Set up our local repo for caching remote objects, making
548 if (!-d $local_repo) {
549 make_path($local_repo) or croak("Error: could not create $local_repo");
551 # This works (exits 0 and doesn't delete fetched objects) even
552 # if $local_repo is already initialized:
553 `$gitcmd init --bare`;
555 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
558 # If $treeish looks like a hash (or abbrev hash) we look it up in
559 # our local cache first, since that's cheaper. (We don't want to
560 # do that with tags/branches though -- those change over time, so
561 # they should always be resolved by the remote repo.)
562 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
563 # Hide stderr because it's normal for this to fail:
564 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
566 # Careful not to resolve a branch named abcdeff to commit 1234567:
567 $sha1 =~ /^$treeish/ &&
568 $sha1 =~ /^([0-9a-f]{40})$/s) {
570 Log(undef, "Commit $commit already present in $local_repo");
574 if (!defined $commit) {
575 # If $treeish isn't just a hash or abbrev hash, or isn't here
576 # yet, we need to fetch the remote to resolve it correctly.
578 # First, remove all local heads. This prevents a name that does
579 # not exist on the remote from resolving to (or colliding with)
580 # a previously fetched branch or tag (possibly from a different
582 remove_tree("$local_repo/refs/heads", {keep_root => 1});
584 Log(undef, "Fetching objects from $repo to $local_repo");
585 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
587 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
591 # Now that the data is all here, we will use our local repo for
592 # the rest of our git activities.
596 my $gitcmd = "git --git-dir=\Q$repo\E";
597 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
598 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
599 croak("`$gitcmd rev-list` exited "
601 .", '$treeish' not found, giving up");
604 Log(undef, "Version $treeish is commit $commit");
606 if ($commit ne $Job->{'script_version'}) {
607 # Record the real commit id in the database, frozentokey, logs,
608 # etc. -- instead of an abbreviation or a branch name which can
609 # become ambiguous or point to a different commit in the future.
610 if (!$Job->update_attributes('script_version' => $commit)) {
611 croak("Error: failed to update job's script_version attribute");
615 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
616 add_git_archive("$gitcmd archive ''\Q$commit\E");
619 my $git_archive = combined_git_archive();
620 if (!defined $git_archive) {
621 Log(undef, "Skip install phase (no git archive)");
623 Log(undef, "Warning: This probably means workers have no source tree!");
628 my $install_script_tries_left = 3;
629 for (my $attempts = 0; $attempts < 3; $attempts++) {
630 Log(undef, "Run install script on all workers");
632 my @srunargs = ("srun",
633 "--nodelist=$nodelist",
634 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
635 my @execargs = ("sh", "-c",
636 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
638 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
639 my ($install_stderr_r, $install_stderr_w);
640 pipe $install_stderr_r, $install_stderr_w or croak("pipe() failed: $!");
641 set_nonblocking($install_stderr_r);
642 my $installpid = fork();
643 if ($installpid == 0)
645 close($install_stderr_r);
646 fcntl($install_stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
647 open(STDOUT, ">&", $install_stderr_w);
648 open(STDERR, ">&", $install_stderr_w);
649 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
652 close($install_stderr_w);
653 # Tell freeze_if_want_freeze how to kill the child, otherwise the
654 # "waitpid(installpid)" loop won't get interrupted by a freeze:
655 $proc{$installpid} = {};
657 # Track whether anything appears on stderr other than slurm errors
658 # ("srun: ...") and the "starting: ..." message printed by the
659 # srun subroutine itself:
660 my $stderr_anything_from_script = 0;
661 my $match_our_own_errors = '^(srun: error: |starting: \[)';
662 while ($installpid != waitpid(-1, WNOHANG)) {
663 freeze_if_want_freeze ($installpid);
664 # Wait up to 0.1 seconds for something to appear on stderr, then
665 # do a non-blocking read.
666 my $bits = fhbits($install_stderr_r);
667 select ($bits, undef, $bits, 0.1);
668 if (0 < sysread ($install_stderr_r, $stderr_buf, 8192, length($stderr_buf)))
670 while ($stderr_buf =~ /^(.*?)\n/) {
672 substr $stderr_buf, 0, 1+length($line), "";
673 Log(undef, "stderr $line");
674 if ($line !~ /$match_our_own_errors/) {
675 $stderr_anything_from_script = 1;
680 delete $proc{$installpid};
681 $install_exited = $?;
682 close($install_stderr_r);
683 if (length($stderr_buf) > 0) {
684 if ($stderr_buf !~ /$match_our_own_errors/) {
685 $stderr_anything_from_script = 1;
687 Log(undef, "stderr $stderr_buf")
690 Log (undef, "Install script exited ".exit_status_s($install_exited));
691 last if $install_exited == 0 || $main::please_freeze;
692 # If the install script fails but doesn't print an error message,
693 # the next thing anyone is likely to do is just run it again in
694 # case it was a transient problem like "slurm communication fails
695 # because the network isn't reliable enough". So we'll just do
696 # that ourselves (up to 3 attempts in total). OTOH, if there is an
697 # error message, the problem is more likely to have a real fix and
698 # we should fail the job so the fixing process can start, instead
699 # of doing 2 more attempts.
