7 (ARGV.any? ? ARGV : ['--jobs', '--pipelines']).each do |arg|
10 $options[:jobs] = true
12 $options[:pipelines] = true
14 abort "Unrecognized command line option '#{arg}'"
17 if not ($options[:jobs] or $options[:pipelines])
18 abort "Nothing to do. Please specify at least one of: --jobs, --pipelines."
21 ARGV.reject! { |a| a =~ /--jobs|--pipelines/ }
25 %w{TERM INT}.each do |sig|
28 $stderr.puts "Received #{signame} signal"
33 if ENV["CRUNCH_DISPATCH_LOCKFILE"]
34 lockfilename = ENV.delete "CRUNCH_DISPATCH_LOCKFILE"
35 lockfile = File.open(lockfilename, File::RDWR|File::CREAT, 0644)
36 unless lockfile.flock File::LOCK_EX|File::LOCK_NB
37 abort "Lock unavailable on #{lockfilename} - exit"
41 ENV["RAILS_ENV"] = ARGV[0] || ENV["RAILS_ENV"] || "development"
43 require File.dirname(__FILE__) + '/../config/boot'
44 require File.dirname(__FILE__) + '/../config/environment'
49 self.utc.strftime "%Y-%m-%d_%H:%M:%S"
54 include ApplicationHelper
59 @crunch_job_bin = (ENV['CRUNCH_JOB_BIN'] || `which arv-crunch-job`.strip)
60 if @crunch_job_bin.empty?
61 raise "No CRUNCH_JOB_BIN env var, and crunch-job not in path."
64 @arvados_internal = Rails.configuration.git_internal_dir
65 if not File.exists? @arvados_internal
66 $stderr.puts `mkdir -p #{@arvados_internal.shellescape} && git init --bare #{@arvados_internal.shellescape}`
67 raise "No internal git repository available" unless ($? == 0)
70 @repo_root = Rails.configuration.git_repositories_dir
71 @arvados_repo_path = Repository.where(name: "arvados").first.server_path
77 @pipe_auth_tokens = {}
84 return act_as_system_user
89 @todo = Job.queue.select(&:repository)
91 if $options[:pipelines]
92 @todo_pipelines = PipelineInstance.queue
96 def each_slurm_line(cmd, outfmt, max_fields=nil)
97 max_fields ||= outfmt.split(":").size
98 max_fields += 1 # To accommodate the node field we add
99 @@slurm_version ||= Gem::Version.new(`sinfo --version`.match(/\b[\d\.]+\b/)[0])
100 if Gem::Version.new('2.3') <= @@slurm_version
101 `#{cmd} --noheader -o '%n:#{outfmt}'`.each_line do |line|
102 yield line.chomp.split(":", max_fields)
105 # Expand rows with hostname ranges (like "foo[1-3,5,9-12]:idle")
106 # into multiple rows with one hostname each.
107 `#{cmd} --noheader -o '%N:#{outfmt}'`.each_line do |line|
108 tokens = line.chomp.split(":", max_fields)
109 if (re = tokens[0].match /^(.*?)\[([-,\d]+)\]$/)
111 re[2].split(",").each do |range|
112 range = range.split("-").collect(&:to_i)
113 (range[0]..range[-1]).each do |n|
114 yield [re[1] + n.to_s] + tokens
126 each_slurm_line("sinfo", "%t") do |hostname, state|
127 # Treat nodes in idle* state as down, because the * means that slurm
128 # hasn't been able to communicate with it recently.
