Blame view

Scripts/utils/queue.pl 11.8 KB
ec85f8892   bigot benjamin   first commit
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
  #!/usr/bin/perl
  # Copyright 2012  Johns Hopkins University (Author: Daniel Povey).
  # Apache 2.0.
  use File::Basename;
  use Cwd;
  
  # queue.pl has the same functionality as run.pl, except that
  # it runs the job in question on the queue (Sun GridEngine).
  # This version of queue.pl uses the task array functionality
  # of the grid engine.  Note: it's different from the queue.pl
  # in the s4 and earlier scripts.
  
  $qsub_opts = "";
  $sync = 0;
  
  for ($x = 1; $x <= 3; $x++) { # This for-loop is to 
    # allow the JOB=1:n option to be interleaved with the
    # options to qsub.
    while (@ARGV >= 2 && $ARGV[0] =~ m:^-:) {
      $switch = shift @ARGV;
      if ($switch eq "-V") {
        $qsub_opts .= "-V ";
      } else {
        $option = shift @ARGV;
        if ($switch eq "-sync" && $option =~ m/^[yY]/) {
          $sync = 1;
        }
        $qsub_opts .= "$switch $option ";
        if ($switch eq "-pe") { # e.g. -pe smp 5
          $option2 = shift @ARGV;
          $qsub_opts .= "$option2 ";
        }
      }
    }
    if ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+):(\d+)$/) {
      $jobname = $1;
      $jobstart = $2;
      $jobend = $3;
      shift;
      if ($jobstart > $jobend) {
        die "queue.pl: invalid job range $ARGV[0]";
      }
    } elsif ($ARGV[0] =~ m/^([\w_][\w\d_]*)+=(\d+)$/) { # e.g. JOB=1.
      $jobname = $1;
      $jobstart = $2;
      $jobend = $2;
      shift;
    } elsif ($ARGV[0] =~ m/.+\=.*\:.*$/) {
      print STDERR "Warning: suspicious first argument to queue.pl: $ARGV[0]
  ";
    }
  }
  
  
  if (@ARGV < 2) {
    print STDERR
     "Usage: queue.pl [options to qsub] [JOB=1:n] log-file command-line arguments...
  " .
     "e.g.: queue.pl foo.log echo baz
  " .
     " (which will echo \"baz\", with stdout and stderr directed to foo.log)
  " .
     "or: queue.pl -q all.q\@xyz foo.log echo bar \| sed s/bar/baz/ 
  " .
     " (which is an example of using a pipe; you can provide other escaped bash constructs)
  " .
     "or: queue.pl -q all.q\@qyz JOB=1:10 foo.JOB.log echo JOB 
  " .
     " (which illustrates the mechanism to submit parallel jobs; note, you can use 
  " .
     "  another string other than JOB)
  " .
     "Note: if you pass the \"-sync y\" option to qsub, this script will take note
  " .
     "and change its behavior.  Otherwise it uses qstat to work out when the job finished
  ";
    exit 1;
  }
  
  $cwd = getcwd();
  $logfile = shift @ARGV;
  
  if (defined $jobname && $logfile !~ m/$jobname/
      && $jobend > $jobstart) {
    print STDERR "queue.pl: you are trying to run a parallel job but "
      . "you are putting the output into just one log file ($logfile)
  ";
    exit(1);
  }
  
  #
  # Work out the command; quote escaping is done here.
  # Note: the rules for escaping stuff are worked out pretty
  # arbitrarily, based on what we want it to do.  Some things that
  # we pass as arguments to queue.pl, such as "|", we want to be
  # interpreted by bash, so we don't escape them.  Other things,
  # such as archive specifiers like 'ark:gunzip -c foo.gz|', we want
  # to be passed, in quotes, to the Kaldi program.  Our heuristic
  # is that stuff with spaces in should be quoted.  This doesn't
  # always work.
  #
  $cmd = "";
  
  foreach $x (@ARGV) { 
    if ($x =~ m/^\S+$/) { $cmd .= $x . " "; } # If string contains no spaces, take
                                              # as-is.
    elsif ($x =~ m:\":) { $cmd .= "'\''$x'\'' "; } # else if no dbl-quotes, use single
    else { $cmd .= "\"$x\" "; }  # else use double.
  }
  