700 last if $stderr_anything_from_script;
703 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
704 unlink($tar_filename);
707 if ($install_exited != 0) {
712 foreach (qw (script script_version script_parameters runtime_constraints))
716 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
718 foreach (split (/\n/, $Job->{knobs}))
720 Log (undef, "knob " . $_);
725 $main::success = undef;
731 my $thisround_succeeded = 0;
732 my $thisround_failed = 0;
733 my $thisround_failed_multiple = 0;
734 my $working_slot_count = scalar(@slot);
736 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
737 or $a <=> $b } @jobstep_todo;
738 my $level = $jobstep[$jobstep_todo[0]]->{level};
740 my $initial_tasks_this_level = 0;
741 foreach my $id (@jobstep_todo) {
742 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
745 # If the number of tasks scheduled at this level #T is smaller than the number
746 # of slots available #S, only use the first #T slots, or the first slot on
747 # each node, whichever number is greater.
749 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
750 # based on these numbers. Using fewer slots makes more resources available
751 # to each individual task, which should normally be a better strategy when
752 # there are fewer of them running with less parallelism.
754 # Note that this calculation is not redone if the initial tasks at
755 # this level queue more tasks at the same level. This may harm
756 # overall task throughput for that level.
758 if ($initial_tasks_this_level < @node) {
759 @freeslot = (0..$#node);
760 } elsif ($initial_tasks_this_level < @slot) {
761 @freeslot = (0..$initial_tasks_this_level - 1);
763 @freeslot = (0..$#slot);
765 my $round_num_freeslots = scalar(@freeslot);
767 my %round_max_slots = ();
768 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
769 my $this_slot = $slot[$freeslot[$ii]];
770 my $node_name = $this_slot->{node}->{name};
771 $round_max_slots{$node_name} ||= $this_slot->{cpu};
772 last if (scalar(keys(%round_max_slots)) >= @node);
775 Log(undef, "start level $level with $round_num_freeslots slots");
778 my $progress_is_dirty = 1;
779 my $progress_stats_updated = 0;
781 update_progress_stats();
785 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
787 my $id = $jobstep_todo[$todo_ptr];
788 my $Jobstep = $jobstep[$id];
789 if ($Jobstep->{level} != $level)
794 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
795 set_nonblocking($reader{$id});
797 my $childslot = $freeslot[0];
798 my $childnode = $slot[$childslot]->{node};
799 my $childslotname = join (".",
800 $slot[$childslot]->{node}->{name},
801 $slot[$childslot]->{cpu});
803 my $childpid = fork();
806 $SIG{'INT'} = 'DEFAULT';
807 $SIG{'QUIT'} = 'DEFAULT';
808 $SIG{'TERM'} = 'DEFAULT';
810 foreach (values (%reader))
814 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
815 open(STDOUT,">&writer");
816 open(STDERR,">&writer");
821 delete $ENV{"GNUPGHOME"};
822 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
823 $ENV{"TASK_QSEQUENCE"} = $id;
824 $ENV{"TASK_SEQUENCE"} = $level;
825 $ENV{"JOB_SCRIPT"} = $Job->{script};
826 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
827 $param =~ tr/a-z/A-Z/;
828 $ENV{"JOB_PARAMETER_$param"} = $value;
830 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
831 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
832 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
833 $ENV{"HOME"} = $ENV{"TASK_WORK"};
834 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
835 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
836 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
837 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
843 "--nodelist=".$childnode->{name},
844 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
845 "--job-name=$job_id.$id.$$",
848 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
849 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
850 ."&& cd $ENV{CRUNCH_TMP} "
851 # These environment variables get used explicitly later in
852 # $command. No tool is expected to read these values directly.
853 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
854 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
855 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
856 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
857 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
860 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
861 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
862 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
863 # We only set memory limits if Docker lets us limit both memory and swap.
864 # Memory limits alone have been supported longer, but subprocesses tend
865 # to get SIGKILL if they exceed that without any swap limit set.
866 # See #5642 for additional background.
867 if ($docker_limitmem) {
868 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
871 # Dynamically configure the container to use the host system as its
872 # DNS server. Get the host's global addresses from the ip command,
873 # and turn them into docker --dns options using gawk.
875 q{$(ip -o address show scope global |
876 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
878 # The source tree and $destdir directory (which we have
879 # installed on the worker host) are available in the container,
880 # under the same path.
881 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
882 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
884 # Currently, we make arv-mount's mount point appear at /keep
885 # inside the container (instead of using the same path as the
886 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
887 # crunch scripts and utilities must not rely on this. They must
888 # use $TASK_KEEPMOUNT.
889 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
890 $ENV{TASK_KEEPMOUNT} = "/keep";
892 # TASK_WORK is almost exactly like a docker data volume: it
893 # starts out empty, is writable, and persists until no
894 # containers use it any more. We don't use --volumes-from to
895 # share it with other containers: it is only accessible to this
896 # task, and it goes away when this task stops.
898 # However, a docker data volume is writable only by root unless
899 # the mount point already happens to exist in the container with
900 # different permissions. Therefore, we [1] assume /tmp already
901 # exists in the image and is writable by the crunch user; [2]
902 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
903 # writable if they are created by docker while setting up the
904 # other --volumes); and [3] create $TASK_WORK inside the
905 # container using $build_script.
906 $command .= "--volume=/tmp ";
907 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
908 $ENV{"HOME"} = $ENV{"TASK_WORK"};
909 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
911 # TODO: Share a single JOB_WORK volume across all task
912 # containers on a given worker node, and delete it when the job
913 # ends (and, in case that doesn't work, when the next job
916 # For now, use the same approach as TASK_WORK above.
