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 EX_TEMPFAIL => 75;
101 $ENV{"TMPDIR"} ||= "/tmp";
102 unless (defined $ENV{"CRUNCH_TMP"}) {
103 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
104 if ($ENV{"USER"} ne "crunch" && $< != 0) {
105 # use a tmp dir unique for my uid
106 $ENV{"CRUNCH_TMP"} .= "-$<";
110 # Create the tmp directory if it does not exist
111 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
112 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
115 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
116 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
117 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
118 mkdir ($ENV{"JOB_WORK"});
126 GetOptions('force-unlock' => \$force_unlock,
127 'git-dir=s' => \$git_dir,
128 'job=s' => \$jobspec,
129 'job-api-token=s' => \$job_api_token,
130 'no-clear-tmp' => \$no_clear_tmp,
131 'resume-stash=s' => \$resume_stash,
134 if (defined $job_api_token) {
135 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
138 my $have_slurm = exists $ENV{SLURM_JOBID} && exists $ENV{SLURM_NODELIST};
144 $main::ENV{CRUNCH_DEBUG} = 1;
148 $main::ENV{CRUNCH_DEBUG} = 0;
153 my $arv = Arvados->new('apiVersion' => 'v1');
161 my $User = api_call("users/current");
163 if ($jobspec =~ /^[-a-z\d]+$/)
165 # $jobspec is an Arvados UUID, not a JSON job specification
166 $Job = api_call("jobs/get", uuid => $jobspec);
167 if (!$force_unlock) {
168 # Claim this job, and make sure nobody else does
169 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
171 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
178 $Job = JSON::decode_json($jobspec);
182 map { croak ("No $_ specified") unless $Job->{$_} }
183 qw(script script_version script_parameters);
186 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
187 $Job->{'started_at'} = gmtime;
188 $Job->{'state'} = 'Running';
190 $Job = api_call("jobs/create", job => $Job);
192 $job_id = $Job->{'uuid'};
194 my $keep_logfile = $job_id . '.log.txt';
195 log_writer_start($keep_logfile);
197 $Job->{'runtime_constraints'} ||= {};
198 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
199 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
201 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
203 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
204 chomp($gem_versions);
205 chop($gem_versions); # Closing parentheses
210 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
212 Log (undef, "check slurm allocation");
215 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
219 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
220 push @sinfo, "$localcpus localhost";
222 if (exists $ENV{SLURM_NODELIST})
224 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
228 my ($ncpus, $slurm_nodelist) = split;
229 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
232 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
235 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
238 foreach (split (",", $ranges))
251 push @nodelist, map {
253 $n =~ s/\[[-,\d]+\]/$_/;
260 push @nodelist, $nodelist;
263 foreach my $nodename (@nodelist)
265 Log (undef, "node $nodename - $ncpus slots");
266 my $node = { name => $nodename,
270 foreach my $cpu (1..$ncpus)
272 push @slot, { node => $node,
276 push @node, @nodelist;
281 # Ensure that we get one jobstep running on each allocated node before
282 # we start overloading nodes with concurrent steps
284 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
287 $Job->update_attributes(
288 'tasks_summary' => { 'failed' => 0,
293 Log (undef, "start");
294 $SIG{'INT'} = sub { $main::please_freeze = 1; };
295 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
296 $SIG{'TERM'} = \&croak;
297 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
298 $SIG{'ALRM'} = sub { $main::please_info = 1; };
299 $SIG{'CONT'} = sub { $main::please_continue = 1; };
300 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
302 $main::please_freeze = 0;
303 $main::please_info = 0;
304 $main::please_continue = 0;
305 $main::please_refresh = 0;
306 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
308 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
309 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
310 $ENV{"JOB_UUID"} = $job_id;
313 my @jobstep_todo = ();
314 my @jobstep_done = ();
315 my @jobstep_tomerge = ();
316 my $jobstep_tomerge_level = 0;
318 my $squeue_kill_checked;
319 my $latest_refresh = scalar time;
323 if (defined $Job->{thawedfromkey})
325 thaw ($Job->{thawedfromkey});
329 my $first_task = api_call("job_tasks/create", job_task => {
330 'job_uuid' => $Job->{'uuid'},
335 push @jobstep, { 'level' => 0,
337 'arvados_task' => $first_task,
339 push @jobstep_todo, 0;
345 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
348 my $build_script = handle_readall(\*DATA);
349 my $nodelist = join(",", @node);
350 my $git_tar_count = 0;
352 if (!defined $no_clear_tmp) {
353 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
354 Log (undef, "Clean work dirs");
356 my $cleanpid = fork();
359 # Find FUSE mounts that look like Keep mounts (the mount path has the
360 # word "keep") and unmount them. Then clean up work directories.
361 # TODO: When #5036 is done and widely deployed, we can get rid of the
362 # regular expression and just unmount everything with type fuse.keep.
363 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
364 ['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*']);
369 last if $cleanpid == waitpid (-1, WNOHANG);
370 freeze_if_want_freeze ($cleanpid);
371 select (undef, undef, undef, 0.1);
373 Log (undef, "Cleanup command exited ".exit_status_s($?));
376 # If this job requires a Docker image, install that.
