services/api/script/crunch-dispatch.rb

   1 #!/usr/bin/env ruby
   2
   3 include Process
   4
   5 $warned = {}
   6 $signal = {}
   7 %w{TERM INT}.each do |sig|
   8   signame = sig
   9   Signal.trap(sig) do
  10     $stderr.puts "Received #{signame} signal"
  11     $signal[:term] = true
  12   end
  13 end
  14
  15 if ENV["CRUNCH_DISPATCH_LOCKFILE"]
  16   lockfilename = ENV.delete "CRUNCH_DISPATCH_LOCKFILE"
  17   lockfile = File.open(lockfilename, File::RDWR|File::CREAT, 0644)
  18   unless lockfile.flock File::LOCK_EX|File::LOCK_NB
  19     abort "Lock unavailable on #{lockfilename} - exit"
  20   end
  21 end
  22
  23 ENV["RAILS_ENV"] = ARGV[0] || ENV["RAILS_ENV"] || "development"
  24
  25 require File.dirname(__FILE__) + '/../config/boot'
  26 require File.dirname(__FILE__) + '/../config/environment'
  27 require 'open3'
  28
  29 $redis ||= Redis.new
  30 LOG_BUFFER_SIZE = 2**20
  31
  32 class Dispatcher
  33   include ApplicationHelper
  34
  35   def sysuser
  36     return act_as_system_user
  37   end
  38
  39   def refresh_todo
  40     @todo = Job.queue
  41     @todo_pipelines = PipelineInstance.queue
  42   end
  43
  44   def sinfo
  45     @@slurm_version ||= Gem::Version.new(`sinfo --version`.match(/\b[\d\.]+\b/)[0])
  46     if Gem::Version.new('2.3') <= @@slurm_version
  47       `sinfo --noheader -o '%n:%t'`.strip
  48     else
  49       # Expand rows with hostname ranges (like "foo[1-3,5,9-12]:idle")
  50       # into multiple rows with one hostname each.
  51       `sinfo --noheader -o '%N:%t'`.split("\n").collect do |line|
  52         tokens = line.split ":"
  53         if (re = tokens[0].match /^(.*?)\[([-,\d]+)\]$/)
  54           re[2].split(",").collect do |range|
  55             range = range.split("-").collect(&:to_i)
  56             (range[0]..range[-1]).collect do |n|
  57               [re[1] + n.to_s, tokens[1..-1]].join ":"
  58             end
  59           end
  60         else
  61           tokens.join ":"
  62         end
  63       end.flatten.join "\n"
  64     end
  65   end
  66
  67   def update_node_status
  68     if Server::Application.config.crunch_job_wrapper.to_s.match /^slurm/
  69       @nodes_in_state = {idle: 0, alloc: 0, down: 0}
  70       @node_state ||= {}
  71       node_seen = {}
  72       begin
  73         sinfo.split("\n").
  74           each do |line|
  75           re = line.match /(\S+?):+(idle|alloc|down)/
  76           next if !re
  77
  78           # sinfo tells us about a node N times if it is shared by N partitions
  79           next if node_seen[re[1]]
  80           node_seen[re[1]] = true
  81
  82           # count nodes in each state
  83           @nodes_in_state[re[2].to_sym] += 1
  84
  85           # update our database (and cache) when a node's state changes
  86           if @node_state[re[1]] != re[2]
  87             @node_state[re[1]] = re[2]
  88             node = Node.where('hostname=?', re[1]).first
  89             if node
  90               $stderr.puts "dispatch: update #{re[1]} state to #{re[2]}"
  91               node.info[:slurm_state] = re[2]
  92               node.save
  93             elsif re[2] != 'down'
  94               $stderr.puts "dispatch: sinfo reports '#{re[1]}' is not down, but no node has that name"
  95             end
  96           end
  97         end
  98       rescue
  99       end
 100     end
 101   end
 102
 103   def start_jobs
 104     @todo.each do |job|
 105
 106       min_nodes = 1
 107       begin
 108         if job.runtime_constraints['min_nodes']
 109           min_nodes = begin job.runtime_constraints['min_nodes'].to_i rescue 1 end
 110         end
 111       end
 112
 113       begin
 114         next if @nodes_in_state[:idle] < min_nodes
 115       rescue
 116       end
 117
 118       next if @running[job.uuid]
 119       next if !take(job)
 120
 121       cmd_args = nil
 122       case Server::Application.config.crunch_job_wrapper
 123       when :none
 124         cmd_args = []
 125       when :slurm_immediate
 126         cmd_args = ["salloc",
 127                     "--chdir=/",
 128                     "--immediate",
 129                     "--exclusive",
 130                     "--no-kill",
 131                     "--job-name=#{job.uuid}",
 132                     "--nodes=#{min_nodes}"]
 133       else
 134         raise "Unknown crunch_job_wrapper: #{Server::Application.config.crunch_job_wrapper}"
 135       end
 136
 137       if Server::Application.config.crunch_job_user
 138         cmd_args.unshift("sudo", "-E", "-u",
 139                          Server::Application.config.crunch_job_user,
 140                          "PERLLIB=#{ENV['PERLLIB']}")
 141       end
 142
 143       job_auth = ApiClientAuthorization.