917 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
919 while (my ($env_key, $env_val) = each %ENV)
921 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
922 $command .= "--env=\Q$env_key=$env_val\E ";
925 $command .= "--env=\QHOME=$ENV{HOME}\E ";
926 $command .= "\Q$docker_hash\E ";
927 $command .= "stdbuf --output=0 --error=0 ";
928 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
931 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
932 $command .= "stdbuf --output=0 --error=0 ";
933 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
936 my @execargs = ('bash', '-c', $command);
937 srun (\@srunargs, \@execargs, undef, $build_script);
938 # exec() failed, we assume nothing happened.
939 die "srun() failed on build script\n";
942 if (!defined $childpid)
949 $proc{$childpid} = { jobstep => $id,
952 jobstepname => "$job_id.$id.$childpid",
954 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
955 $slot[$childslot]->{pid} = $childpid;
957 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
958 Log ($id, "child $childpid started on $childslotname");
959 $Jobstep->{starttime} = time;
960 $Jobstep->{node} = $childnode->{name};
961 $Jobstep->{slotindex} = $childslot;
962 delete $Jobstep->{stderr};
963 delete $Jobstep->{finishtime};
964 delete $Jobstep->{tempfail};
966 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
967 $Jobstep->{'arvados_task'}->save;
969 splice @jobstep_todo, $todo_ptr, 1;
972 $progress_is_dirty = 1;
976 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
978 last THISROUND if $main::please_freeze || defined($main::success);
979 if ($main::please_info)
981 $main::please_info = 0;
983 create_output_collection();
985 update_progress_stats();
992 check_refresh_wanted();
994 update_progress_stats();
995 select (undef, undef, undef, 0.1);
997 elsif (time - $progress_stats_updated >= 30)
999 update_progress_stats();
1001 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1002 $_->{node}->{hold_count} < 4 } @slot);
1003 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1004 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1006 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1007 .($thisround_failed+$thisround_succeeded)
1008 .") -- giving up on this round";
1009 Log (undef, $message);
1013 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1014 for (my $i=$#freeslot; $i>=0; $i--) {
1015 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1016 push @holdslot, (splice @freeslot, $i, 1);
1019 for (my $i=$#holdslot; $i>=0; $i--) {
1020 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1021 push @freeslot, (splice @holdslot, $i, 1);
1025 # give up if no nodes are succeeding
1026 if ($working_slot_count < 1) {
1027 Log(undef, "Every node has failed -- giving up");
1034 push @freeslot, splice @holdslot;
1035 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1038 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1041 if ($main::please_continue) {
1042 $main::please_continue = 0;
1045 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1047 if (!reapchildren())
1049 check_refresh_wanted();
1051 update_progress_stats();
1052 select (undef, undef, undef, 0.1);
1053 killem (keys %proc) if $main::please_freeze;
1057 update_progress_stats();
1058 freeze_if_want_freeze();
1061 if (!defined $main::success)
1063 if (!@jobstep_todo) {
1065 } elsif ($working_slot_count < 1) {
1066 save_output_collection();
1068 exit(EX_RETRY_UNLOCKED);
1069 } elsif ($thisround_succeeded == 0 &&
1070 ($thisround_failed == 0 || $thisround_failed > 4)) {
1071 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1072 Log (undef, $message);
1077 goto ONELEVEL if !defined $main::success;
1080 release_allocation();
1082 my $collated_output = save_output_collection();
1083 Log (undef, "finish");
1088 if ($collated_output && $main::success) {
1089 $final_state = 'Complete';
1091 $final_state = 'Failed';
1093 $Job->update_attributes('state' => $final_state);
1095 exit (($final_state eq 'Complete') ? 0 : 1);
1099 sub update_progress_stats
1101 $progress_stats_updated = time;
1102 return if !$progress_is_dirty;
1103 my ($todo, $done, $running) = (scalar @jobstep_todo,
1104 scalar @jobstep_done,
1105 scalar @slot - scalar @freeslot - scalar @holdslot);
1106 $Job->{'tasks_summary'} ||= {};
1107 $Job->{'tasks_summary'}->{'todo'} = $todo;
1108 $Job->{'tasks_summary'}->{'done'} = $done;
1109 $Job->{'tasks_summary'}->{'running'} = $running;
1110 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1111 Log (undef, "status: $done done, $running running, $todo todo");
1112 $progress_is_dirty = 0;
1119 my $pid = waitpid (-1, WNOHANG);
1120 return 0 if $pid <= 0;
1122 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1124 . $slot[$proc{$pid}->{slot}]->{cpu});
1125 my $jobstepid = $proc{$pid}->{jobstep};
1126 my $elapsed = time - $proc{$pid}->{time};
1127 my $Jobstep = $jobstep[$jobstepid];
1129 my $childstatus = $?;
1130 my $exitvalue = $childstatus >> 8;
1131 my $exitinfo = "exit ".exit_status_s($childstatus);
1132 $Jobstep->{'arvados_task'}->reload;
1133 my $task_success = $Jobstep->{'arvados_task'}->{success};
1135 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1137 if (!defined $task_success) {
1138 # task did not indicate one way or the other --> fail
1139 $Jobstep->{'arvados_task'}->{success} = 0;
1140 $Jobstep->{'arvados_task'}->save;
1147 $temporary_fail ||= $Jobstep->{tempfail};
1148 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1150 ++$thisround_failed;
1151 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1153 # Check for signs of a failed or misconfigured node
1154 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1155 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1156 # Don't count this against jobstep failure thresholds if this
1157 # node is already suspected faulty and srun exited quickly
1158 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1160 Log ($jobstepid, "blaming failure on suspect node " .