377 my $docker_bin = "/usr/bin/docker.io";
378 my ($docker_locator, $docker_stream, $docker_hash);
379 if ($docker_locator = $Job->{docker_image_locator}) {
380 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
383 croak("No Docker image hash found from locator $docker_locator");
385 $docker_stream =~ s/^\.//;
386 my $docker_install_script = qq{
387 if ! $docker_bin images -q --no-trunc | grep -qxF \Q$docker_hash\E; then
388 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
391 my $docker_pid = fork();
392 if ($docker_pid == 0)
394 srun (["srun", "--nodelist=" . join(',', @node)],
395 ["/bin/sh", "-ec", $docker_install_script]);
400 last if $docker_pid == waitpid (-1, WNOHANG);
401 freeze_if_want_freeze ($docker_pid);
402 select (undef, undef, undef, 0.1);
406 croak("Installing Docker image from $docker_locator exited "
410 if ($Job->{arvados_sdk_version}) {
411 # The job also specifies an Arvados SDK version. Add the SDKs to the
412 # tar file for the build script to install.
413 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
414 $Job->{arvados_sdk_version}));
415 add_git_archive("git", "--git-dir=$git_dir", "archive",
416 "--prefix=.arvados.sdk/",
417 $Job->{arvados_sdk_version}, "sdk");
421 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
422 # If script_version looks like an absolute path, *and* the --git-dir
423 # argument was not given -- which implies we were not invoked by
424 # crunch-dispatch -- we will use the given path as a working
425 # directory instead of resolving script_version to a git commit (or
426 # doing anything else with git).
427 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
428 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
431 # Resolve the given script_version to a git commit sha1. Also, if
432 # the repository is remote, clone it into our local filesystem: this
433 # ensures "git archive" will work, and is necessary to reliably
434 # resolve a symbolic script_version like "master^".
435 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
437 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
439 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
441 # If we're running under crunch-dispatch, it will have already
442 # pulled the appropriate source tree into its own repository, and
443 # given us that repo's path as $git_dir.
445 # If we're running a "local" job, we might have to fetch content
446 # from a remote repository.
448 # (Currently crunch-dispatch gives a local path with --git-dir, but
449 # we might as well accept URLs there too in case it changes its
451 my $repo = $git_dir || $Job->{'repository'};
453 # Repository can be remote or local. If remote, we'll need to fetch it
454 # to a local dir before doing `git log` et al.
457 if ($repo =~ m{://|^[^/]*:}) {
458 # $repo is a git url we can clone, like git:// or https:// or
459 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
460 # not recognized here because distinguishing that from a local
461 # path is too fragile. If you really need something strange here,
462 # use the ssh:// form.
463 $repo_location = 'remote';
464 } elsif ($repo =~ m{^\.*/}) {
465 # $repo is a local path to a git index. We'll also resolve ../foo
466 # to ../foo/.git if the latter is a directory. To help
467 # disambiguate local paths from named hosted repositories, this
468 # form must be given as ./ or ../ if it's a relative path.
469 if (-d "$repo/.git") {
470 $repo = "$repo/.git";
472 $repo_location = 'local';
474 # $repo is none of the above. It must be the name of a hosted
476 my $arv_repo_list = api_call("repositories/list",
477 'filters' => [['name','=',$repo]]);
478 my @repos_found = @{$arv_repo_list->{'items'}};
479 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
481 Log(undef, "Repository '$repo' -> "
482 . join(", ", map { $_->{'uuid'} } @repos_found));
485 croak("Error: Found $n_found repositories with name '$repo'.");
487 $repo = $repos_found[0]->{'fetch_url'};
488 $repo_location = 'remote';
490 Log(undef, "Using $repo_location repository '$repo'");
491 $ENV{"CRUNCH_SRC_URL"} = $repo;
493 # Resolve given script_version (we'll call that $treeish here) to a
494 # commit sha1 ($commit).
495 my $treeish = $Job->{'script_version'};
497 if ($repo_location eq 'remote') {
498 # We minimize excess object-fetching by re-using the same bare
499 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
500 # just keep adding remotes to it as needed.
501 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
502 my $gitcmd = "git --git-dir=\Q$local_repo\E";
504 # Set up our local repo for caching remote objects, making
506 if (!-d $local_repo) {
507 make_path($local_repo) or croak("Error: could not create $local_repo");
509 # This works (exits 0 and doesn't delete fetched objects) even
510 # if $local_repo is already initialized:
511 `$gitcmd init --bare`;
513 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
516 # If $treeish looks like a hash (or abbrev hash) we look it up in
517 # our local cache first, since that's cheaper. (We don't want to
518 # do that with tags/branches though -- those change over time, so
519 # they should always be resolved by the remote repo.)
520 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
521 # Hide stderr because it's normal for this to fail:
522 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
524 # Careful not to resolve a branch named abcdeff to commit 1234567:
525 $sha1 =~ /^$treeish/ &&
526 $sha1 =~ /^([0-9a-f]{40})$/s) {
528 Log(undef, "Commit $commit already present in $local_repo");
532 if (!defined $commit) {
533 # If $treeish isn't just a hash or abbrev hash, or isn't here
534 # yet, we need to fetch the remote to resolve it correctly.
536 # First, remove all local heads. This prevents a name that does
537 # not exist on the remote from resolving to (or colliding with)
538 # a previously fetched branch or tag (possibly from a different
540 remove_tree("$local_repo/refs/heads", {keep_root => 1});
542 Log(undef, "Fetching objects from $repo to $local_repo");
543 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
545 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
549 # Now that the data is all here, we will use our local repo for
550 # the rest of our git activities.
554 my $gitcmd = "git --git-dir=\Q$repo\E";
555 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
556 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
557 croak("`$gitcmd rev-list` exited "
559 .", '$treeish' not found. Giving up.");
562 Log(undef, "Version $treeish is commit $commit");
564 if ($commit ne $Job->{'script_version'}) {
565 # Record the real commit id in the database, frozentokey, logs,
566 # etc. -- instead of an abbreviation or a branch name which can
567 # become ambiguous or point to a different commit in the future.