129 state.sub!(/^idle\*/, "down")
130 state.sub!(/\W+$/, "")
131 state = "down" unless %w(idle alloc down).include?(state)
132 slurm_nodes[hostname] = {state: state, job: nil}
134 each_slurm_line("squeue", "%j") do |hostname, job_uuid|
135 slurm_nodes[hostname][:job] = job_uuid if slurm_nodes[hostname]
140 def update_node_status
141 return unless Server::Application.config.crunch_job_wrapper.to_s.match /^slurm/
142 slurm_status.each_pair do |hostname, slurmdata|
143 next if @node_state[hostname] == slurmdata
145 node = Node.where('hostname=?', hostname).order(:last_ping_at).last
147 $stderr.puts "dispatch: update #{hostname} state to #{slurmdata}"
148 node.info["slurm_state"] = slurmdata[:state]
149 node.job_uuid = slurmdata[:job]
151 @node_state[hostname] = slurmdata
153 $stderr.puts "dispatch: failed to update #{node.uuid}: #{node.errors.messages}"
155 elsif slurmdata[:state] != 'down'
156 $stderr.puts "dispatch: SLURM reports '#{hostname}' is not down, but no node has that name"
159 $stderr.puts "dispatch: error updating #{hostname} node status: #{error}"
164 def positive_int(raw_value, default=nil)
165 value = begin raw_value.to_i rescue 0 end
173 NODE_CONSTRAINT_MAP = {
174 # Map Job runtime_constraints keys to the corresponding Node info key.
175 'min_ram_mb_per_node' => 'total_ram_mb',
176 'min_scratch_mb_per_node' => 'total_scratch_mb',
177 'min_cores_per_node' => 'total_cpu_cores',
180 def nodes_available_for_job_now(job)
181 # Find Nodes that satisfy a Job's runtime constraints (by building
182 # a list of Procs and using them to test each Node). If there
183 # enough to run the Job, return an array of their names.
184 # Otherwise, return nil.
185 need_procs = NODE_CONSTRAINT_MAP.each_pair.map do |job_key, node_key|
187 positive_int(node.info[node_key], 0) >=
188 positive_int(job.runtime_constraints[job_key], 0)
191 min_node_count = positive_int(job.runtime_constraints['min_nodes'], 1)
193 Node.find_each do |node|
194 good_node = (node.info['slurm_state'] == 'idle')
195 need_procs.each { |node_test| good_node &&= node_test.call(node) }
198 if usable_nodes.count >= min_node_count
199 return usable_nodes.map { |node| node.hostname }
206 def nodes_available_for_job(job)
207 # Check if there are enough idle nodes with the Job's minimum
208 # hardware requirements to run it. If so, return an array of
209 # their names. If not, up to once per hour, signal start_jobs to
210 # hold off launching Jobs. This delay is meant to give the Node
211 # Manager an opportunity to make new resources available for new
214 # The exact timing parameters here might need to be adjusted for
215 # the best balance between helping the longest-waiting Jobs run,
216 # and making efficient use of immediately available resources.
217 # These are all just first efforts until we have more data to work
219 nodelist = nodes_available_for_job_now(job)
220 if nodelist.nil? and not did_recently(:wait_for_available_nodes, 3600)
221 $stderr.puts "dispatch: waiting for nodes for #{job.uuid}"
222 @node_wait_deadline = Time.now + 5.minutes
227 def fail_job job, message
228 $stderr.puts "dispatch: #{job.uuid}: #{message}"
230 Log.new(object_uuid: job.uuid,
231 event_type: 'dispatch',
232 owner_uuid: job.owner_uuid,
234 properties: {"text" => message}).save!
236 $stderr.puts "dispatch: log.create failed"
240 job.lock @authorizations[job.uuid].user.uuid
243 $stderr.puts "dispatch: save failed setting job #{job.uuid} to failed"
245 rescue ArvadosModel::AlreadyLockedError
246 $stderr.puts "dispatch: tried to mark job #{job.uuid} as failed but it was already locked by someone else"
250 def stdout_s(cmd_a, opts={})
251 IO.popen(cmd_a, "r", opts) do |pipe|
252 return pipe.read.chomp
257 ["git", "--git-dir=#{@arvados_internal}"] + cmd_a
260 def get_authorization(job)
261 if @authorizations[job.uuid] and
262 @authorizations[job.uuid].user.uuid != job.modified_by_user_uuid
263 # We already made a token for this job, but we need a new one
264 # because modified_by_user_uuid has changed (the job will run
265 # as a different user).
266 @authorizations[job.uuid].update_attributes expires_at: Time.now
267 @authorizations[job.uuid] = nil
269 if not @authorizations[job.uuid]
270 auth = ApiClientAuthorization.