1161 $slot[$proc{$pid}->{slot}]->{node}->{name});
1162 $temporary_fail ||= 1;
1164 ban_node_by_slot($proc{$pid}->{slot});
1167 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1168 ++$Jobstep->{'failures'},
1169 $temporary_fail ? 'temporary' : 'permanent',
1172 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1173 # Give up on this task, and the whole job
1176 # Put this task back on the todo queue
1177 push @jobstep_todo, $jobstepid;
1178 $Job->{'tasks_summary'}->{'failed'}++;
1182 ++$thisround_succeeded;
1183 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1184 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1185 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1186 push @jobstep_done, $jobstepid;
1187 Log ($jobstepid, "success in $elapsed seconds");
1189 $Jobstep->{exitcode} = $childstatus;
1190 $Jobstep->{finishtime} = time;
1191 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1192 $Jobstep->{'arvados_task'}->save;
1193 process_stderr ($jobstepid, $task_success);
1194 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1195 length($Jobstep->{'arvados_task'}->{output}),
1196 $Jobstep->{'arvados_task'}->{output}));
1198 close $reader{$jobstepid};
1199 delete $reader{$jobstepid};
1200 delete $slot[$proc{$pid}->{slot}]->{pid};
1201 push @freeslot, $proc{$pid}->{slot};
1204 if ($task_success) {
1206 my $newtask_list = [];
1207 my $newtask_results;
1209 $newtask_results = api_call(
1212 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1214 'order' => 'qsequence',
1215 'offset' => scalar(@$newtask_list),
1217 push(@$newtask_list, @{$newtask_results->{items}});
1218 } while (@{$newtask_results->{items}});
1219 foreach my $arvados_task (@$newtask_list) {
1221 'level' => $arvados_task->{'sequence'},
1223 'arvados_task' => $arvados_task
1225 push @jobstep, $jobstep;
1226 push @jobstep_todo, $#jobstep;
1230 $progress_is_dirty = 1;
1234 sub check_refresh_wanted
1236 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1237 if (@stat && $stat[9] > $latest_refresh) {
1238 $latest_refresh = scalar time;
1239 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1240 for my $attr ('cancelled_at',
1241 'cancelled_by_user_uuid',
1242 'cancelled_by_client_uuid',
1244 $Job->{$attr} = $Job2->{$attr};
1246 if ($Job->{'state'} ne "Running") {
1247 if ($Job->{'state'} eq "Cancelled") {
1248 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1250 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1253 $main::please_freeze = 1;
1260 my $last_squeue_check = $squeue_checked;
1262 # Do not call `squeue` or check the kill list more than once every
1264 return if $last_squeue_check > time - 15;
1265 $squeue_checked = time;
1267 # Look for children from which we haven't received stderr data since
1268 # the last squeue check. If no such children exist, all procs are
1269 # alive and there's no need to even look at squeue.
1271 # As long as the crunchstat poll interval (10s) is shorter than the
1272 # squeue check interval (15s) this should make the squeue check an
1274 my $silent_procs = 0;
1275 for my $jobstep (values %proc)
1277 if ($jobstep->{stderr_at} < $last_squeue_check)
1282 return if $silent_procs == 0;
1284 # use killem() on procs whose killtime is reached
1285 while (my ($pid, $jobstep) = each %proc)
1287 if (exists $jobstep->{killtime}
1288 && $jobstep->{killtime} <= time
1289 && $jobstep->{stderr_at} < $last_squeue_check)
1292 if ($jobstep->{stderr_at}) {
1293 $sincewhen = " in last " . (time - $jobstep->{stderr_at}) . "s";
1295 Log($jobstep->{jobstep}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1302 # here is an opportunity to check for mysterious problems with local procs
1306 # Get a list of steps still running. Note: squeue(1) says --steps
1307 # selects a format (which we override anyway) and allows us to
1308 # specify which steps we're interested in (which we don't).
1309 # Importantly, it also changes the meaning of %j from "job name" to
1310 # "step name" and (although this isn't mentioned explicitly in the
1311 # docs) switches from "one line per job" mode to "one line per step"
1312 # mode. Without it, we'd just get a list of one job, instead of a
1314 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1317 Log(undef, "warning: squeue exit status $? ($!)");
1322 # which of my jobsteps are running, according to squeue?
1324 for my $jobstepname (@squeue)
1326 $ok{$jobstepname} = 1;
1329 # Check for child procs >60s old and not mentioned by squeue.
1330 while (my ($pid, $jobstep) = each %proc)
1332 if ($jobstep->{time} < time - 60
1333 && $jobstep->{jobstepname}
1334 && !exists $ok{$jobstep->{jobstepname}}
1335 && !exists $jobstep->{killtime})
1337 # According to slurm, this task has ended (successfully or not)
1338 # -- but our srun child hasn't exited. First we must wait (30
1339 # seconds) in case this is just a race between communication
1340 # channels. Then, if our srun child process still hasn't
1341 # terminated, we'll conclude some slurm communication
1342 # error/delay has caused the task to die without notifying srun,
1343 # and we'll kill srun ourselves.