568 if (!$Job->update_attributes('script_version' => $commit)) {
569 croak("Error: failed to update job's script_version attribute");
573 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
574 add_git_archive("$gitcmd archive ''\Q$commit\E");
577 my $git_archive = combined_git_archive();
578 if (!defined $git_archive) {
579 Log(undef, "Skip install phase (no git archive)");
581 Log(undef, "Warning: This probably means workers have no source tree!");
585 Log(undef, "Run install script on all workers");
587 my @srunargs = ("srun",
588 "--nodelist=$nodelist",
589 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
590 my @execargs = ("sh", "-c",
591 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
593 my $installpid = fork();
594 if ($installpid == 0)
596 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
601 last if $installpid == waitpid (-1, WNOHANG);
602 freeze_if_want_freeze ($installpid);
603 select (undef, undef, undef, 0.1);
605 my $install_exited = $?;
606 Log (undef, "Install script exited ".exit_status_s($install_exited));
607 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
608 unlink($tar_filename);
610 exit (1) if $install_exited != 0;
613 foreach (qw (script script_version script_parameters runtime_constraints))
617 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
619 foreach (split (/\n/, $Job->{knobs}))
621 Log (undef, "knob " . $_);
626 $main::success = undef;
632 my $thisround_succeeded = 0;
633 my $thisround_failed = 0;
634 my $thisround_failed_multiple = 0;
636 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
637 or $a <=> $b } @jobstep_todo;
638 my $level = $jobstep[$jobstep_todo[0]]->{level};
639 Log (undef, "start level $level");
644 my @freeslot = (0..$#slot);
647 my $progress_is_dirty = 1;
648 my $progress_stats_updated = 0;
650 update_progress_stats();
655 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
657 my $id = $jobstep_todo[$todo_ptr];
658 my $Jobstep = $jobstep[$id];
659 if ($Jobstep->{level} != $level)
664 pipe $reader{$id}, "writer" or croak ($!);
665 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
666 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
668 my $childslot = $freeslot[0];
669 my $childnode = $slot[$childslot]->{node};
670 my $childslotname = join (".",
671 $slot[$childslot]->{node}->{name},
672 $slot[$childslot]->{cpu});
673 my $childpid = fork();
676 $SIG{'INT'} = 'DEFAULT';
677 $SIG{'QUIT'} = 'DEFAULT';
678 $SIG{'TERM'} = 'DEFAULT';
680 foreach (values (%reader))
684 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
685 open(STDOUT,">&writer");
686 open(STDERR,">&writer");
691 delete $ENV{"GNUPGHOME"};
692 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
693 $ENV{"TASK_QSEQUENCE"} = $id;
694 $ENV{"TASK_SEQUENCE"} = $level;
695 $ENV{"JOB_SCRIPT"} = $Job->{script};
696 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
697 $param =~ tr/a-z/A-Z/;
698 $ENV{"JOB_PARAMETER_$param"} = $value;
700 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
701 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
702 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
703 $ENV{"HOME"} = $ENV{"TASK_WORK"};
704 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
705 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
706 $ENV{"CRUNCH_NODE_SLOTS"} = $slot[$childslot]->{node}->{ncpus};
707 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
713 "--nodelist=".$childnode->{name},
714 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
715 "--job-name=$job_id.$id.$$",
718 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
719 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
720 ."&& cd $ENV{CRUNCH_TMP} ";
721 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
724 my $cidfile = "$ENV{CRUNCH_TMP}/$ENV{TASK_UUID}.cid";
725 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
726 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
728 # Dynamically configure the container to use the host system as its
729 # DNS server. Get the host's global addresses from the ip command,
730 # and turn them into docker --dns options using gawk.
732 q{$(ip -o address show scope global |
733 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
735 # The source tree and $destdir directory (which we have
736 # installed on the worker host) are available in the container,
737 # under the same path.
738 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
739 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
741 # Currently, we make arv-mount's mount point appear at /keep
742 # inside the container (instead of using the same path as the
743 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
744 # crunch scripts and utilities must not rely on this. They must
745 # use $TASK_KEEPMOUNT.
746 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
747 $ENV{TASK_KEEPMOUNT} = "/keep";
749 # TASK_WORK is almost exactly like a docker data volume: it
750 # starts out empty, is writable, and persists until no
751 # containers use it any more. We don't use --volumes-from to
752 # share it with other containers: it is only accessible to this
753 # task, and it goes away when this task stops.
755 # However, a docker data volume is writable only by root unless
756 # the mount point already happens to exist in the container with
757 # different permissions. Therefore, we [1] assume /tmp already
758 # exists in the image and is writable by the crunch user; [2]
759 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
760 # writable if they are created by docker while setting up the
761 # other --volumes); and [3] create $TASK_WORK inside the
762 # container using $build_script.
763 $command .= "--volume=/tmp ";
764 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
765 $ENV{"HOME"} = $ENV{"TASK_WORK"};
766 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
768 # TODO: Share a single JOB_WORK volume across all task
769 # containers on a given worker node, and delete it when the job
770 # ends (and, in case that doesn't work, when the next job
773 # For now, use the same approach as TASK_WORK above.