271 new(user: User.where('uuid=?', job.modified_by_user_uuid).first,
274 $stderr.puts "dispatch: auth.save failed for #{job.uuid}"
276 @authorizations[job.uuid] = auth
279 @authorizations[job.uuid]
282 def get_commit(src_repo, commit_hash)
283 # @fetched_commits[V]==true if we know commit V exists in the
284 # arvados_internal git repository.
285 if !@fetched_commits[commit_hash]
286 # check if the commit needs to be fetched or not
287 commit_rev = stdout_s(git_cmd("rev-list", "-n1", commit_hash),
289 unless $? == 0 and commit_rev == commit_hash
290 # commit does not exist in internal repository, so import the source repository using git fetch-pack
291 cmd = git_cmd("fetch-pack", "--no-progress", "--all", src_repo)
292 $stderr.puts "dispatch: #{cmd}"
293 $stderr.puts(stdout_s(cmd))
295 fail_job job, "git fetch-pack failed"
299 @fetched_commits[commit_hash] = true
301 @fetched_commits[commit_hash]
304 def tag_commit(commit_hash, tag_name)
305 # @git_tags[T]==V if we know commit V has been tagged T in the
306 # arvados_internal repository.
307 if not @git_tags[tag_name]
308 cmd = git_cmd("tag", tag_name, commit_hash)
309 $stderr.puts "dispatch: #{cmd}"
310 $stderr.puts(stdout_s(cmd, err: "/dev/null"))
312 # git tag failed. This may be because the tag already exists, so check for that.
313 tag_rev = stdout_s(git_cmd("rev-list", "-n1", tag_name))
315 # We got a revision back
316 if tag_rev != commit_hash
317 # Uh oh, the tag doesn't point to the revision we were expecting.
318 # Someone has been monkeying with the job record and/or git.
319 fail_job job, "Existing tag #{tag_name} points to commit #{tag_rev} but expected commit #{commit_hash}"
322 # we're okay (fall through to setting @git_tags below)
324 # git rev-list failed for some reason.
325 fail_job job, "'git tag' for #{tag_name} failed but did not find any existing tag using 'git rev-list'"
329 # 'git tag' was successful, or there is an existing tag that points to the same revision.
330 @git_tags[tag_name] = commit_hash
331 elsif @git_tags[tag_name] != commit_hash
332 fail_job job, "Existing tag #{tag_name} points to commit #{@git_tags[tag_name]} but this job uses commit #{commit_hash}"
340 next if @running[job.uuid]
343 case Server::Application.config.crunch_job_wrapper
346 # Don't run more than one at a time.
350 when :slurm_immediate
351 nodelist = nodes_available_for_job(job)
353 if Time.now < @node_wait_deadline
359 cmd_args = ["salloc",
364 "--job-name=#{job.uuid}",
365 "--nodelist=#{nodelist.join(',')}"]
367 raise "Unknown crunch_job_wrapper: #{Server::Application.config.crunch_job_wrapper}"
370 if Server::Application.config.crunch_job_user
371 cmd_args.unshift("sudo", "-E", "-u",
372 Server::Application.config.crunch_job_user,
373 "PATH=#{ENV['PATH']}",
374 "PERLLIB=#{ENV['PERLLIB']}",
375 "PYTHONPATH=#{ENV['PYTHONPATH']}",
376 "RUBYLIB=#{ENV['RUBYLIB']}",
377 "GEM_PATH=#{ENV['GEM_PATH']}")
380 repo = Repository.where(name: job.repository).first
381 if repo.nil? or repo.server_path.nil?