1344 $jobstep->{killtime} = time + 30;
1345 Log($jobstep->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
1351 sub release_allocation
1355 Log (undef, "release job allocation");
1356 system "scancel $ENV{SLURM_JOB_ID}";
1364 foreach my $job (keys %reader)
1367 while (0 < sysread ($reader{$job}, $buf, 8192))
1369 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1370 $jobstep[$job]->{stderr_at} = time;
1371 $jobstep[$job]->{stderr} .= $buf;
1372 preprocess_stderr ($job);
1373 if (length ($jobstep[$job]->{stderr}) > 16384)
1375 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1384 sub preprocess_stderr
1388 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1390 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1391 Log ($job, "stderr $line");
1392 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1394 $main::please_freeze = 1;
1396 elsif ($line =~ /srun: error: Node failure on/) {
1397 my $job_slot_index = $jobstep[$job]->{slotindex};
1398 $slot[$job_slot_index]->{node}->{fail_count}++;
1399 $jobstep[$job]->{tempfail} = 1;
1400 ban_node_by_slot($job_slot_index);
1402 elsif ($line =~ /srun: error: (Unable to create job step|.*: Communication connection failure)/) {
1403 $jobstep[$job]->{tempfail} = 1;
1404 ban_node_by_slot($jobstep[$job]->{slotindex});
1406 elsif ($line =~ /arvados\.errors\.Keep/) {
1407 $jobstep[$job]->{tempfail} = 1;
1416 my $task_success = shift;
1417 preprocess_stderr ($job);
1420 Log ($job, "stderr $_");
1421 } split ("\n", $jobstep[$job]->{stderr});
1428 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1429 Log(undef, "fetch_block run error from arv-get $hash: $!");
1432 my $output_block = "";
1435 my $bytes = sysread($keep, $buf, 1024 * 1024);
1436 if (!defined $bytes) {
1437 Log(undef, "fetch_block read error from arv-get: $!");
1438 $output_block = undef;
1440 } elsif ($bytes == 0) {
1441 # sysread returns 0 at the end of the pipe.
1444 # some bytes were read into buf.
1445 $output_block .= $buf;
1450 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1451 $output_block = undef;
1453 return $output_block;
1456 # Create a collection by concatenating the output of all tasks (each
1457 # task's output is either a manifest fragment, a locator for a
1458 # manifest fragment stored in Keep, or nothing at all). Return the
1459 # portable_data_hash of the new collection.
1460 sub create_output_collection
1462 Log (undef, "collate");
1464 my ($child_out, $child_in);
1465 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1468 print (arvados.api("v1").collections().
1469 create(body={"manifest_text": sys.stdin.read()}).
1470 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1474 my $manifest_size = 0;
1478 my $output = $_->{'arvados_task'}->{output};
1479 next if (!defined($output));
1481 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1482 $next_write = fetch_block($output);
1484 $next_write = $output;
1486 if (defined($next_write)) {
1487 if (!defined(syswrite($child_in, $next_write))) {
1488 # There's been an error writing. Stop the loop.
1489 # We'll log details about the exit code later.
1492 $manifest_size += length($next_write);
1495 my $uuid = $_->{'arvados_task'}->{'uuid'};
1496 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1501 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1504 my $s = IO::Select->new($child_out);
1505 if ($s->can_read(120)) {
1506 sysread($child_out, $joboutput, 1024 * 1024);
1509 Log(undef, "output collection creation exited " . exit_status_s($?));
1515 Log (undef, "timed out while creating output collection");
1516 foreach my $signal (2, 2, 2, 15, 15, 9) {
1517 kill($signal, $pid);
1518 last if waitpid($pid, WNOHANG) == -1;
1527 # Calls create_output_collection, logs the result, and returns it.
1528 # If that was successful, save that as the output in the job record.
1529 sub save_output_collection {
1530 my $collated_output = create_output_collection();
1532 if (!$collated_output) {
1533 Log(undef, "Failed to write output collection");
1536 Log(undef, "job output $collated_output");
1537 $Job->update_attributes('output' => $collated_output);
1539 return $collated_output;
1546 my $sig = 2; # SIGINT first
1547 if (exists $proc{$_}->{"sent_$sig"} &&
1548 time - $proc{$_}->{"sent_$sig"} > 4)
1550 $sig = 15; # SIGTERM if SIGINT doesn't work
1552 if (exists $proc{$_}->{"sent_$sig"} &&
1553 time - $proc{$_}->{"sent_$sig"} > 4)
1555 $sig = 9; # SIGKILL if SIGTERM doesn't work
1557 if (!exists $proc{$_}->{"sent_$sig"})
1559 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1561 select (undef, undef, undef, 0.1);
1564 kill $sig, $_; # srun wants two SIGINT to really interrupt
1566 $proc{$_}->{"sent_$sig"} = time;
1567 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1577 vec($bits,fileno($_),1) = 1;
1583 # Send log output to Keep via arv-put.
1585 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1586 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1587 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1588 # $log_pipe_pid is the pid of the arv-put subprocess.
1590 # The only functions that should access these variables directly are:
1592 # log_writer_start($logfilename)
1593 # Starts an arv-put pipe, reading data on stdin and writing it to
1594 # a $logfilename file in an output collection.
1596 # log_writer_read_output([$timeout])
1597 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1598 # Passes $timeout to the select() call, with a default of 0.01.