774 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
776 while (my ($env_key, $env_val) = each %ENV)
778 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
779 $command .= "--env=\Q$env_key=$env_val\E ";
782 $command .= "--env=\QHOME=$ENV{HOME}\E ";
783 $command .= "\Q$docker_hash\E ";
784 $command .= "stdbuf --output=0 --error=0 ";
785 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
788 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
789 $command .= "stdbuf --output=0 --error=0 ";
790 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
793 my @execargs = ('bash', '-c', $command);
794 srun (\@srunargs, \@execargs, undef, $build_script);
795 # exec() failed, we assume nothing happened.
796 die "srun() failed on build script\n";
799 if (!defined $childpid)
806 $proc{$childpid} = { jobstep => $id,
809 jobstepname => "$job_id.$id.$childpid",
811 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
812 $slot[$childslot]->{pid} = $childpid;
814 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
815 Log ($id, "child $childpid started on $childslotname");
816 $Jobstep->{starttime} = time;
817 $Jobstep->{node} = $childnode->{name};
818 $Jobstep->{slotindex} = $childslot;
819 delete $Jobstep->{stderr};
820 delete $Jobstep->{finishtime};
822 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
823 $Jobstep->{'arvados_task'}->save;
825 splice @jobstep_todo, $todo_ptr, 1;
828 $progress_is_dirty = 1;
832 (@slot > @freeslot && $todo_ptr+1 > $#jobstep_todo))
834 last THISROUND if $main::please_freeze;
835 if ($main::please_info)
837 $main::please_info = 0;
839 create_output_collection();
841 update_progress_stats();
848 check_refresh_wanted();
850 update_progress_stats();
851 select (undef, undef, undef, 0.1);
853 elsif (time - $progress_stats_updated >= 30)
855 update_progress_stats();
857 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
858 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
860 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
861 .($thisround_failed+$thisround_succeeded)
862 .") -- giving up on this round";
863 Log (undef, $message);
867 # move slots from freeslot to holdslot (or back to freeslot) if necessary
868 for (my $i=$#freeslot; $i>=0; $i--) {
869 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
870 push @holdslot, (splice @freeslot, $i, 1);
873 for (my $i=$#holdslot; $i>=0; $i--) {
874 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
875 push @freeslot, (splice @holdslot, $i, 1);
879 # give up if no nodes are succeeding
880 if (!grep { $_->{node}->{losing_streak} == 0 &&
881 $_->{node}->{hold_count} < 4 } @slot) {
882 my $message = "Every node has failed -- giving up on this round";
883 Log (undef, $message);
890 push @freeslot, splice @holdslot;
891 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
894 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
897 if ($main::please_continue) {
898 $main::please_continue = 0;
901 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
905 check_refresh_wanted();
907 update_progress_stats();
908 select (undef, undef, undef, 0.1);
909 killem (keys %proc) if $main::please_freeze;
913 update_progress_stats();
914 freeze_if_want_freeze();
917 if (!defined $main::success)
920 $thisround_succeeded == 0 &&
921 ($thisround_failed == 0 || $thisround_failed > 4))
923 my $message = "stop because $thisround_failed tasks failed and none succeeded";
924 Log (undef, $message);
933 goto ONELEVEL if !defined $main::success;
936 release_allocation();
938 my $collated_output = &create_output_collection();
940 if (!$collated_output) {
941 Log (undef, "Failed to write output collection");
944 Log(undef, "output hash " . $collated_output);
945 $Job->update_attributes('output' => $collated_output);
948 Log (undef, "finish");
953 if ($collated_output && $main::success) {
954 $final_state = 'Complete';
956 $final_state = 'Failed';
958 $Job->update_attributes('state' => $final_state);
960 exit (($final_state eq 'Complete') ? 0 : 1);
964 sub update_progress_stats
966 $progress_stats_updated = time;
967 return if !$progress_is_dirty;
968 my ($todo, $done, $running) = (scalar @jobstep_todo,
969 scalar @jobstep_done,
970 scalar @slot - scalar @freeslot - scalar @holdslot);
971 $Job->{'tasks_summary'} ||= {};
972 $Job->{'tasks_summary'}->{'todo'} = $todo;
973 $Job->{'tasks_summary'}->{'done'} = $done;
974 $Job->{'tasks_summary'}->{'running'} = $running;
975 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
976 Log (undef, "status: $done done, $running running, $todo todo");
977 $progress_is_dirty = 0;
984 my $pid = waitpid (-1, WNOHANG);
985 return 0 if $pid <= 0;
987 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
989 . $slot[$proc{$pid}->{slot}]->{cpu});
990 my $jobstepid = $proc{$pid}->{jobstep};
991 my $elapsed = time - $proc{$pid}->{time};
992 my $Jobstep = $jobstep[$jobstepid];
994 my $childstatus = $?;
995 my $exitvalue = $childstatus >> 8;
996 my $exitinfo = "exit ".exit_status_s($childstatus);
997 $Jobstep->{'arvados_task'}->reload;
998 my $task_success = $Jobstep->{'arvados_task'}->{success};
1000 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1002 if (!defined $task_success) {
1003 # task did not indicate one way or the other --> fail
1004 $Jobstep->{'arvados_task'}->{success} = 0;
1005 $Jobstep->{'arvados_task'}->save;
1012 $temporary_fail ||= $Jobstep->{node_fail};
1013 $temporary_fail ||= ($exitvalue == 111);
1015 ++$thisround_failed;
1016 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1018 # Check for signs of a failed or misconfigured node
1019 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1020 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1021 # Don't count this against jobstep failure thresholds if this
1022 # node is already suspected faulty and srun exited quickly
1023 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1025 Log ($jobstepid, "blaming failure on suspect node " .
