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"});
127 GetOptions('force-unlock' => \$force_unlock,
128 'git-dir=s' => \$git_dir,
129 'job=s' => \$jobspec,
130 'job-api-token=s' => \$job_api_token,
131 'no-clear-tmp' => \$no_clear_tmp,
132 'resume-stash=s' => \$resume_stash,
135 if (defined $job_api_token) {
136 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
139 my $have_slurm = exists $ENV{SLURM_JOBID} && exists $ENV{SLURM_NODELIST};
145 $main::ENV{CRUNCH_DEBUG} = 1;
149 $main::ENV{CRUNCH_DEBUG} = 0;
154 my $arv = Arvados->new('apiVersion' => 'v1');
162 my $User = api_call("users/current");
164 if ($jobspec =~ /^[-a-z\d]+$/)
166 # $jobspec is an Arvados UUID, not a JSON job specification
167 $Job = api_call("jobs/get", uuid => $jobspec);
168 if (!$force_unlock) {
169 # Claim this job, and make sure nobody else does
170 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
172 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
179 $Job = JSON::decode_json($jobspec);
183 map { croak ("No $_ specified") unless $Job->{$_} }
184 qw(script script_version script_parameters);
187 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
188 $Job->{'started_at'} = gmtime;
189 $Job->{'state'} = 'Running';
191 $Job = api_call("jobs/create", job => $Job);
193 $job_id = $Job->{'uuid'};
195 my $keep_logfile = $job_id . '.log.txt';
196 log_writer_start($keep_logfile);
198 $Job->{'runtime_constraints'} ||= {};
199 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
200 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
202 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
204 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
205 chomp($gem_versions);
206 chop($gem_versions); # Closing parentheses
211 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
213 Log (undef, "check slurm allocation");
216 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
220 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
221 push @sinfo, "$localcpus localhost";
223 if (exists $ENV{SLURM_NODELIST})
225 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
229 my ($ncpus, $slurm_nodelist) = split;
230 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
233 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
236 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
239 foreach (split (",", $ranges))
252 push @nodelist, map {
254 $n =~ s/\[[-,\d]+\]/$_/;
261 push @nodelist, $nodelist;
264 foreach my $nodename (@nodelist)
266 Log (undef, "node $nodename - $ncpus slots");
267 my $node = { name => $nodename,
271 foreach my $cpu (1..$ncpus)
273 push @slot, { node => $node,
277 push @node, @nodelist;
282 # Ensure that we get one jobstep running on each allocated node before
283 # we start overloading nodes with concurrent steps
285 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
288 $Job->update_attributes(
289 'tasks_summary' => { 'failed' => 0,
294 Log (undef, "start");
295 $SIG{'INT'} = sub { $main::please_freeze = 1; };
296 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
297 $SIG{'TERM'} = \&croak;
298 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
299 $SIG{'ALRM'} = sub { $main::please_info = 1; };
300 $SIG{'CONT'} = sub { $main::please_continue = 1; };
301 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
303 $main::please_freeze = 0;
304 $main::please_info = 0;
305 $main::please_continue = 0;
306 $main::please_refresh = 0;
307 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
309 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
310 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
311 $ENV{"JOB_UUID"} = $job_id;
314 my @jobstep_todo = ();
315 my @jobstep_done = ();
316 my @jobstep_tomerge = ();
317 my $jobstep_tomerge_level = 0;
319 my $squeue_kill_checked;
320 my $latest_refresh = scalar time;
324 if (defined $Job->{thawedfromkey})
326 thaw ($Job->{thawedfromkey});
330 my $first_task = api_call("job_tasks/create", job_task => {
331 'job_uuid' => $Job->{'uuid'},
336 push @jobstep, { 'level' => 0,
338 'arvados_task' => $first_task,
340 push @jobstep_todo, 0;
346 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
349 my $build_script = handle_readall(\*DATA);
350 my $nodelist = join(",", @node);
351 my $git_tar_count = 0;
353 if (!defined $no_clear_tmp) {
354 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
355 Log (undef, "Clean work dirs");
357 my $cleanpid = fork();
360 # Find FUSE mounts that look like Keep mounts (the mount path has the
361 # word "keep") and unmount them. Then clean up work directories.
362 # TODO: When #5036 is done and widely deployed, we can get rid of the
363 # regular expression and just unmount everything with type fuse.keep.
364 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
365 ['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']);
370 last if $cleanpid == waitpid (-1, WNOHANG);
371 freeze_if_want_freeze ($cleanpid);
372 select (undef, undef, undef, 0.1);
374 Log (undef, "Cleanup command exited ".exit_status_s($?));
377 # If this job requires a Docker image, install that.
378 my $docker_bin = "/usr/bin/docker.io";
379 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem);
380 if ($docker_locator = $Job->{docker_image_locator}) {
381 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
384 croak("No Docker image hash found from locator $docker_locator");
386 $docker_stream =~ s/^\.//;
387 my $docker_install_script = qq{
388 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
389 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
392 my $docker_pid = fork();
393 if ($docker_pid == 0)
395 srun (["srun", "--nodelist=" . join(',', @node)],
396 ["/bin/sh", "-ec", $docker_install_script]);
401 last if $docker_pid == waitpid (-1, WNOHANG);
402 freeze_if_want_freeze ($docker_pid);
403 select (undef, undef, undef, 0.1);
407 croak("Installing Docker image from $docker_locator exited "
411 # Determine whether this version of Docker supports memory+swap limits.
412 srun(["srun", "--nodelist=" . $node[0]],
413 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
415 $docker_limitmem = ($? == 0);
417 if ($Job->{arvados_sdk_version}) {
418 # The job also specifies an Arvados SDK version. Add the SDKs to the
419 # tar file for the build script to install.