382 fail_job "Repository #{job.repository} not found under #{@repo_root}"
386 ready = (get_authorization(job) and
387 get_commit(repo.server_path, job.script_version) and
388 tag_commit(job.script_version, job.uuid))
389 if ready and job.arvados_sdk_version
390 ready = (get_commit(@arvados_repo_path, job.arvados_sdk_version) and
391 tag_commit(job.arvados_sdk_version, "#{job.uuid}-arvados-sdk"))
395 cmd_args += [@crunch_job_bin,
396 '--job-api-token', @authorizations[job.uuid].api_token,
398 '--git-dir', @arvados_internal]
400 $stderr.puts "dispatch: #{cmd_args.join ' '}"
403 i, o, e, t = Open3.popen3(*cmd_args)
405 $stderr.puts "dispatch: popen3: #{$!}"
410 $stderr.puts "dispatch: job #{job.uuid}"
411 start_banner = "dispatch: child #{t.pid} start #{LogTime.now}"
412 $stderr.puts start_banner
414 @running[job.uuid] = {
420 buf: {stderr: '', stdout: ''},
423 job_auth: @authorizations[job.uuid],
424 stderr_buf_to_flush: '',
425 stderr_flushed_at: Time.new(0),
428 log_throttle_is_open: true,
429 log_throttle_reset_time: Time.now + Rails.configuration.crunch_log_throttle_period,
430 log_throttle_bytes_so_far: 0,
431 log_throttle_lines_so_far: 0,
432 log_throttle_bytes_skipped: 0,
439 # Test for hard cap on total output and for log throttling. Returns whether
440 # the log line should go to output or not. Modifies "line" in place to
441 # replace it with an error if a logging limit is tripped.
442 def rate_limit running_job, line
445 if running_job[:log_throttle_is_open]
446 running_job[:log_throttle_lines_so_far] += 1
447 running_job[:log_throttle_bytes_so_far] += linesize
448 running_job[:bytes_logged] += linesize
450 if (running_job[:bytes_logged] >
451 Rails.configuration.crunch_limit_log_bytes_per_job)
452 message = "Exceeded log limit #{Rails.configuration.crunch_limit_log_bytes_per_job} bytes (crunch_limit_log_bytes_per_job). Log will be truncated."
453 running_job[:log_throttle_reset_time] = Time.now + 100.years
454 running_job[:log_throttle_is_open] = false
456 elsif (running_job[:log_throttle_bytes_so_far] >
457 Rails.configuration.crunch_log_throttle_bytes)
458 remaining_time = running_job[:log_throttle_reset_time] - Time.now
459 message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_bytes} bytes per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_bytes). Logging will be silenced for the next #{remaining_time.round} seconds.\n"
460 running_job[:log_throttle_is_open] = false
462 elsif (running_job[:log_throttle_lines_so_far] >
463 Rails.configuration.crunch_log_throttle_lines)
464 remaining_time = running_job[:log_throttle_reset_time] - Time.now
465 message = "Exceeded rate #{Rails.configuration.crunch_log_throttle_lines} lines per #{Rails.configuration.crunch_log_throttle_period} seconds (crunch_log_throttle_lines), logging will be silenced for the next #{remaining_time.round} seconds.\n"
466 running_job[:log_throttle_is_open] = false
470 if not running_job[:log_throttle_is_open]
471 # Don't log anything if any limit has been exceeded. Just count lossage.
472 running_job[:log_throttle_bytes_skipped] += linesize
476 # Yes, write to logs, but use our "rate exceeded" message
477 # instead of the log message that exceeded the limit.
481 running_job[:log_throttle_is_open]
486 @running.each do |job_uuid, j|
490 if now > j[:log_throttle_reset_time]
491 # It has been more than throttle_period seconds since the last
492 # checkpoint so reset the throttle
493 if j[:log_throttle_bytes_skipped] > 0
494 message = "#{job_uuid} ! Skipped #{j[:log_throttle_bytes_skipped]} bytes of log"
496 j[:stderr_buf_to_flush] << "#{LogTime.now} #{message}\n"
499 j[:log_throttle_reset_time] = now + Rails.configuration.crunch_log_throttle_period
500 j[:log_throttle_bytes_so_far] = 0
501 j[:log_throttle_lines_so_far] = 0
502 j[:log_throttle_bytes_skipped] = 0
503 j[:log_throttle_is_open] = true
506 j[:buf].each do |stream, streambuf|
507 # Read some data from the child stream
510 # It's important to use a big enough buffer here. When we're
511 # being flooded with logs, we must read and discard many
512 # bytes at once. Otherwise, we can easily peg a CPU with
513 # time-checking and other loop overhead. (Quick tests show a
514 # 1MiB buffer working 2.5x as fast as a 64 KiB buffer.)