1026 $slot[$proc{$pid}->{slot}]->{node}->{name});
1027 $temporary_fail ||= 1;
1029 ban_node_by_slot($proc{$pid}->{slot});
1032 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1033 ++$Jobstep->{'failures'},
1034 $temporary_fail ? 'temporary ' : 'permanent',
1037 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1038 # Give up on this task, and the whole job
1041 # Put this task back on the todo queue
1042 push @jobstep_todo, $jobstepid;
1043 $Job->{'tasks_summary'}->{'failed'}++;
1047 ++$thisround_succeeded;
1048 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1049 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1050 push @jobstep_done, $jobstepid;
1051 Log ($jobstepid, "success in $elapsed seconds");
1053 $Jobstep->{exitcode} = $childstatus;
1054 $Jobstep->{finishtime} = time;
1055 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1056 $Jobstep->{'arvados_task'}->save;
1057 process_stderr ($jobstepid, $task_success);
1058 Log ($jobstepid, "output " . $Jobstep->{'arvados_task'}->{output});
1060 close $reader{$jobstepid};
1061 delete $reader{$jobstepid};
1062 delete $slot[$proc{$pid}->{slot}]->{pid};
1063 push @freeslot, $proc{$pid}->{slot};
1066 if ($task_success) {
1068 my $newtask_list = [];
1069 my $newtask_results;
1071 $newtask_results = api_call(
1074 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1076 'order' => 'qsequence',
1077 'offset' => scalar(@$newtask_list),
1079 push(@$newtask_list, @{$newtask_results->{items}});
1080 } while (@{$newtask_results->{items}});
1081 foreach my $arvados_task (@$newtask_list) {
1083 'level' => $arvados_task->{'sequence'},
1085 'arvados_task' => $arvados_task
1087 push @jobstep, $jobstep;
1088 push @jobstep_todo, $#jobstep;
1092 $progress_is_dirty = 1;
1096 sub check_refresh_wanted
1098 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1099 if (@stat && $stat[9] > $latest_refresh) {
1100 $latest_refresh = scalar time;
1101 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1102 for my $attr ('cancelled_at',
1103 'cancelled_by_user_uuid',
1104 'cancelled_by_client_uuid',
1106 $Job->{$attr} = $Job2->{$attr};
1108 if ($Job->{'state'} ne "Running") {
1109 if ($Job->{'state'} eq "Cancelled") {
1110 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1112 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1115 $main::please_freeze = 1;
1122 # return if the kill list was checked <4 seconds ago
1123 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
1127 $squeue_kill_checked = time;
1129 # use killem() on procs whose killtime is reached
1132 if (exists $proc{$_}->{killtime}
1133 && $proc{$_}->{killtime} <= time)
1139 # return if the squeue was checked <60 seconds ago
1140 if (defined $squeue_checked && $squeue_checked > time - 60)
1144 $squeue_checked = time;
1148 # here is an opportunity to check for mysterious problems with local procs
1152 # get a list of steps still running
1153 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
1155 if ($squeue[-1] ne "ok")
1161 # which of my jobsteps are running, according to squeue?
1165 if (/^(\d+)\.(\d+) (\S+)/)
1167 if ($1 eq $ENV{SLURM_JOBID})
1174 # which of my active child procs (>60s old) were not mentioned by squeue?
1175 foreach (keys %proc)
1177 if ($proc{$_}->{time} < time - 60
1178 && !exists $ok{$proc{$_}->{jobstepname}}
1179 && !exists $proc{$_}->{killtime})
1181 # kill this proc if it hasn't exited in 30 seconds
1182 $proc{$_}->{killtime} = time + 30;
1188 sub release_allocation
1192 Log (undef, "release job allocation");
1193 system "scancel $ENV{SLURM_JOBID}";
1201 foreach my $job (keys %reader)
1204 while (0 < sysread ($reader{$job}, $buf, 8192))
1206 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1207 $jobstep[$job]->{stderr} .= $buf;
1208 preprocess_stderr ($job);
1209 if (length ($jobstep[$job]->{stderr}) > 16384)
1211 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1220 sub preprocess_stderr
1224 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1226 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1227 Log ($job, "stderr $line");
1228 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1230 $main::please_freeze = 1;
1232 elsif ($line =~ /srun: error: (Node failure on|Unable to create job step) /) {
1233 $jobstep[$job]->{node_fail} = 1;
1234 ban_node_by_slot($jobstep[$job]->{slotindex});
1243 my $task_success = shift;
1244 preprocess_stderr ($job);
1247 Log ($job, "stderr $_");
1248 } split ("\n", $jobstep[$job]->{stderr});
1255 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1256 Log(undef, "fetch_block run error from arv-get $hash: $!");
1259 my $output_block = "";
1262 my $bytes = sysread($keep, $buf, 1024 * 1024);
1263 if (!defined $bytes) {
1264 Log(undef, "fetch_block read error from arv-get: $!");
1265 $output_block = undef;
1267 } elsif ($bytes == 0) {
1268 # sysread returns 0 at the end of the pipe.
1271 # some bytes were read into buf.
1272 $output_block .= $buf;
1277 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1278 $output_block = undef;
1280 return $output_block;
1283 # Create a collection by concatenating the output of all tasks (each
1284 # task's output is either a manifest fragment, a locator for a
1285 # manifest fragment stored in Keep, or nothing at all). Return the
1286 # portable_data_hash of the new collection.
1287 sub create_output_collection
1289 Log (undef, "collate");
1291 my ($child_out, $child_in);
1292 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1295 print (arvados.api("v1").collections().