420 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
421 $Job->{arvados_sdk_version}));
422 add_git_archive("git", "--git-dir=$git_dir", "archive",
423 "--prefix=.arvados.sdk/",
424 $Job->{arvados_sdk_version}, "sdk");
428 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
429 # If script_version looks like an absolute path, *and* the --git-dir
430 # argument was not given -- which implies we were not invoked by
431 # crunch-dispatch -- we will use the given path as a working
432 # directory instead of resolving script_version to a git commit (or
433 # doing anything else with git).
434 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
435 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
438 # Resolve the given script_version to a git commit sha1. Also, if
439 # the repository is remote, clone it into our local filesystem: this
440 # ensures "git archive" will work, and is necessary to reliably
441 # resolve a symbolic script_version like "master^".
442 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
444 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
446 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
448 # If we're running under crunch-dispatch, it will have already
449 # pulled the appropriate source tree into its own repository, and
450 # given us that repo's path as $git_dir.
452 # If we're running a "local" job, we might have to fetch content
453 # from a remote repository.
455 # (Currently crunch-dispatch gives a local path with --git-dir, but
456 # we might as well accept URLs there too in case it changes its
458 my $repo = $git_dir || $Job->{'repository'};
460 # Repository can be remote or local. If remote, we'll need to fetch it
461 # to a local dir before doing `git log` et al.
464 if ($repo =~ m{://|^[^/]*:}) {
465 # $repo is a git url we can clone, like git:// or https:// or
466 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
467 # not recognized here because distinguishing that from a local
468 # path is too fragile. If you really need something strange here,
469 # use the ssh:// form.
470 $repo_location = 'remote';
471 } elsif ($repo =~ m{^\.*/}) {
472 # $repo is a local path to a git index. We'll also resolve ../foo
473 # to ../foo/.git if the latter is a directory. To help
474 # disambiguate local paths from named hosted repositories, this
475 # form must be given as ./ or ../ if it's a relative path.
476 if (-d "$repo/.git") {
477 $repo = "$repo/.git";
479 $repo_location = 'local';
481 # $repo is none of the above. It must be the name of a hosted
483 my $arv_repo_list = api_call("repositories/list",
484 'filters' => [['name','=',$repo]]);
485 my @repos_found = @{$arv_repo_list->{'items'}};
486 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
488 Log(undef, "Repository '$repo' -> "
489 . join(", ", map { $_->{'uuid'} } @repos_found));
492 croak("Error: Found $n_found repositories with name '$repo'.");
494 $repo = $repos_found[0]->{'fetch_url'};
495 $repo_location = 'remote';
497 Log(undef, "Using $repo_location repository '$repo'");
498 $ENV{"CRUNCH_SRC_URL"} = $repo;
500 # Resolve given script_version (we'll call that $treeish here) to a
501 # commit sha1 ($commit).
502 my $treeish = $Job->{'script_version'};
504 if ($repo_location eq 'remote') {
505 # We minimize excess object-fetching by re-using the same bare
506 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
507 # just keep adding remotes to it as needed.
508 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
509 my $gitcmd = "git --git-dir=\Q$local_repo\E";
511 # Set up our local repo for caching remote objects, making
513 if (!-d $local_repo) {
514 make_path($local_repo) or croak("Error: could not create $local_repo");
516 # This works (exits 0 and doesn't delete fetched objects) even
517 # if $local_repo is already initialized:
518 `$gitcmd init --bare`;
520 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
523 # If $treeish looks like a hash (or abbrev hash) we look it up in
524 # our local cache first, since that's cheaper. (We don't want to
525 # do that with tags/branches though -- those change over time, so
526 # they should always be resolved by the remote repo.)
527 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
528 # Hide stderr because it's normal for this to fail:
529 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
531 # Careful not to resolve a branch named abcdeff to commit 1234567:
532 $sha1 =~ /^$treeish/ &&
533 $sha1 =~ /^([0-9a-f]{40})$/s) {
535 Log(undef, "Commit $commit already present in $local_repo");
539 if (!defined $commit) {
540 # If $treeish isn't just a hash or abbrev hash, or isn't here
541 # yet, we need to fetch the remote to resolve it correctly.
543 # First, remove all local heads. This prevents a name that does
544 # not exist on the remote from resolving to (or colliding with)
545 # a previously fetched branch or tag (possibly from a different
547 remove_tree("$local_repo/refs/heads", {keep_root => 1});
549 Log(undef, "Fetching objects from $repo to $local_repo");
550 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
552 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
556 # Now that the data is all here, we will use our local repo for
557 # the rest of our git activities.
561 my $gitcmd = "git --git-dir=\Q$repo\E";
562 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
563 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
564 croak("`$gitcmd rev-list` exited "
566 .", '$treeish' not found. Giving up.");
569 Log(undef, "Version $treeish is commit $commit");
571 if ($commit ne $Job->{'script_version'}) {
572 # Record the real commit id in the database, frozentokey, logs,
573 # etc. -- instead of an abbreviation or a branch name which can
574 # become ambiguous or point to a different commit in the future.