  #
  # Work out the location of the script file, and open it for writing.
  #
  $dir = dirname($logfile);
  $base = basename($logfile);
  $qdir = "$dir/q";
  $qdir =~ s:/(log|LOG)/*q:/q:; # If qdir ends in .../log/q, make it just .../q.
  $queue_logfile = "$qdir/$base";
  
  if (!-d $dir) { system "mkdir $dir 2>/dev/null"; } # another job may be doing this...
  if (!-d $dir) { die "Cannot make the directory $dir
  "; }
  # make a directory called "q",
  # where we will put the log created by qsub... normally this doesn't contain
  # anything interesting, evertyhing goes to $logfile.
  if (! -d "$qdir") { 
    system "mkdir $qdir 2>/dev/null";
    sleep(5); ## This is to fix an issue we encountered in denominator lattice creation,
    ## where if e.g. the exp/tri2b_denlats/log/15/q directory had just been
    ## created and the job immediately ran, it would die with an error because nfs
    ## had not yet synced.  I'm also decreasing the acdirmin and acdirmax in our
    ## NFS settings to something like 5 seconds.
  } 
  
  if (defined $jobname) { # It's an array job.
    $queue_array_opt = "-t $jobstart:$jobend"; 
    $logfile =~ s/$jobname/\$SGE_TASK_ID/g; # This variable will get 
    # replaced by qsub, in each job, with the job-id.
    $cmd =~ s/$jobname/\$SGE_TASK_ID/g; # same for the command...
    $queue_logfile =~ s/\.?$jobname//; # the log file in the q/ subdirectory
    # is for the queue to put its log, and this doesn't need the task array subscript
    # so we remove it.
  }
  
  # queue_scriptfile is as $queue_logfile [e.g. dir/q/foo.log] but
  # with the suffix .sh.
  $queue_scriptfile = $queue_logfile;
  ($queue_scriptfile =~ s/\.[a-zA-Z]{1,5}$/.sh/) || ($queue_scriptfile .= ".sh");
  if ($queue_scriptfile !~ m:^/:) {
    $queue_scriptfile = $cwd . "/" . $queue_scriptfile; # just in case.
  }
  
  # We'll write to the standard input of "qsub" (the file-handle Q),
  # the job that we want it to execute.
  # Also keep our current PATH around, just in case there was something
  # in it that we need (although we also source ./path.sh)
  
  $syncfile = "$qdir/done.$$";
  
  system("rm $queue_logfile $syncfile 2>/dev/null");
  #
  # Write to the script file, and then close it.
  #
  open(Q, ">$queue_scriptfile") || die "Failed to write to $queue_scriptfile";
  
  print Q "#!/bin/bash
  ";
  print Q "cd $cwd
  ";
  print Q ". ./path.sh
  ";
  print Q "( echo '#' Running on \`hostname\`
  ";
  print Q "  echo '#' Started at \`date\`
  ";
  print Q "  echo -n '# '; cat <<EOF
  ";
  print Q "$cmd
  "; # this is a way of echoing the command into a comment in the log file,
  print Q "EOF
  "; # without having to escape things like "|" and quote characters.
  print Q ") >$logfile
  ";
  print Q " ( $cmd ) 2>>$logfile >>$logfile
  ";
  print Q "ret=\$?
  ";
  print Q "echo '#' Finished at \`date\` with status \$ret >>$logfile
  ";
  print Q "[ \$ret -eq 137 ] && exit 100;
  "; # If process was killed (e.g. oom) it will exit with status 137; 
    # let the script return with status 100 which will put it to E state; more easily rerunnable.
  if (!defined $jobname) { # not an array job
    print Q "touch $syncfile
  "; # so we know it's done.
  } else {
    print Q "touch $syncfile.\$SGE_TASK_ID
  "; # touch a bunch of sync-files.
  }
  print Q "exit \$[\$ret ? 1 : 0]
  "; # avoid status 100 which grid-engine
  print Q "## submitted with:
  ";       # treats specially.
  $qsub_cmd = "qsub -S /bin/bash -v PATH -cwd -j y -o $queue_logfile $qsub_opts $queue_array_opt $queue_scriptfile >>$queue_logfile 2>&1";
  print Q "# $qsub_cmd
  ";
  if (!close(Q)) { # close was not successful... || die "Could not close script file $shfile";
    die "Failed to close the script file (full disk?)";
  }
  