 144         new(user: User.where('uuid=?', job.modified_by_user_uuid).first,
 145             api_client_id: 0)
 146       job_auth.save
 147
 148       crunch_job_bin = (ENV['CRUNCH_JOB_BIN'] || `which arv-crunch-job`.strip)
 149       if crunch_job_bin == ''
 150         raise "No CRUNCH_JOB_BIN env var, and crunch-job not in path."
 151       end
 152
 153       cmd_args << crunch_job_bin
 154       cmd_args << '--job-api-token'
 155       cmd_args << job_auth.api_token
 156       cmd_args << '--job'
 157       cmd_args << job.uuid
 158
 159       commit = Commit.where(sha1: job.script_version).first
 160       if commit
 161         cmd_args << '--git-dir'
 162         if File.exists?(File.
 163                         join(Rails.configuration.git_repositories_dir,
 164                              commit.repository_name + '.git'))
 165           cmd_args << File.
 166             join(Rails.configuration.git_repositories_dir,
 167                  commit.repository_name + '.git')
 168         else
 169           cmd_args << File.
 170             join(Rails.configuration.git_repositories_dir,
 171                  commit.repository_name, '.git')
 172         end
 173       end
 174
 175       $stderr.puts "dispatch: #{cmd_args.join ' '}"
 176
 177       begin
 178         i, o, e, t = Open3.popen3(*cmd_args)
 179       rescue
 180         $stderr.puts "dispatch: popen3: #{$!}"
 181         sleep 1
 182         untake(job)
 183         next
 184       end
 185
 186       $stderr.puts "dispatch: job #{job.uuid}"
 187       start_banner = "dispatch: child #{t.pid} start #{Time.now.ctime.to_s}"
 188       $stderr.puts start_banner
 189       $redis.set job.uuid, start_banner + "\n"
 190       $redis.publish job.uuid, start_banner
 191       $redis.publish job.owner_uuid, start_banner
 192
 193       @running[job.uuid] = {
 194         stdin: i,
 195         stdout: o,
 196         stderr: e,
 197         wait_thr: t,
 198         job: job,
 199         stderr_buf: '',
 200         started: false,
 201         sent_int: 0,
 202         job_auth: job_auth
 203       }
 204       i.close
 205     end
 206   end
 207
 208   def take(job)
 209     # no-op -- let crunch-job take care of locking.
 210     true
 211   end
 212
 213   def untake(job)
 214     # no-op -- let crunch-job take care of locking.
 215     true
 216   end
 217
 218   def read_pipes
 219     @running.each do |job_uuid, j|
 220       job = j[:job]
 221
 222       # Throw away child stdout
 223       begin
 224         j[:stdout].read_nonblock(2**20)
 225       rescue Errno::EAGAIN, EOFError
 226       end
 227
 228       # Read whatever is available from child stderr
 229       stderr_buf = false
 230       begin
 231         stderr_buf = j[:stderr].read_nonblock(2**20)
 232       rescue Errno::EAGAIN, EOFError
 233       end
 234
 235       if stderr_buf
 236         j[:stderr_buf] << stderr_buf
 237         if j[:stderr_buf].index "\n"
 238           lines = j[:stderr_buf].lines("\n").to_a
 239           if j[:stderr_buf][-1] == "\n"
 240             j[:stderr_buf] = ''
 241           else
 242             j[:stderr_buf] = lines.pop
 243           end
 244           lines.each do |line|
 245             $stderr.print "#{job_uuid} ! " unless line.index(job_uuid)
 246             $stderr.puts line
 247             pub_msg = "#{Time.now.ctime.to_s} #{line.strip}"
 248             $redis.publish job.owner_uuid, pub_msg
 249             $redis.publish job_uuid, pub_msg
 250             $redis.append job_uuid, pub_msg + "\n"
 251             if LOG_BUFFER_SIZE < $redis.strlen(job_uuid)
 252               $redis.set(job_uuid,
 253                          $redis
 254                            .getrange(job_uuid, (LOG_BUFFER_SIZE >> 1), -1)
 255                            .sub(/^.*?\n/, ''))
 256             end
 257           end
 258         end
 259       end
 260     end
 261   end
 262
 263   def reap_children
 264     return if 0 == @running.size
 265     pid_done = nil
 266     j_done = nil
 267
 268     if false
 269       begin
 270         pid_done = waitpid(-1, Process::WNOHANG | Process::WUNTRACED)
 271         if pid_done
 272           j_done = @running.values.