575 if (!$Job->update_attributes('script_version' => $commit)) {
576 croak("Error: failed to update job's script_version attribute");
580 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
581 add_git_archive("$gitcmd archive ''\Q$commit\E");
584 my $git_archive = combined_git_archive();
585 if (!defined $git_archive) {
586 Log(undef, "Skip install phase (no git archive)");
588 Log(undef, "Warning: This probably means workers have no source tree!");
592 Log(undef, "Run install script on all workers");
594 my @srunargs = ("srun",
595 "--nodelist=$nodelist",
596 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
597 my @execargs = ("sh", "-c",
598 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
600 my $installpid = fork();
601 if ($installpid == 0)
603 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
608 last if $installpid == waitpid (-1, WNOHANG);
609 freeze_if_want_freeze ($installpid);
610 select (undef, undef, undef, 0.1);
612 my $install_exited = $?;
613 Log (undef, "Install script exited ".exit_status_s($install_exited));
614 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
615 unlink($tar_filename);
617 exit (1) if $install_exited != 0;
620 foreach (qw (script script_version script_parameters runtime_constraints))
624 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
626 foreach (split (/\n/, $Job->{knobs}))
628 Log (undef, "knob " . $_);
633 $main::success = undef;
639 my $thisround_succeeded = 0;
640 my $thisround_failed = 0;
641 my $thisround_failed_multiple = 0;
643 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
644 or $a <=> $b } @jobstep_todo;
645 my $level = $jobstep[$jobstep_todo[0]]->{level};
647 my $initial_tasks_this_level = 0;
648 foreach my $id (@jobstep_todo) {
649 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
652 # If the number of tasks scheduled at this level #T is smaller than the number
653 # of slots available #S, only use the first #T slots, or the first slot on
654 # each node, whichever number is greater.
656 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
657 # based on these numbers. Using fewer slots makes more resources available
658 # to each individual task, which should normally be a better strategy when
659 # there are fewer of them running with less parallelism.
661 # Note that this calculation is not redone if the initial tasks at
662 # this level queue more tasks at the same level. This may harm
663 # overall task throughput for that level.
665 if ($initial_tasks_this_level < @node) {
666 @freeslot = (0..$#node);
667 } elsif ($initial_tasks_this_level < @slot) {
668 @freeslot = (0..$initial_tasks_this_level - 1);
670 @freeslot = (0..$#slot);
672 my $round_num_freeslots = scalar(@freeslot);
674 my %round_max_slots = ();
675 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
676 my $this_slot = $slot[$freeslot[$ii]];
677 my $node_name = $this_slot->{node}->{name};
678 $round_max_slots{$node_name} ||= $this_slot->{cpu};
679 last if (scalar(keys(%round_max_slots)) >= @node);
682 Log(undef, "start level $level with $round_num_freeslots slots");
686 my $progress_is_dirty = 1;
687 my $progress_stats_updated = 0;
689 update_progress_stats();
693 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
695 my $id = $jobstep_todo[$todo_ptr];
696 my $Jobstep = $jobstep[$id];
697 if ($Jobstep->{level} != $level)
702 pipe $reader{$id}, "writer" or croak ($!);
703 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
704 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
706 my $childslot = $freeslot[0];
707 my $childnode = $slot[$childslot]->{node};
708 my $childslotname = join (".",
709 $slot[$childslot]->{node}->{name},
710 $slot[$childslot]->{cpu});
712 my $childpid = fork();
715 $SIG{'INT'} = 'DEFAULT';
716 $SIG{'QUIT'} = 'DEFAULT';
717 $SIG{'TERM'} = 'DEFAULT';
719 foreach (values (%reader))
723 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
724 open(STDOUT,">&writer");
725 open(STDERR,">&writer");
730 delete $ENV{"GNUPGHOME"};
731 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
732 $ENV{"TASK_QSEQUENCE"} = $id;
733 $ENV{"TASK_SEQUENCE"} = $level;
734 $ENV{"JOB_SCRIPT"} = $Job->{script};
735 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
736 $param =~ tr/a-z/A-Z/;
737 $ENV{"JOB_PARAMETER_$param"} = $value;
739 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
740 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
741 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
742 $ENV{"HOME"} = $ENV{"TASK_WORK"};
743 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
744 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
745 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
746 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
752 "--nodelist=".$childnode->{name},
753 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
754 "--job-name=$job_id.$id.$$",
757 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
758 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
759 ."&& cd $ENV{CRUNCH_TMP} "
760 # These environment variables get used explicitly later in
761 # $command. No tool is expected to read these values directly.
762 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
763 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
764 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
765 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
766 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
769 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
770 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
771 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
772 # We only set memory limits if Docker lets us limit both memory and swap.
773 # Memory limits alone have been supported longer, but subprocesses tend
774 # to get SIGKILL if they exceed that without any swap limit set.
775 # See #5642 for additional background.
776 if ($docker_limitmem) {
777 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
780 # Dynamically configure the container to use the host system as its
781 # DNS server. Get the host's global addresses from the ip command,
782 # and turn them into docker --dns options using gawk.
784 q{$(ip -o address show scope global |
785 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
787 # The source tree and $destdir directory (which we have
788 # installed on the worker host) are available in the container,
789 # under the same path.
790 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
791 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
793 # Currently, we make arv-mount's mount point appear at /keep
794 # inside the container (instead of using the same path as the
795 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
796 # crunch scripts and utilities must not rely on this. They must
797 # use $TASK_KEEPMOUNT.
798 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
799 $ENV{TASK_KEEPMOUNT} = "/keep";
801 # TASK_WORK is almost exactly like a docker data volume: it
802 # starts out empty, is writable, and persists until no
803 # containers use it any more. We don't use --volumes-from to
804 # share it with other containers: it is only accessible to this
805 # task, and it goes away when this task stops.
807 # However, a docker data volume is writable only by root unless
808 # the mount point already happens to exist in the container with
809 # different permissions. Therefore, we [1] assume /tmp already
810 # exists in the image and is writable by the crunch user; [2]
811 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
812 # writable if they are created by docker while setting up the
813 # other --volumes); and [3] create $TASK_WORK inside the
814 # container using $build_script.