516 # So don't reduce this buffer size!
517 buf = j[stream].read_nonblock(2**20)
518 rescue Errno::EAGAIN, EOFError
521 # Short circuit the counting code if we're just going to throw
522 # away the data anyway.
523 if not j[:log_throttle_is_open]
524 j[:log_throttle_bytes_skipped] += streambuf.size + buf.size
531 # Append to incomplete line from previous read, if any
535 streambuf.each_line do |line|
536 if not line.end_with? $/
537 if line.size > Rails.configuration.crunch_log_throttle_bytes
538 # Without a limit here, we'll use 2x an arbitrary amount
539 # of memory, and waste a lot of time copying strings
540 # around, all without providing any feedback to anyone
541 # about what's going on _or_ hitting any of our throttle
544 # Here we leave "line" alone, knowing it will never be
545 # sent anywhere: rate_limit() will reach
546 # crunch_log_throttle_bytes immediately. However, we'll
547 # leave [...] in bufend: if the trailing end of the long
548 # line does end up getting sent anywhere, it will have
549 # some indication that it is incomplete.
552 # If line length is sane, we'll wait for the rest of the
553 # line to appear in the next read_pipes() call.
558 # rate_limit returns true or false as to whether to actually log
559 # the line or not. It also modifies "line" in place to replace
560 # it with an error if a logging limit is tripped.
561 if rate_limit j, line
562 $stderr.print "#{job_uuid} ! " unless line.index(job_uuid)
564 pub_msg = "#{LogTime.now} #{line.strip}\n"
565 j[:stderr_buf_to_flush] << pub_msg
569 # Leave the trailing incomplete line (if any) in streambuf for
571 streambuf.replace bufend
573 # Flush buffered logs to the logs table, if appropriate. We have
574 # to do this even if we didn't collect any new logs this time:
575 # otherwise, buffered data older than seconds_between_events
576 # won't get flushed until new data arrives.
582 return if 0 == @running.size
588 pid_done = waitpid(-1, Process::WNOHANG | Process::WUNTRACED)
590 j_done = @running.values.
591 select { |j| j[:wait_thr].pid == pid_done }.
594 rescue SystemCallError
595 # I have @running processes but system reports I have no
596 # children. This is likely to happen repeatedly if it happens at
597 # all; I will log this no more than once per child process I
599 if 0 < @running.select { |uuid,j| j[:warned_waitpid_error].nil? }.size
600 children = @running.values.collect { |j| j[:wait_thr].pid }.join ' '
601 $stderr.puts "dispatch: IPC bug: waitpid() error (#{$!}), but I have children #{children}"
603 @running.each do |uuid,j| j[:warned_waitpid_error] = true end
606 @running.each do |uuid, j|
607 if j[:wait_thr].status == false
608 pid_done = j[:wait_thr].pid
616 job_done = j_done[:job]
617 $stderr.puts "dispatch: child #{pid_done} exit"
618 $stderr.puts "dispatch: job #{job_done.uuid} end"
620 # Ensure every last drop of stdout and stderr is consumed.
622 # Reset flush timestamp to make sure log gets written.
623 j_done[:stderr_flushed_at] = Time.new(0)
624 # Write any remaining logs.
627 j_done[:buf].each do |stream, streambuf|
629 $stderr.puts streambuf + "\n"
633 # Wait the thread (returns a Process::Status)
634 exit_status = j_done[:wait_thr].value.exitstatus
636 jobrecord = Job.find_by_uuid(job_done.uuid)
637 if exit_status != EXIT_TEMPFAIL and jobrecord.state == "Running"
638 # crunch-job did not return exit code 75 (see below) and left the job in
639 # the "Running" state, which means there was an unhandled error. Fail
641 jobrecord.state = "Failed"
642 if not jobrecord.save
643 $stderr.puts "dispatch: jobrecord.save failed"
646 # Don't fail the job if crunch-job didn't even get as far as
647 # starting it. If the job failed to run due to an infrastructure
648 # issue with crunch-job or slurm, we want the job to stay in the
649 # queue. If crunch-job exited after losing a race to another
650 # crunch-job process, it exits 75 and we should leave the job
651 # record alone so the winner of the race do its thing.