 273             select { |j| j[:wait_thr].pid == pid_done }.
 274             first
 275         end
 276       rescue SystemCallError
 277         # I have @running processes but system reports I have no
 278         # children. This is likely to happen repeatedly if it happens at
 279         # all; I will log this no more than once per child process I
 280         # start.
 281         if 0 < @running.select { |uuid,j| j[:warned_waitpid_error].nil? }.size
 282           children = @running.values.collect { |j| j[:wait_thr].pid }.join ' '
 283           $stderr.puts "dispatch: IPC bug: waitpid() error (#{$!}), but I have children #{children}"
 284         end
 285         @running.each do |uuid,j| j[:warned_waitpid_error] = true end
 286       end
 287     else
 288       @running.each do |uuid, j|
 289         if j[:wait_thr].status == false
 290           pid_done = j[:wait_thr].pid
 291           j_done = j
 292         end
 293       end
 294     end
 295
 296     return if !pid_done
 297
 298     job_done = j_done[:job]
 299     $stderr.puts "dispatch: child #{pid_done} exit"
 300     $stderr.puts "dispatch: job #{job_done.uuid} end"
 301     $redis.publish job_done.uuid, "end"
 302
 303     # Ensure every last drop of stdout and stderr is consumed
 304     read_pipes
 305     if j_done[:stderr_buf] and j_done[:stderr_buf] != ''
 306       $stderr.puts j_done[:stderr_buf] + "\n"
 307     end
 308
 309     # Wait the thread
 310     j_done[:wait_thr].value
 311
 312     # Invalidate the per-job auth token
 313     j_done[:job_auth].update_attributes expires_at: Time.now
 314
 315     @running.delete job_done.uuid
 316   end
 317
 318   def update_pipelines
 319     @todo_pipelines.each do |p|
 320       pipe_auth = ApiClientAuthorization.
 321         new(user: User.where('uuid=?', p.modified_by_user_uuid).first,
 322             api_client_id: 0)
 323       pipe_auth.save
 324
 325       puts `export ARVADOS_API_TOKEN=#{pipe_auth.api_token} && arv-run-pipeline-instance --run-here --no-wait --instance #{p.uuid}`
 326     end
 327   end
 328
 329   def run
 330     act_as_system_user
 331     @running ||= {}
 332     $stderr.puts "dispatch: ready"
 333     while !$signal[:term] or @running.size > 0
 334       read_pipes
 335       if $signal[:term]
 336         @running.each do |uuid, j|
 337           if !j[:started] and j[:sent_int] < 2
 338             begin
 339               Process.kill 'INT', j[:wait_thr].pid
 340             rescue Errno::ESRCH
 341               # No such pid = race condition + desired result is
 342               # already achieved
 343             end
 344             j[:sent_int] += 1
 345           end
 346         end
 347       else
 348         refresh_todo unless did_recently(:refresh_todo, 1.0)
 349         update_node_status
 350         unless @todo.empty? or did_recently(:start_jobs, 1.0) or $signal[:term]
 351           start_jobs
 352         end
 353         unless @todo_pipelines.empty? or did_recently(:update_pipelines, 5.0)
 354           update_pipelines
 355         end
 356       end
 357       reap_children
 358       select(@running.values.collect { |j| [j[:stdout], j[:stderr]] }.flatten,
 359              [], [], 1)
 360     end
 361   end
 362
 363
 364
 365   protected
 366
 367   def did_recently(thing, min_interval)
 368     @did_recently ||= {}
 369     if !@did_recently[thing] or @did_recently[thing] < Time.now - min_interval
 370       @did_recently[thing] = Time.now
 371       false
 372     else
 373       true
 374     end
 375   end
 376 end
 377
 378 # This is how crunch-job child procs know where the "refresh" trigger file is
 379 ENV["CRUNCH_REFRESH_TRIGGER"] = Rails.configuration.crunch_refresh_trigger
 380
 381 Dispatcher.new.run