815 $command .= "--volume=/tmp ";
816 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
817 $ENV{"HOME"} = $ENV{"TASK_WORK"};
818 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
820 # TODO: Share a single JOB_WORK volume across all task
821 # containers on a given worker node, and delete it when the job
822 # ends (and, in case that doesn't work, when the next job
825 # For now, use the same approach as TASK_WORK above.
826 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
828 while (my ($env_key, $env_val) = each %ENV)
830 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
831 $command .= "--env=\Q$env_key=$env_val\E ";
834 $command .= "--env=\QHOME=$ENV{HOME}\E ";
835 $command .= "\Q$docker_hash\E ";
836 $command .= "stdbuf --output=0 --error=0 ";
837 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
840 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
841 $command .= "stdbuf --output=0 --error=0 ";
842 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
845 my @execargs = ('bash', '-c', $command);
846 srun (\@srunargs, \@execargs, undef, $build_script);
847 # exec() failed, we assume nothing happened.
848 die "srun() failed on build script\n";
851 if (!defined $childpid)
858 $proc{$childpid} = { jobstep => $id,
861 jobstepname => "$job_id.$id.$childpid",
863 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
864 $slot[$childslot]->{pid} = $childpid;
866 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
867 Log ($id, "child $childpid started on $childslotname");
868 $Jobstep->{starttime} = time;
869 $Jobstep->{node} = $childnode->{name};
870 $Jobstep->{slotindex} = $childslot;
871 delete $Jobstep->{stderr};
872 delete $Jobstep->{finishtime};
874 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
875 $Jobstep->{'arvados_task'}->save;
877 splice @jobstep_todo, $todo_ptr, 1;
880 $progress_is_dirty = 1;
884 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
886 last THISROUND if $main::please_freeze || defined($main::success);
887 if ($main::please_info)
889 $main::please_info = 0;
891 create_output_collection();
893 update_progress_stats();
900 check_refresh_wanted();
902 update_progress_stats();
903 select (undef, undef, undef, 0.1);
905 elsif (time - $progress_stats_updated >= 30)
907 update_progress_stats();
909 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
910 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
912 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
913 .($thisround_failed+$thisround_succeeded)
914 .") -- giving up on this round";
915 Log (undef, $message);
919 # move slots from freeslot to holdslot (or back to freeslot) if necessary
920 for (my $i=$#freeslot; $i>=0; $i--) {
921 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
922 push @holdslot, (splice @freeslot, $i, 1);
925 for (my $i=$#holdslot; $i>=0; $i--) {
926 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
927 push @freeslot, (splice @holdslot, $i, 1);
931 # give up if no nodes are succeeding
932 if (!grep { $_->{node}->{losing_streak} == 0 &&
933 $_->{node}->{hold_count} < 4 } @slot) {
934 my $message = "Every node has failed -- giving up on this round";
935 Log (undef, $message);
942 push @freeslot, splice @holdslot;
943 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
946 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
949 if ($main::please_continue) {
950 $main::please_continue = 0;
953 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
957 check_refresh_wanted();
959 update_progress_stats();
960 select (undef, undef, undef, 0.1);
961 killem (keys %proc) if $main::please_freeze;
965 update_progress_stats();
966 freeze_if_want_freeze();
969 if (!defined $main::success)
972 $thisround_succeeded == 0 &&
973 ($thisround_failed == 0 || $thisround_failed > 4))
975 my $message = "stop because $thisround_failed tasks failed and none succeeded";
976 Log (undef, $message);
985 goto ONELEVEL if !defined $main::success;
988 release_allocation();
990 my $collated_output = &create_output_collection();
992 if (!$collated_output) {
993 Log (undef, "Failed to write output collection");
996 Log(undef, "job output $collated_output");
997 $Job->update_attributes('output' => $collated_output);
1000 Log (undef, "finish");
1005 if ($collated_output && $main::success) {
1006 $final_state = 'Complete';
1008 $final_state = 'Failed';
1010 $Job->update_attributes('state' => $final_state);
1012 exit (($final_state eq 'Complete') ? 0 : 1);
1016 sub update_progress_stats
1018 $progress_stats_updated = time;
1019 return if !$progress_is_dirty;
1020 my ($todo, $done, $running) = (scalar @jobstep_todo,
1021 scalar @jobstep_done,
1022 scalar @slot - scalar @freeslot - scalar @holdslot);
1023 $Job->{'tasks_summary'} ||= {};
1024 $Job->{'tasks_summary'}->{'todo'} = $todo;
1025 $Job->{'tasks_summary'}->{'done'} = $done;
1026 $Job->{'tasks_summary'}->{'running'} = $running;
1027 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1028 Log (undef, "status: $done done, $running running, $todo todo");
1029 $progress_is_dirty = 0;
1036 my $pid = waitpid (-1, WNOHANG);
1037 return 0 if $pid <= 0;
1039 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1041 . $slot[$proc{$pid}->{slot}]->{cpu});
1042 my $jobstepid = $proc{$pid}->{jobstep};
1043 my $elapsed = time - $proc{$pid}->{time};
1044 my $Jobstep = $jobstep[$jobstepid];
1046 my $childstatus = $?;
1047 my $exitvalue = $childstatus >> 8;
1048 my $exitinfo = "exit ".exit_status_s($childstatus);
1049 $Jobstep->{'arvados_task'}->reload;
1050 my $task_success = $Jobstep->{'arvados_task'}->{success};
1052 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1054 if (!defined $task_success) {
1055 # task did not indicate one way or the other --> fail
1056 $Jobstep->{'arvados_task'}->{success} = 0;
1057 $Jobstep->{'arvados_task'}->save;
1064 $temporary_fail ||= $Jobstep->{node_fail};
1065 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1067 ++$thisround_failed;
1068 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1070 # Check for signs of a failed or misconfigured node
1071 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1072 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1073 # Don't count this against jobstep failure thresholds if this
1074 # node is already suspected faulty and srun exited quickly
1075 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1077 Log ($jobstepid, "blaming failure on suspect node " .