1296 create(body={"manifest_text": sys.stdin.read()}).
1297 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1304 my $output = $_->{'arvados_task'}->{output};
1305 next if (!defined($output));
1307 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1308 $next_write = fetch_block($output);
1310 $next_write = $output;
1312 if (defined($next_write)) {
1313 if (!defined(syswrite($child_in, $next_write))) {
1314 # There's been an error writing. Stop the loop.
1315 # We'll log details about the exit code later.
1319 my $uuid = $_->{'arvados_task'}->{'uuid'};
1320 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1327 my $s = IO::Select->new($child_out);
1328 if ($s->can_read(120)) {
1329 sysread($child_out, $joboutput, 64 * 1024 * 1024);
1332 Log(undef, "output collection creation exited " . exit_status_s($?));
1338 Log (undef, "timed out while creating output collection");
1339 foreach my $signal (2, 2, 2, 15, 15, 9) {
1340 kill($signal, $pid);
1341 last if waitpid($pid, WNOHANG) == -1;
1355 my $sig = 2; # SIGINT first
1356 if (exists $proc{$_}->{"sent_$sig"} &&
1357 time - $proc{$_}->{"sent_$sig"} > 4)
1359 $sig = 15; # SIGTERM if SIGINT doesn't work
1361 if (exists $proc{$_}->{"sent_$sig"} &&
1362 time - $proc{$_}->{"sent_$sig"} > 4)
1364 $sig = 9; # SIGKILL if SIGTERM doesn't work
1366 if (!exists $proc{$_}->{"sent_$sig"})
1368 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1370 select (undef, undef, undef, 0.1);
1373 kill $sig, $_; # srun wants two SIGINT to really interrupt
1375 $proc{$_}->{"sent_$sig"} = time;
1376 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1386 vec($bits,fileno($_),1) = 1;
1392 # Send log output to Keep via arv-put.
1394 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1395 # $log_pipe_pid is the pid of the arv-put subprocess.
1397 # The only functions that should access these variables directly are:
1399 # log_writer_start($logfilename)
1400 # Starts an arv-put pipe, reading data on stdin and writing it to
1401 # a $logfilename file in an output collection.
1403 # log_writer_send($txt)
1404 # Writes $txt to the output log collection.
1406 # log_writer_finish()
1407 # Closes the arv-put pipe and returns the output that it produces.
1409 # log_writer_is_active()
1410 # Returns a true value if there is currently a live arv-put
1411 # process, false otherwise.
1413 my ($log_pipe_in, $log_pipe_out, $log_pipe_pid);
1415 sub log_writer_start($)
1417 my $logfilename = shift;
1418 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1420 '--portable-data-hash',
1421 '--project-uuid', $Job->{owner_uuid},
1423 '--name', $logfilename,
1424 '--filename', $logfilename,
1428 sub log_writer_send($)
1431 print $log_pipe_in $txt;
1434 sub log_writer_finish()
1436 return unless $log_pipe_pid;
1438 close($log_pipe_in);
1441 my $s = IO::Select->new($log_pipe_out);
1442 if ($s->can_read(120)) {
1443 sysread($log_pipe_out, $arv_put_output, 1024);
1444 chomp($arv_put_output);
1446 Log (undef, "timed out reading from 'arv-put'");
1449 waitpid($log_pipe_pid, 0);
1450 $log_pipe_pid = $log_pipe_in = $log_pipe_out = undef;
1452 Log("log_writer_finish: arv-put exited ".exit_status_s($?))
1455 return $arv_put_output;
1458 sub log_writer_is_active() {
1459 return $log_pipe_pid;
1462 sub Log # ($jobstep_id, $logmessage)
1464 if ($_[1] =~ /\n/) {
1465 for my $line (split (/\n/, $_[1])) {
1470 my $fh = select STDERR; $|=1; select $fh;
1471 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1472 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1475 if (log_writer_is_active() || -t STDERR) {
1476 my @gmtime = gmtime;
1477 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1478 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1480 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1482 if (log_writer_is_active()) {
1483 log_writer_send($datetime . " " . $message);
1490 my ($package, $file, $line) = caller;
1491 my $message = "@_ at $file line $line\n";
1492 Log (undef, $message);
1493 freeze() if @jobstep_todo;
1494 create_output_collection() if @jobstep_todo;
1504 if ($Job->{'state'} eq 'Cancelled') {
1505 $Job->update_attributes('finished_at' => scalar gmtime);
1507 $Job->update_attributes('state' => 'Failed');
1514 my $justcheckpoint = shift; # false if this will be the last meta saved
1515 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1516 return unless log_writer_is_active();
1518 my $loglocator = log_writer_finish();
1519 Log (undef, "log manifest is $loglocator");
1520 $Job->{'log'} = $loglocator;
1521 $Job->update_attributes('log', $loglocator);
1525 sub freeze_if_want_freeze
1527 if ($main::please_freeze)
1529 release_allocation();
1532 # kill some srun procs before freeze+stop
1533 map { $proc{$_} = {} } @_;
1536 killem (keys %proc);
1537 select (undef, undef, undef, 0.1);
1539 while (($died = waitpid (-1, WNOHANG)) > 0)
1541 delete $proc{$died};
1546 create_output_collection();
1556 Log (undef, "Freeze not implemented");
1563 croak ("Thaw not implemented");
1579 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1586 my $srunargs = shift;
1587 my $execargs = shift;
1588 my $opts = shift || {};
1590 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1592 $Data::Dumper::Terse = 1;
1593 $Data::Dumper::Indent = 0;
1594 my $show_cmd = Dumper($args);
1595 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1596 $show_cmd =~ s/\n/ /g;
1597 warn "starting: $show_cmd\n";
1599 if (defined $stdin) {
1600 my $child = open STDIN, "-|";
1601 defined $child or die "no fork: $!";
1603 print $stdin or die $!;
1604 close STDOUT or die $!;
1609 return system (@$args) if $opts->{fork};
1612 warn "ENV size is ".length(join(" ",%ENV));
1613 die "exec failed: $!: @$args";
1617 sub ban_node_by_slot {
1618 # Don't start any new jobsteps on this node for 60 seconds
1620 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1621 $slot[$slotid]->{node}->{hold_count}++;
1622 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1627 my ($lockfile, $error_message) = @_;
1628 open L, ">", $lockfile or croak("$lockfile: $!");
1629 if (!flock L, LOCK_EX|LOCK_NB) {
1630 croak("Can't lock $lockfile: $error_message\n");
1634 sub find_docker_image {
1635 # Given a Keep locator, check to see if it contains a Docker image.