653 # There is still an unhandled race condition: If our crunch-job
654 # process is about to lose a race with another crunch-job
655 # process, but crashes before getting to its "exit 75" (for
656 # example, "cannot fork" or "cannot reach API server") then we
657 # will assume incorrectly that it's our process's fault
658 # jobrecord.started_at is non-nil, and mark the job as failed
659 # even though the winner of the race is probably still doing
663 # Invalidate the per-job auth token, unless the job is still queued and we
664 # might want to try it again.
665 if jobrecord.state != "Queued"
666 j_done[:job_auth].update_attributes expires_at: Time.now
669 @running.delete job_done.uuid
673 expire_tokens = @pipe_auth_tokens.dup
674 @todo_pipelines.each do |p|
675 pipe_auth = (@pipe_auth_tokens[p.uuid] ||= ApiClientAuthorization.
676 create(user: User.where('uuid=?', p.modified_by_user_uuid).first,
678 puts `export ARVADOS_API_TOKEN=#{pipe_auth.api_token} && arv-run-pipeline-instance --run-pipeline-here --no-wait --instance #{p.uuid}`
679 expire_tokens.delete p.uuid
682 expire_tokens.each do |k, v|
683 v.update_attributes expires_at: Time.now
684 @pipe_auth_tokens.delete k
690 $stderr.puts "dispatch: ready"
691 while !$signal[:term] or @running.size > 0
694 @running.each do |uuid, j|
695 if !j[:started] and j[:sent_int] < 2
697 Process.kill 'INT', j[:wait_thr].pid
699 # No such pid = race condition + desired result is
706 refresh_todo unless did_recently(:refresh_todo, 1.0)
707 update_node_status unless did_recently(:update_node_status, 1.0)
708 unless @todo.empty? or did_recently(:start_jobs, 1.0) or $signal[:term]
711 unless (@todo_pipelines.empty? and @pipe_auth_tokens.empty?) or did_recently(:update_pipelines, 5.0)
716 select(@running.values.collect { |j| [j[:stdout], j[:stderr]] }.flatten,
723 def did_recently(thing, min_interval)
724 if !@did_recently[thing] or @did_recently[thing] < Time.now - min_interval
725 @did_recently[thing] = Time.now
732 # send message to log table. we want these records to be transient
733 def write_log running_job
734 return if running_job[:stderr_buf_to_flush] == ''
736 # Send out to log event if buffer size exceeds the bytes per event or if
737 # it has been at least crunch_log_seconds_between_events seconds since
739 if running_job[:stderr_buf_to_flush].size > Rails.configuration.crunch_log_bytes_per_event or
740 (Time.now - running_job[:stderr_flushed_at]) >= Rails.configuration.crunch_log_seconds_between_events
742 log = Log.new(object_uuid: running_job[:job].uuid,
743 event_type: 'stderr',
744 owner_uuid: running_job[:job].owner_uuid,
745 properties: {"text" => running_job[:stderr_buf_to_flush]})
747 running_job[:events_logged] += 1
749 $stderr.puts "Failed to write logs"
750 $stderr.puts exception.backtrace
752 running_job[:stderr_buf_to_flush] = ''
753 running_job[:stderr_flushed_at] = Time.now
758 # This is how crunch-job child procs know where the "refresh" trigger file is
759 ENV["CRUNCH_REFRESH_TRIGGER"] = Rails.configuration.crunch_refresh_trigger
761 # If salloc can't allocate resources immediately, make it use our temporary
762 # failure exit code. This ensures crunch-dispatch won't mark a job failed
763 # because of an issue with node allocation. This often happens when
764 # another dispatcher wins the race to allocate nodes.
765 ENV["SLURM_EXIT_IMMEDIATE"] = Dispatcher::EXIT_TEMPFAIL.to_s