1078 $slot[$proc{$pid}->{slot}]->{node}->{name});
1079 $temporary_fail ||= 1;
1081 ban_node_by_slot($proc{$pid}->{slot});
1084 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1085 ++$Jobstep->{'failures'},
1086 $temporary_fail ? 'temporary ' : 'permanent',
1089 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1090 # Give up on this task, and the whole job
1093 # Put this task back on the todo queue
1094 push @jobstep_todo, $jobstepid;
1095 $Job->{'tasks_summary'}->{'failed'}++;
1099 ++$thisround_succeeded;
1100 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1101 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1102 push @jobstep_done, $jobstepid;
1103 Log ($jobstepid, "success in $elapsed seconds");
1105 $Jobstep->{exitcode} = $childstatus;
1106 $Jobstep->{finishtime} = time;
1107 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1108 $Jobstep->{'arvados_task'}->save;
1109 process_stderr ($jobstepid, $task_success);
1110 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1111 length($Jobstep->{'arvados_task'}->{output}),
1112 $Jobstep->{'arvados_task'}->{output}));
1114 close $reader{$jobstepid};
1115 delete $reader{$jobstepid};
1116 delete $slot[$proc{$pid}->{slot}]->{pid};
1117 push @freeslot, $proc{$pid}->{slot};
1120 if ($task_success) {
1122 my $newtask_list = [];
1123 my $newtask_results;
1125 $newtask_results = api_call(
1128 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1130 'order' => 'qsequence',
1131 'offset' => scalar(@$newtask_list),
1133 push(@$newtask_list, @{$newtask_results->{items}});
1134 } while (@{$newtask_results->{items}});
1135 foreach my $arvados_task (@$newtask_list) {
1137 'level' => $arvados_task->{'sequence'},
1139 'arvados_task' => $arvados_task
1141 push @jobstep, $jobstep;
1142 push @jobstep_todo, $#jobstep;
1146 $progress_is_dirty = 1;
1150 sub check_refresh_wanted
1152 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1153 if (@stat && $stat[9] > $latest_refresh) {
1154 $latest_refresh = scalar time;
1155 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1156 for my $attr ('cancelled_at',
1157 'cancelled_by_user_uuid',
1158 'cancelled_by_client_uuid',
1160 $Job->{$attr} = $Job2->{$attr};
1162 if ($Job->{'state'} ne "Running") {
1163 if ($Job->{'state'} eq "Cancelled") {
1164 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1166 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1169 $main::please_freeze = 1;
1176 # return if the kill list was checked <4 seconds ago
1177 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
1181 $squeue_kill_checked = time;
1183 # use killem() on procs whose killtime is reached
1186 if (exists $proc{$_}->{killtime}
1187 && $proc{$_}->{killtime} <= time)
1193 # return if the squeue was checked <60 seconds ago
1194 if (defined $squeue_checked && $squeue_checked > time - 60)
1198 $squeue_checked = time;
1202 # here is an opportunity to check for mysterious problems with local procs
1206 # get a list of steps still running
1207 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
1209 if ($squeue[-1] ne "ok")
1215 # which of my jobsteps are running, according to squeue?
1219 if (/^(\d+)\.(\d+) (\S+)/)
1221 if ($1 eq $ENV{SLURM_JOBID})
1228 # which of my active child procs (>60s old) were not mentioned by squeue?
1229 foreach (keys %proc)
1231 if ($proc{$_}->{time} < time - 60
1232 && !exists $ok{$proc{$_}->{jobstepname}}
1233 && !exists $proc{$_}->{killtime})
1235 # kill this proc if it hasn't exited in 30 seconds
1236 $proc{$_}->{killtime} = time + 30;
1242 sub release_allocation
1246 Log (undef, "release job allocation");
1247 system "scancel $ENV{SLURM_JOBID}";
1255 foreach my $job (keys %reader)
1258 while (0 < sysread ($reader{$job}, $buf, 8192))
1260 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1261 $jobstep[$job]->{stderr} .= $buf;
1262 preprocess_stderr ($job);
1263 if (length ($jobstep[$job]->{stderr}) > 16384)
1265 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1274 sub preprocess_stderr
1278 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1280 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1281 Log ($job, "stderr $line");
1282 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1284 $main::please_freeze = 1;
1286 elsif ($line =~ /(srun: error: (Node failure on|Unable to create job step|.*: Communication connection failure))|arvados.errors.Keep/) {
1287 $jobstep[$job]->{node_fail} = 1;
1288 ban_node_by_slot($jobstep[$job]->{slotindex});
1297 my $task_success = shift;
1298 preprocess_stderr ($job);
1301 Log ($job, "stderr $_");
1302 } split ("\n", $jobstep[$job]->{stderr});
1309 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1310 Log(undef, "fetch_block run error from arv-get $hash: $!");
1313 my $output_block = "";
1316 my $bytes = sysread($keep, $buf, 1024 * 1024);
1317 if (!defined $bytes) {
1318 Log(undef, "fetch_block read error from arv-get: $!");
1319 $output_block = undef;
1321 } elsif ($bytes == 0) {
1322 # sysread returns 0 at the end of the pipe.
1325 # some bytes were read into buf.