1636 # If so, return its stream name and Docker hash.
1637 # If not, return undef for both values.
1638 my $locator = shift;
1639 my ($streamname, $filename);
1640 my $image = api_call("collections/get", uuid => $locator);
1642 foreach my $line (split(/\n/, $image->{manifest_text})) {
1643 my @tokens = split(/\s+/, $line);
1645 $streamname = shift(@tokens);
1646 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1647 if (defined($filename)) {
1648 return (undef, undef); # More than one file in the Collection.
1650 $filename = (split(/:/, $filedata, 3))[2];
1655 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1656 return ($streamname, $1);
1658 return (undef, undef);
1663 # Calculate the number of times an operation should be retried,
1664 # assuming exponential backoff, and that we're willing to retry as
1665 # long as tasks have been running. Enforce a minimum of 3 retries.
1666 my ($starttime, $endtime, $timediff, $retries);
1668 $starttime = $jobstep[0]->{starttime};
1669 $endtime = $jobstep[-1]->{finishtime};
1671 if (!defined($starttime)) {
1673 } elsif (!defined($endtime)) {
1674 $timediff = time - $starttime;
1676 $timediff = ($endtime - $starttime) - (time - $endtime);
1678 if ($timediff > 0) {
1679 $retries = int(log($timediff) / log(2));
1681 $retries = 1; # Use the minimum.
1683 return ($retries > 3) ? $retries : 3;
1687 # Pass in two function references.
1688 # This method will be called with the remaining arguments.
1689 # If it dies, retry it with exponential backoff until it succeeds,
1690 # or until the current retry_count is exhausted. After each failure
1691 # that can be retried, the second function will be called with
1692 # the current try count (0-based), next try time, and error message.
1693 my $operation = shift;
1694 my $retry_callback = shift;
1695 my $retries = retry_count();
1696 foreach my $try_count (0..$retries) {
1697 my $next_try = time + (2 ** $try_count);
1698 my $result = eval { $operation->(@_); };
1701 } elsif ($try_count < $retries) {
1702 $retry_callback->($try_count, $next_try, $@);
1703 my $sleep_time = $next_try - time;
1704 sleep($sleep_time) if ($sleep_time > 0);
1707 # Ensure the error message ends in a newline, so Perl doesn't add
1708 # retry_op's line number to it.
1714 # Pass in a /-separated API method name, and arguments for it.
1715 # This function will call that method, retrying as needed until
1716 # the current retry_count is exhausted, with a log on the first failure.
1717 my $method_name = shift;
1718 my $log_api_retry = sub {
1719 my ($try_count, $next_try_at, $errmsg) = @_;
1720 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1721 $errmsg =~ s/\s/ /g;
1722 $errmsg =~ s/\s+$//;
1724 if ($next_try_at < time) {
1725 $retry_msg = "Retrying.";
1727 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1728 $retry_msg = "Retrying at $next_try_fmt.";
1730 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1733 foreach my $key (split(/\//, $method_name)) {
1734 $method = $method->{$key};
1736 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1740 # Given a $?, return a human-readable exit code string like "0" or
1741 # "1" or "0 with signal 1" or "1 with signal 11".
1742 my $exitcode = shift;
1743 my $s = $exitcode >> 8;
1744 if ($exitcode & 0x7f) {
1745 $s .= " with signal " . ($exitcode & 0x7f);
1747 if ($exitcode & 0x80) {
1748 $s .= " with core dump";
1753 sub handle_readall {
1754 # Pass in a glob reference to a file handle.
1755 # Read all its contents and return them as a string.
1756 my $fh_glob_ref = shift;
1758 return <$fh_glob_ref>;
1761 sub tar_filename_n {
1763 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1766 sub add_git_archive {
1767 # Pass in a git archive command as a string or list, a la system().
1768 # This method will save its output to be included in the archive sent to the
1772 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1773 croak("Failed to save git archive: $!");
1775 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1777 waitpid($git_pid, 0);
1780 croak("Failed to save git archive: git exited " . exit_status_s($?));
1784 sub combined_git_archive {
1785 # Combine all saved tar archives into a single archive, then return its
1786 # contents in a string. Return undef if no archives have been saved.
1787 if ($git_tar_count < 1) {
1790 my $base_tar_name = tar_filename_n(1);
1791 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
1792 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
1793 if ($tar_exit != 0) {
1794 croak("Error preparing build archive: tar -A exited " .