1599 # Returns the result of the last read() call on $log_pipe_out, or
1600 # -1 if read() wasn't called because select() timed out.
1601 # Only other log_writer_* functions should need to call this.
1603 # log_writer_send($txt)
1604 # Writes $txt to the output log collection.
1606 # log_writer_finish()
1607 # Closes the arv-put pipe and returns the output that it produces.
1609 # log_writer_is_active()
1610 # Returns a true value if there is currently a live arv-put
1611 # process, false otherwise.
1613 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1616 sub log_writer_start($)
1618 my $logfilename = shift;
1619 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1623 '--filename', $logfilename,
1625 $log_pipe_out_buf = "";
1626 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1629 sub log_writer_read_output {
1630 my $timeout = shift || 0.01;
1632 while ($read && $log_pipe_out_select->can_read($timeout)) {
1633 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1634 length($log_pipe_out_buf));
1636 if (!defined($read)) {
1637 Log(undef, "error reading log manifest from arv-put: $!");
1642 sub log_writer_send($)
1645 print $log_pipe_in $txt;
1646 log_writer_read_output();
1649 sub log_writer_finish()
1651 return unless $log_pipe_pid;
1653 close($log_pipe_in);
1655 my $read_result = log_writer_read_output(120);
1656 if ($read_result == -1) {
1657 Log (undef, "timed out reading from 'arv-put'");
1658 } elsif ($read_result != 0) {
1659 Log(undef, "failed to read arv-put log manifest to EOF");
1662 waitpid($log_pipe_pid, 0);
1664 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1667 close($log_pipe_out);
1668 my $arv_put_output = $log_pipe_out_buf;
1669 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1670 $log_pipe_out_select = undef;
1672 return $arv_put_output;
1675 sub log_writer_is_active() {
1676 return $log_pipe_pid;
1679 sub Log # ($jobstep_id, $logmessage)
1681 if ($_[1] =~ /\n/) {
1682 for my $line (split (/\n/, $_[1])) {
1687 my $fh = select STDERR; $|=1; select $fh;
1688 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1689 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1692 if (log_writer_is_active() || -t STDERR) {
1693 my @gmtime = gmtime;
1694 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1695 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1697 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1699 if (log_writer_is_active()) {
1700 log_writer_send($datetime . " " . $message);
1707 my ($package, $file, $line) = caller;
1708 my $message = "@_ at $file line $line\n";
1709 Log (undef, $message);
1710 freeze() if @jobstep_todo;
1711 create_output_collection() if @jobstep_todo;
1721 if ($Job->{'state'} eq 'Cancelled') {
1722 $Job->update_attributes('finished_at' => scalar gmtime);
1724 $Job->update_attributes('state' => 'Failed');
1731 my $justcheckpoint = shift; # false if this will be the last meta saved
1732 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1733 return unless log_writer_is_active();
1735 my $log_manifest = "";
1737 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1738 $log_manifest .= $prev_log_coll->{manifest_text};
1740 $log_manifest .= log_writer_finish();
1742 my $log_coll = api_call(
1743 "collections/create", ensure_unique_name => 1, collection => {
1744 manifest_text => $log_manifest,
1745 owner_uuid => $Job->{owner_uuid},
1746 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1748 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1749 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1753 sub freeze_if_want_freeze
1755 if ($main::please_freeze)
1757 release_allocation();
1760 # kill some srun procs before freeze+stop
1761 map { $proc{$_} = {} } @_;
1764 killem (keys %proc);
1765 select (undef, undef, undef, 0.1);
1767 while (($died = waitpid (-1, WNOHANG)) > 0)
1769 delete $proc{$died};
1774 create_output_collection();
1784 Log (undef, "Freeze not implemented");
1791 croak ("Thaw not implemented");
1807 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1814 my $srunargs = shift;
1815 my $execargs = shift;
1816 my $opts = shift || {};
1818 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1820 $Data::Dumper::Terse = 1;
1821 $Data::Dumper::Indent = 0;
1822 my $show_cmd = Dumper($args);
1823 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1824 $show_cmd =~ s/\n/ /g;
1825 if ($opts->{fork}) {
1826 Log(undef, "starting: $show_cmd");
1828 # This is a child process: parent is in charge of reading our
1829 # stderr and copying it to Log() if needed.
1830 warn "starting: $show_cmd\n";
1833 if (defined $stdin) {
1834 my $child = open STDIN, "-|";
1835 defined $child or die "no fork: $!";
1837 print $stdin or die $!;
1838 close STDOUT or die $!;
1843 return system (@$args) if $opts->{fork};
1846 warn "ENV size is ".length(join(" ",%ENV));
1847 die "exec failed: $!: @$args";
1851 sub ban_node_by_slot {
1852 # Don't start any new jobsteps on this node for 60 seconds
1854 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1855 $slot[$slotid]->{node}->{hold_count}++;
1856 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1861 my ($lockfile, $error_message) = @_;
1862 open L, ">", $lockfile or croak("$lockfile: $!");
1863 if (!flock L, LOCK_EX|LOCK_NB) {
1864 croak("Can't lock $lockfile: $error_message\n");
1868 sub find_docker_image {
1869 # Given a Keep locator, check to see if it contains a Docker image.
1870 # If so, return its stream name and Docker hash.
1871 # If not, return undef for both values.