1326 $output_block .= $buf;
1331 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1332 $output_block = undef;
1334 return $output_block;
1337 # Create a collection by concatenating the output of all tasks (each
1338 # task's output is either a manifest fragment, a locator for a
1339 # manifest fragment stored in Keep, or nothing at all). Return the
1340 # portable_data_hash of the new collection.
1341 sub create_output_collection
1343 Log (undef, "collate");
1345 my ($child_out, $child_in);
1346 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1349 print (arvados.api("v1").collections().
1350 create(body={"manifest_text": sys.stdin.read()}).
1351 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1355 my $manifest_size = 0;
1359 my $output = $_->{'arvados_task'}->{output};
1360 next if (!defined($output));
1362 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1363 $next_write = fetch_block($output);
1365 $next_write = $output;
1367 if (defined($next_write)) {
1368 if (!defined(syswrite($child_in, $next_write))) {
1369 # There's been an error writing. Stop the loop.
1370 # We'll log details about the exit code later.
1373 $manifest_size += length($next_write);
1376 my $uuid = $_->{'arvados_task'}->{'uuid'};
1377 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1382 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1385 my $s = IO::Select->new($child_out);
1386 if ($s->can_read(120)) {
1387 sysread($child_out, $joboutput, 1024 * 1024);
1390 Log(undef, "output collection creation exited " . exit_status_s($?));
1396 Log (undef, "timed out while creating output collection");
1397 foreach my $signal (2, 2, 2, 15, 15, 9) {
1398 kill($signal, $pid);
1399 last if waitpid($pid, WNOHANG) == -1;
1413 my $sig = 2; # SIGINT first
1414 if (exists $proc{$_}->{"sent_$sig"} &&
1415 time - $proc{$_}->{"sent_$sig"} > 4)
1417 $sig = 15; # SIGTERM if SIGINT doesn't work
1419 if (exists $proc{$_}->{"sent_$sig"} &&
1420 time - $proc{$_}->{"sent_$sig"} > 4)
1422 $sig = 9; # SIGKILL if SIGTERM doesn't work
1424 if (!exists $proc{$_}->{"sent_$sig"})
1426 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1428 select (undef, undef, undef, 0.1);
1431 kill $sig, $_; # srun wants two SIGINT to really interrupt
1433 $proc{$_}->{"sent_$sig"} = time;
1434 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1444 vec($bits,fileno($_),1) = 1;
1450 # Send log output to Keep via arv-put.
1452 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1453 # $log_pipe_pid is the pid of the arv-put subprocess.
1455 # The only functions that should access these variables directly are:
1457 # log_writer_start($logfilename)
1458 # Starts an arv-put pipe, reading data on stdin and writing it to
1459 # a $logfilename file in an output collection.
1461 # log_writer_send($txt)
1462 # Writes $txt to the output log collection.
1464 # log_writer_finish()
1465 # Closes the arv-put pipe and returns the output that it produces.
1467 # log_writer_is_active()
1468 # Returns a true value if there is currently a live arv-put
1469 # process, false otherwise.
1471 my ($log_pipe_in, $log_pipe_out, $log_pipe_pid);
1473 sub log_writer_start($)
1475 my $logfilename = shift;
1476 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1478 '--portable-data-hash',
1479 '--project-uuid', $Job->{owner_uuid},
1481 '--name', $logfilename,
1482 '--filename', $logfilename,
1486 sub log_writer_send($)
1489 print $log_pipe_in $txt;
1492 sub log_writer_finish()
1494 return unless $log_pipe_pid;
1496 close($log_pipe_in);
1499 my $s = IO::Select->new($log_pipe_out);
1500 if ($s->can_read(120)) {
1501 sysread($log_pipe_out, $arv_put_output, 1024);
1502 chomp($arv_put_output);
1504 Log (undef, "timed out reading from 'arv-put'");
1507 waitpid($log_pipe_pid, 0);
1508 $log_pipe_pid = $log_pipe_in = $log_pipe_out = undef;
1510 Log("log_writer_finish: arv-put exited ".exit_status_s($?))
1513 return $arv_put_output;
1516 sub log_writer_is_active() {
1517 return $log_pipe_pid;
1520 sub Log # ($jobstep_id, $logmessage)
1522 if ($_[1] =~ /\n/) {
1523 for my $line (split (/\n/, $_[1])) {
1528 my $fh = select STDERR; $|=1; select $fh;
1529 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1530 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1533 if (log_writer_is_active() || -t STDERR) {
1534 my @gmtime = gmtime;
1535 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1536 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1538 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1540 if (log_writer_is_active()) {
1541 log_writer_send($datetime . " " . $message);
1548 my ($package, $file, $line) = caller;
1549 my $message = "@_ at $file line $line\n";
1550 Log (undef, $message);
1551 freeze() if @jobstep_todo;
1552 create_output_collection() if @jobstep_todo;
1562 if ($Job->{'state'} eq 'Cancelled') {
1563 $Job->update_attributes('finished_at' => scalar gmtime);
1565 $Job->update_attributes('state' => 'Failed');
1572 my $justcheckpoint = shift; # false if this will be the last meta saved
1573 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1574 return unless log_writer_is_active();
1576 my $loglocator = log_writer_finish();
1577 Log (undef, "log manifest is $loglocator");
1578 $Job->{'log'} = $loglocator;
1579 $Job->update_attributes('log', $loglocator);
1583 sub freeze_if_want_freeze
1585 if ($main::please_freeze)
1587 release_allocation();
1590 # kill some srun procs before freeze+stop
1591 map { $proc{$_} = {} } @_;
1594 killem (keys %proc);
1595 select (undef, undef, undef, 0.1);
1597 while (($died = waitpid (-1, WNOHANG)) > 0)
1599 delete $proc{$died};
1604 create_output_collection();
1614 Log (undef, "Freeze not implemented");
1621 croak ("Thaw not implemented");
1637 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1644 my $srunargs = shift;
1645 my $execargs = shift;
1646 my $opts = shift || {};
1648 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1650 $Data::Dumper::Terse = 1;
1651 $Data::Dumper::Indent = 0;
1652 my $show_cmd = Dumper($args);
1653 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1654 $show_cmd =~ s/\n/ /g;
1655 warn "starting: $show_cmd\n";
1657 if (defined $stdin) {
1658 my $child = open STDIN, "-|";
1659 defined $child or die "no fork: $!";
1661 print $stdin or die $!;
1662 close STDOUT or die $!;
1667 return system (@$args) if $opts->{fork};
1670 warn "ENV size is ".length(join(" ",%ENV));
1671 die "exec failed: $!: @$args";
1675 sub ban_node_by_slot {
1676 # Don't start any new jobsteps on this node for 60 seconds
1678 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1679 $slot[$slotid]->{node}->{hold_count}++;
1680 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1685 my ($lockfile, $error_message) = @_;
1686 open L, ">", $lockfile or croak("$lockfile: $!");
1687 if (!flock L, LOCK_EX|LOCK_NB) {
1688 croak("Can't lock $lockfile: $error_message\n");
1692 sub find_docker_image {
1693 # Given a Keep locator, check to see if it contains a Docker image.