1795 exit_status_s($tar_exit));
1798 if (!open(GIT_TAR, "<", $base_tar_name)) {
1799 croak("Could not open build archive: $!");
1801 my $tar_contents = handle_readall(\*GIT_TAR);
1803 return $tar_contents;
1809 # This is crunch-job's internal dispatch script. crunch-job running on the API
1810 # server invokes this script on individual compute nodes, or localhost if we're
1811 # running a job locally. It gets called in two modes:
1813 # * No arguments: Installation mode. Read a tar archive from the DATA
1814 # file handle; it includes the Crunch script's source code, and
1815 # maybe SDKs as well. Those should be installed in the proper
1816 # locations. This runs outside of any Docker container, so don't try to
1817 # introspect Crunch's runtime environment.
1819 # * With arguments: Crunch script run mode. This script should set up the
1820 # environment, then run the command specified in the arguments. This runs
1821 # inside any Docker container.
1824 use File::Path qw( make_path remove_tree );
1825 use POSIX qw(getcwd);
1827 # Map SDK subdirectories to the path environments they belong to.
1828 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
1830 my $destdir = $ENV{"CRUNCH_SRC"};
1831 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1832 my $repo = $ENV{"CRUNCH_SRC_URL"};
1833 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
1834 my $job_work = $ENV{"JOB_WORK"};
1835 my $task_work = $ENV{"TASK_WORK"};
1837 for my $dir ($destdir, $job_work, $task_work) {
1840 -e $dir or die "Failed to create temporary directory ($dir): $!";
1845 remove_tree($task_work, {keep_root => 1});
1848 open(STDOUT_ORIG, ">&", STDOUT);
1849 open(STDERR_ORIG, ">&", STDERR);
1850 open(STDOUT, ">>", "$destdir.log");
1851 open(STDERR, ">&", STDOUT);
1853 ### Crunch script run mode
1855 # We want to do routine logging during task 0 only. This gives the user
1856 # the information they need, but avoids repeating the information for every
1859 if ($ENV{TASK_SEQUENCE} eq "0") {
1862 printf STDERR_ORIG "[Crunch] $msg\n", @_;
1868 my $python_src = "$install_dir/python";
1869 my $venv_dir = "$job_work/.arvados.venv";
1870 my $venv_built = -e "$venv_dir/bin/activate";
1871 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
1872 shell_or_die("virtualenv", "--quiet", "--system-site-packages",
1873 "--python=python2.7", $venv_dir);
1874 shell_or_die("$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
1876 $Log->("Built Python SDK virtualenv");
1879 my $pip_bin = "pip";
1881 $Log->("Running in Python SDK virtualenv");
1882 $pip_bin = "$venv_dir/bin/pip";
1883 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
1884 @ARGV = ("/bin/sh", "-ec",
1885 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
1886 } elsif (-d $python_src) {
1887 $Log->("Warning: virtualenv not found inside Docker container default " .
1888 "\$PATH. Can't install Python SDK.");
1891 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
1893 $Log->("Using Arvados SDK:");
1894 foreach my $line (split /\n/, $pkgs) {
1898 $Log->("Arvados SDK packages not found");
1901 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
1902 my $sdk_path = "$install_dir/$sdk_dir";
1904 if ($ENV{$sdk_envkey}) {
1905 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
1907 $ENV{$sdk_envkey} = $sdk_path;
1909 $Log->("Arvados SDK added to %s", $sdk_envkey);
1915 open(STDOUT, ">&", STDOUT_ORIG);
1916 open(STDERR, ">&", STDERR_ORIG);
1918 die "Cannot exec `@ARGV`: $!";
1921 ### Installation mode
1922 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1924 if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
1925 # This version already installed -> nothing to do.
1929 unlink "$destdir.commit";
1932 if (!open(TARX, "|-", "tar", "-xC", $destdir)) {
1933 die "Error launching 'tar -xC $destdir': $!";
1935 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
1936 # get SIGPIPE. We must feed it data incrementally.
1938 while (read(DATA, $tar_input, 65536)) {
1939 print TARX $tar_input;
1942 die "'tar -xC $destdir' exited $?: $!";
1947 my $sdk_root = "$destdir/.arvados.sdk/sdk";
1949 foreach my $sdk_lang (("python",
1950 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
1951 if (-d "$sdk_root/$sdk_lang") {
1952 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
1953 die "Failed to install $sdk_lang SDK: $!";
1959 my $python_dir = "$install_dir/python";
1960 if ((-d $python_dir) and can_run("python2.7") and
1961 (system("python2.7", "$python_dir/setup.py", "--quiet", "egg_info") != 0)) {
1962 # egg_info failed, probably when it asked git for a build tag.
1963 # Specify no build tag.
1964 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
1965 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
1969 if (-e "$destdir/crunch_scripts/install") {
1970 shell_or_die ("$destdir/crunch_scripts/install", $install_dir);
1971 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
1973 shell_or_die ("./tests/autotests.sh", $install_dir);
1974 } elsif (-e "./install.sh") {
1975 shell_or_die ("./install.sh", $install_dir);
1979 unlink "$destdir.commit.new";
1980 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
1981 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
1987 my $command_name = shift;
1988 open(my $which, "-|", "which", $command_name);
1989 while (<$which>) { }
1996 if ($ENV{"DEBUG"}) {
1997 print STDERR "@_\n";
1999 if (system (@_) != 0) {
2001 my $exitstatus = sprintf("exit %d signal %d", $? >> 8, $? & 0x7f);
2002 open STDERR, ">&STDERR_ORIG";
2003 system ("cat $destdir.log >&2");
2004 die "@_ failed ($err): $exitstatus";