1872 my $locator = shift;
1873 my ($streamname, $filename);
1874 my $image = api_call("collections/get", uuid => $locator);
1876 foreach my $line (split(/\n/, $image->{manifest_text})) {
1877 my @tokens = split(/\s+/, $line);
1879 $streamname = shift(@tokens);
1880 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1881 if (defined($filename)) {
1882 return (undef, undef); # More than one file in the Collection.
1884 $filename = (split(/:/, $filedata, 3))[2];
1889 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1890 return ($streamname, $1);
1892 return (undef, undef);
1897 # Calculate the number of times an operation should be retried,
1898 # assuming exponential backoff, and that we're willing to retry as
1899 # long as tasks have been running. Enforce a minimum of 3 retries.
1900 my ($starttime, $endtime, $timediff, $retries);
1902 $starttime = $jobstep[0]->{starttime};
1903 $endtime = $jobstep[-1]->{finishtime};
1905 if (!defined($starttime)) {
1907 } elsif (!defined($endtime)) {
1908 $timediff = time - $starttime;
1910 $timediff = ($endtime - $starttime) - (time - $endtime);
1912 if ($timediff > 0) {
1913 $retries = int(log($timediff) / log(2));
1915 $retries = 1; # Use the minimum.
1917 return ($retries > 3) ? $retries : 3;
1921 # Pass in two function references.
1922 # This method will be called with the remaining arguments.
1923 # If it dies, retry it with exponential backoff until it succeeds,
1924 # or until the current retry_count is exhausted. After each failure
1925 # that can be retried, the second function will be called with
1926 # the current try count (0-based), next try time, and error message.
1927 my $operation = shift;
1928 my $retry_callback = shift;
1929 my $retries = retry_count();
1930 foreach my $try_count (0..$retries) {
1931 my $next_try = time + (2 ** $try_count);
1932 my $result = eval { $operation->(@_); };
1935 } elsif ($try_count < $retries) {
1936 $retry_callback->($try_count, $next_try, $@);
1937 my $sleep_time = $next_try - time;
1938 sleep($sleep_time) if ($sleep_time > 0);
1941 # Ensure the error message ends in a newline, so Perl doesn't add
1942 # retry_op's line number to it.
1948 # Pass in a /-separated API method name, and arguments for it.
1949 # This function will call that method, retrying as needed until
1950 # the current retry_count is exhausted, with a log on the first failure.
1951 my $method_name = shift;
1952 my $log_api_retry = sub {
1953 my ($try_count, $next_try_at, $errmsg) = @_;
1954 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1955 $errmsg =~ s/\s/ /g;
1956 $errmsg =~ s/\s+$//;
1958 if ($next_try_at < time) {
1959 $retry_msg = "Retrying.";
1961 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1962 $retry_msg = "Retrying at $next_try_fmt.";
1964 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1967 foreach my $key (split(/\//, $method_name)) {
1968 $method = $method->{$key};
1970 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1974 # Given a $?, return a human-readable exit code string like "0" or
1975 # "1" or "0 with signal 1" or "1 with signal 11".
1976 my $exitcode = shift;
1977 my $s = $exitcode >> 8;
1978 if ($exitcode & 0x7f) {
1979 $s .= " with signal " . ($exitcode & 0x7f);
1981 if ($exitcode & 0x80) {
1982 $s .= " with core dump";
1987 sub handle_readall {
1988 # Pass in a glob reference to a file handle.
1989 # Read all its contents and return them as a string.
1990 my $fh_glob_ref = shift;
1992 return <$fh_glob_ref>;
1995 sub tar_filename_n {
1997 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2000 sub add_git_archive {
2001 # Pass in a git archive command as a string or list, a la system().
2002 # This method will save its output to be included in the archive sent to the
2006 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2007 croak("Failed to save git archive: $!");
2009 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2011 waitpid($git_pid, 0);
2014 croak("Failed to save git archive: git exited " . exit_status_s($?));
2018 sub combined_git_archive {
2019 # Combine all saved tar archives into a single archive, then return its
2020 # contents in a string. Return undef if no archives have been saved.
2021 if ($git_tar_count < 1) {
2024 my $base_tar_name = tar_filename_n(1);
2025 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2026 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2027 if ($tar_exit != 0) {
2028 croak("Error preparing build archive: tar -A exited " .
2029 exit_status_s($tar_exit));
2032 if (!open(GIT_TAR, "<", $base_tar_name)) {
2033 croak("Could not open build archive: $!");
2035 my $tar_contents = handle_readall(\*GIT_TAR);
2037 return $tar_contents;
2040 sub set_nonblocking {
2042 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2043 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2049 # This is crunch-job's internal dispatch script. crunch-job running on the API
2050 # server invokes this script on individual compute nodes, or localhost if we're
2051 # running a job locally. It gets called in two modes:
2053 # * No arguments: Installation mode. Read a tar archive from the DATA
2054 # file handle; it includes the Crunch script's source code, and
2055 # maybe SDKs as well. Those should be installed in the proper
2056 # locations. This runs outside of any Docker container, so don't try to
2057 # introspect Crunch's runtime environment.
2059 # * With arguments: Crunch script run mode. This script should set up the
2060 # environment, then run the command specified in the arguments. This runs
2061 # inside any Docker container.
2064 use File::Path qw( make_path remove_tree );
2065 use POSIX qw(getcwd);
2067 use constant TASK_TEMPFAIL => 111;
2069 # Map SDK subdirectories to the path environments they belong to.