1694 # If so, return its stream name and Docker hash.
1695 # If not, return undef for both values.
1696 my $locator = shift;
1697 my ($streamname, $filename);
1698 my $image = api_call("collections/get", uuid => $locator);
1700 foreach my $line (split(/\n/, $image->{manifest_text})) {
1701 my @tokens = split(/\s+/, $line);
1703 $streamname = shift(@tokens);
1704 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1705 if (defined($filename)) {
1706 return (undef, undef); # More than one file in the Collection.
1708 $filename = (split(/:/, $filedata, 3))[2];
1713 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1714 return ($streamname, $1);
1716 return (undef, undef);
1721 # Calculate the number of times an operation should be retried,
1722 # assuming exponential backoff, and that we're willing to retry as
1723 # long as tasks have been running. Enforce a minimum of 3 retries.
1724 my ($starttime, $endtime, $timediff, $retries);
1726 $starttime = $jobstep[0]->{starttime};
1727 $endtime = $jobstep[-1]->{finishtime};
1729 if (!defined($starttime)) {
1731 } elsif (!defined($endtime)) {
1732 $timediff = time - $starttime;
1734 $timediff = ($endtime - $starttime) - (time - $endtime);
1736 if ($timediff > 0) {
1737 $retries = int(log($timediff) / log(2));
1739 $retries = 1; # Use the minimum.
1741 return ($retries > 3) ? $retries : 3;
1745 # Pass in two function references.
1746 # This method will be called with the remaining arguments.
1747 # If it dies, retry it with exponential backoff until it succeeds,
1748 # or until the current retry_count is exhausted. After each failure
1749 # that can be retried, the second function will be called with
1750 # the current try count (0-based), next try time, and error message.
1751 my $operation = shift;
1752 my $retry_callback = shift;
1753 my $retries = retry_count();
1754 foreach my $try_count (0..$retries) {
1755 my $next_try = time + (2 ** $try_count);
1756 my $result = eval { $operation->(@_); };
1759 } elsif ($try_count < $retries) {
1760 $retry_callback->($try_count, $next_try, $@);
1761 my $sleep_time = $next_try - time;
1762 sleep($sleep_time) if ($sleep_time > 0);
1765 # Ensure the error message ends in a newline, so Perl doesn't add
1766 # retry_op's line number to it.
1772 # Pass in a /-separated API method name, and arguments for it.
1773 # This function will call that method, retrying as needed until
1774 # the current retry_count is exhausted, with a log on the first failure.
1775 my $method_name = shift;
1776 my $log_api_retry = sub {
1777 my ($try_count, $next_try_at, $errmsg) = @_;
1778 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1779 $errmsg =~ s/\s/ /g;
1780 $errmsg =~ s/\s+$//;
1782 if ($next_try_at < time) {
1783 $retry_msg = "Retrying.";
1785 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1786 $retry_msg = "Retrying at $next_try_fmt.";
1788 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1791 foreach my $key (split(/\//, $method_name)) {
1792 $method = $method->{$key};
1794 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1798 # Given a $?, return a human-readable exit code string like "0" or
1799 # "1" or "0 with signal 1" or "1 with signal 11".
1800 my $exitcode = shift;
1801 my $s = $exitcode >> 8;
1802 if ($exitcode & 0x7f) {
1803 $s .= " with signal " . ($exitcode & 0x7f);
1805 if ($exitcode & 0x80) {
1806 $s .= " with core dump";
1811 sub handle_readall {
1812 # Pass in a glob reference to a file handle.
1813 # Read all its contents and return them as a string.
1814 my $fh_glob_ref = shift;
1816 return <$fh_glob_ref>;
1819 sub tar_filename_n {
1821 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1824 sub add_git_archive {
1825 # Pass in a git archive command as a string or list, a la system().
1826 # This method will save its output to be included in the archive sent to the
1830 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1831 croak("Failed to save git archive: $!");
1833 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1835 waitpid($git_pid, 0);
1838 croak("Failed to save git archive: git exited " . exit_status_s($?));
1842 sub combined_git_archive {
1843 # Combine all saved tar archives into a single archive, then return its
1844 # contents in a string. Return undef if no archives have been saved.