2070 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2072 my $destdir = $ENV{"CRUNCH_SRC"};
2073 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2074 my $repo = $ENV{"CRUNCH_SRC_URL"};
2075 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2076 my $job_work = $ENV{"JOB_WORK"};
2077 my $task_work = $ENV{"TASK_WORK"};
2079 open(STDOUT_ORIG, ">&", STDOUT);
2080 open(STDERR_ORIG, ">&", STDERR);
2082 for my $dir ($destdir, $job_work, $task_work) {
2085 -e $dir or die "Failed to create temporary directory ($dir): $!";
2090 remove_tree($task_work, {keep_root => 1});
2093 ### Crunch script run mode
2095 # We want to do routine logging during task 0 only. This gives the user
2096 # the information they need, but avoids repeating the information for every
2099 if ($ENV{TASK_SEQUENCE} eq "0") {
2102 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2108 my $python_src = "$install_dir/python";
2109 my $venv_dir = "$job_work/.arvados.venv";
2110 my $venv_built = -e "$venv_dir/bin/activate";
2111 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2112 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2113 "--python=python2.7", $venv_dir);
2114 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2116 $Log->("Built Python SDK virtualenv");
2119 my $pip_bin = "pip";
2121 $Log->("Running in Python SDK virtualenv");
2122 $pip_bin = "$venv_dir/bin/pip";
2123 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2124 @ARGV = ("/bin/sh", "-ec",
2125 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2126 } elsif (-d $python_src) {
2127 $Log->("Warning: virtualenv not found inside Docker container default " .
2128 "\$PATH. Can't install Python SDK.");
2131 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
2133 $Log->("Using Arvados SDK:");
2134 foreach my $line (split /\n/, $pkgs) {
2138 $Log->("Arvados SDK packages not found");
2141 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2142 my $sdk_path = "$install_dir/$sdk_dir";
2144 if ($ENV{$sdk_envkey}) {
2145 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2147 $ENV{$sdk_envkey} = $sdk_path;
2149 $Log->("Arvados SDK added to %s", $sdk_envkey);
2154 die "Cannot exec `@ARGV`: $!";
2157 ### Installation mode
2158 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2160 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2161 # This exact git archive (source + arvados sdk) is already installed
2162 # here, so there's no need to reinstall it.
2164 # We must consume our DATA section, though: otherwise the process
2165 # feeding it to us will get SIGPIPE.
2167 while (read(DATA, $buf, 65536)) { }
2172 unlink "$destdir.archive_hash";
2176 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2177 local $SIG{PIPE} = "IGNORE";
2178 warn "Extracting archive: $archive_hash\n";
2179 # --ignore-zeros is necessary sometimes: depending on how much NUL
2180 # padding tar -A put on our combined archive (which in turn depends
2181 # on the length of the component archives) tar without
2182 # --ignore-zeros will exit before consuming stdin and cause close()
2183 # to fail on the resulting SIGPIPE.
2184 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2185 die "Error launching 'tar -xC $destdir': $!";
2187 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2188 # get SIGPIPE. We must feed it data incrementally.
2190 while (read(DATA, $tar_input, 65536)) {
2191 print TARX $tar_input;
2194 die "'tar -xC $destdir' exited $?: $!";
2200 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2202 foreach my $sdk_lang (("python",
2203 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2204 if (-d "$sdk_root/$sdk_lang") {
2205 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2206 die "Failed to install $sdk_lang SDK: $!";
2212 my $python_dir = "$install_dir/python";
2213 if ((-d $python_dir) and can_run("python2.7")) {
2214 open(my $egg_info_pipe, "-|",
2215 "python2.7 \Q$python_dir/setup.py\E --quiet egg_info 2>&1 >/dev/null");
2216 my @egg_info_errors = <$egg_info_pipe>;
2217 close($egg_info_pipe);
2219 if (@egg_info_errors and ($egg_info_errors[-1] =~ /\bgit\b/)) {
2220 # egg_info apparently failed because it couldn't ask git for a build tag.
2221 # Specify no build tag.
2222 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2223 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2226 my $egg_info_exit = $? >> 8;
2227 foreach my $errline (@egg_info_errors) {
2228 print STDERR_ORIG $errline;
2230 warn "python setup.py egg_info failed: exit $egg_info_exit";
2231 exit ($egg_info_exit || 1);
2236 # Hide messages from the install script (unless it fails: shell_or_die
2237 # will show $destdir.log in that case).
2238 open(STDOUT, ">>", "$destdir.log");
2239 open(STDERR, ">&", STDOUT);
2241 if (-e "$destdir/crunch_scripts/install") {
2242 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2243 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2245 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2246 } elsif (-e "./install.sh") {
2247 shell_or_die (undef, "./install.sh", $install_dir);
2250 if ($archive_hash) {
2251 unlink "$destdir.archive_hash.new";
2252 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2253 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2259 my $command_name = shift;
2260 open(my $which, "-|", "which", $command_name);
2261 while (<$which>) { }
2268 my $exitcode = shift;
2270 if ($ENV{"DEBUG"}) {
2271 print STDERR "@_\n";
2273 if (system (@_) != 0) {
2276 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2277 open STDERR, ">&STDERR_ORIG";
2278 system ("cat $destdir.log >&2");
2279 warn "@_ failed ($err): $exitstatus";
2280 if (defined($exitcode)) {
2284 exit (($code >> 8) || 1);