1845 if ($git_tar_count < 1) {
1848 my $base_tar_name = tar_filename_n(1);
1849 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
1850 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
1851 if ($tar_exit != 0) {
1852 croak("Error preparing build archive: tar -A exited " .
1853 exit_status_s($tar_exit));
1856 if (!open(GIT_TAR, "<", $base_tar_name)) {
1857 croak("Could not open build archive: $!");
1859 my $tar_contents = handle_readall(\*GIT_TAR);
1861 return $tar_contents;
1867 # This is crunch-job's internal dispatch script. crunch-job running on the API
1868 # server invokes this script on individual compute nodes, or localhost if we're
1869 # running a job locally. It gets called in two modes:
1871 # * No arguments: Installation mode. Read a tar archive from the DATA
1872 # file handle; it includes the Crunch script's source code, and
1873 # maybe SDKs as well. Those should be installed in the proper
1874 # locations. This runs outside of any Docker container, so don't try to
1875 # introspect Crunch's runtime environment.
1877 # * With arguments: Crunch script run mode. This script should set up the
1878 # environment, then run the command specified in the arguments. This runs
1879 # inside any Docker container.
1882 use File::Path qw( make_path remove_tree );
1883 use POSIX qw(getcwd);
1885 use constant TASK_TEMPFAIL => 111;
1887 # Map SDK subdirectories to the path environments they belong to.
1888 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
1890 my $destdir = $ENV{"CRUNCH_SRC"};
1891 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1892 my $repo = $ENV{"CRUNCH_SRC_URL"};
1893 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
1894 my $job_work = $ENV{"JOB_WORK"};
1895 my $task_work = $ENV{"TASK_WORK"};
1897 for my $dir ($destdir, $job_work, $task_work) {
1900 -e $dir or die "Failed to create temporary directory ($dir): $!";
1905 remove_tree($task_work, {keep_root => 1});
1908 open(STDOUT_ORIG, ">&", STDOUT);
1909 open(STDERR_ORIG, ">&", STDERR);
1910 open(STDOUT, ">>", "$destdir.log");
1911 open(STDERR, ">&", STDOUT);
1913 ### Crunch script run mode
1915 # We want to do routine logging during task 0 only. This gives the user
1916 # the information they need, but avoids repeating the information for every
1919 if ($ENV{TASK_SEQUENCE} eq "0") {
1922 printf STDERR_ORIG "[Crunch] $msg\n", @_;
1928 my $python_src = "$install_dir/python";
1929 my $venv_dir = "$job_work/.arvados.venv";
1930 my $venv_built = -e "$venv_dir/bin/activate";
1931 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
1932 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
1933 "--python=python2.7", $venv_dir);
1934 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
1936 $Log->("Built Python SDK virtualenv");
1939 my $pip_bin = "pip";
1941 $Log->("Running in Python SDK virtualenv");
1942 $pip_bin = "$venv_dir/bin/pip";
1943 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
1944 @ARGV = ("/bin/sh", "-ec",
1945 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
1946 } elsif (-d $python_src) {
1947 $Log->("Warning: virtualenv not found inside Docker container default " .
1948 "\$PATH. Can't install Python SDK.");
1951 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
1953 $Log->("Using Arvados SDK:");
1954 foreach my $line (split /\n/, $pkgs) {
1958 $Log->("Arvados SDK packages not found");
1961 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
1962 my $sdk_path = "$install_dir/$sdk_dir";
1964 if ($ENV{$sdk_envkey}) {
1965 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
1967 $ENV{$sdk_envkey} = $sdk_path;
1969 $Log->("Arvados SDK added to %s", $sdk_envkey);
1975 open(STDOUT, ">&", STDOUT_ORIG);
1976 open(STDERR, ">&", STDERR_ORIG);
1978 die "Cannot exec `@ARGV`: $!";
1981 ### Installation mode
1982 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1984 if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
1985 # This version already installed -> nothing to do.
1989 unlink "$destdir.commit";
1992 if (!open(TARX, "|-", "tar", "-xC", $destdir)) {
1993 die "Error launching 'tar -xC $destdir': $!";
1995 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
1996 # get SIGPIPE. We must feed it data incrementally.
1998 while (read(DATA, $tar_input, 65536)) {
1999 print TARX $tar_input;
2002 die "'tar -xC $destdir' exited $?: $!";
2007 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2009 foreach my $sdk_lang (("python",
2010 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2011 if (-d "$sdk_root/$sdk_lang") {
2012 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2013 die "Failed to install $sdk_lang SDK: $!";
2019 my $python_dir = "$install_dir/python";
2020 if ((-d $python_dir) and can_run("python2.7") and
2021 (system("python2.7", "$python_dir/setup.py", "--quiet", "egg_info") != 0)) {
2022 # egg_info failed, probably when it asked git for a build tag.
2023 # Specify no build tag.
2024 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2025 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2029 if (-e "$destdir/crunch_scripts/install") {
2030 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2031 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2033 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2034 } elsif (-e "./install.sh") {
2035 shell_or_die (undef, "./install.sh", $install_dir);
2039 unlink "$destdir.commit.new";
2040 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
2041 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
2047 my $command_name = shift;
2048 open(my $which, "-|", "which", $command_name);
2049 while (<$which>) { }
2056 my $exitcode = shift;
2058 if ($ENV{"DEBUG"}) {
2059 print STDERR "@_\n";
2061 if (system (@_) != 0) {
2064 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2065 open STDERR, ">&STDERR_ORIG";
2066 system ("cat $destdir.log >&2");
2067 warn "@_ failed ($err): $exitstatus";
2068 if (defined($exitcode)) {
2072 exit (($code >> 8) || 1);