  $ret = system ($qsub_cmd);
  if ($ret != 0) {
    if ($sync && $ret == 256) { # this is the exit status when a job failed (bad exit status)
      if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/*/g; }
      print STDERR "queue.pl: job writing to $logfile failed
  ";
    } else {
      print STDERR "queue.pl: error submitting jobs to queue (return status was $ret)
  ";
      print STDERR `tail $queue_logfile`;
    }
    exit(1);
  }
  
  if (! $sync) { # We're not submitting with -sync y, so we
    # need to wait for the jobs to finish.  We wait for the
    # sync-files we "touched" in the script to exist.
    @syncfiles = ();
    if (!defined $jobname) { # not an array job.
      push @syncfiles, $syncfile;
    } else {
      for ($jobid = $jobstart; $jobid <= $jobend; $jobid++) {
        push @syncfiles, "$syncfile.$jobid";
      }
    }
    # We will need the sge_job_id, to check that job still exists
    $sge_job_id=`grep "Your job" $queue_logfile | awk '{ print \$3 }' | sed 's|\\\..*||'`;
    chomp($sge_job_id);
    $check_sge_job_ctr=1;
    #
    $wait = 0.1;
    foreach $f (@syncfiles) {
      # wait for them to finish one by one.
      while (! -f $f) {
        sleep($wait);
        $wait *= 1.2;
        if ($wait > 3.0) {
          $wait = 3.0; # never wait more than 3 seconds.
          if (rand() > 0.5) {
            system("touch $qdir/.kick");
          } else {
            system("rm $qdir/.kick 2>/dev/null");
          }
          # This seems to kick NFS in the teeth to cause it to refresh the
          # directory.  I've seen cases where it would indefinitely fail to get
          # updated, even though the file exists on the server.
          system("ls $qdir >/dev/null");
        }
  
        # Check that the job exists in SGE. Job can be killed if duration 
        # exceeds some hard limit, or in case of a machine shutdown. 
        if(($check_sge_job_ctr++ % 10) == 0) { # Don't run qstat too often, avoid stress on SGE.
          if ( -f $f ) { next; }; #syncfile appeared, ok
          $ret = system("qstat -j $sge_job_id >/dev/null 2>/dev/null");
          if($ret != 0) {
            # Don't consider immediately missing job as error, first wait some  
            # time to make sure it is not just delayed creation of the syncfile.
            sleep(3);
            if ( -f $f ) { next; }; #syncfile appeared, ok
            sleep(7);
            if ( -f $f ) { next; }; #syncfile appeared, ok
            sleep(20);
            if ( -f $f ) { next; }; #syncfile appeared, ok
            #Otherwise it is an error
            if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/*/g; }
            print STDERR "queue.pl: Error, unfinished job no longer exists, log is in $logfile
  ";
            print STDERR "          Possible reasons: a) Exceeded time limit? -> Use more jobs! b) Shutdown/Frozen machine? -> Run again!
  ";
            exit(1);
          }
        }
      }
    }
    $all_syncfiles = join(" ", @syncfiles);
    system("rm $all_syncfiles 2>/dev/null");
  }
  
  # OK, at this point we are synced; we know the job is done.
  # But we don't know about its exit status.  We'll look at $logfile for this.
  # First work out an array @logfiles of file-locations we need to
  # read (just one, unless it's an array job).
  @logfiles = ();
  if (!defined $jobname) { # not an array job.
    push @logfiles, $logfile;
  } else {
    for ($jobid = $jobstart; $jobid <= $jobend; $jobid++) {
      $l = $logfile; 
      $l =~ s/\$SGE_TASK_ID/$jobid/g;
      push @logfiles, $l;
    }
  }
  
  $num_failed = 0;
  foreach $l (@logfiles) {
    @wait_times = (0.1, 0.2, 0.2, 0.3, 0.5, 0.5, 1.0, 2.0, 5.0, 5.0, 5.0, 10.0, 25.0);
    for ($iter = 0; $iter <= @wait_times; $iter++) {
      $line = `tail -10 $l 2>/dev/null`; # Note: although this line should be the last
      # line of the file, I've seen cases where it was not quite the last line because
      # of delayed output by the process that was running, or processes it had called.
      # so tail -10 gives it a little leeway.
      if ($line =~ m/with status (\d+)/) {
        $status = $1;
        last;
      } else {
        if ($iter < @wait_times) {
          sleep($wait_times[$iter]);
        } else {
          if (! -f $l) {
            print STDERR "Log-file $l does not exist.
  ";
          } else {
            print STDERR "The last line of log-file $l does not seem to indicate the "
              . "return status as expected
  ";
          }
          exit(1);                # Something went wrong with the queue, or the
          # machine it was running on, probably.
        }
      }
    }
    # OK, now we have $status, which is the return-status of
    # the command in the job.
    if ($status != 0) { $num_failed++; }
  }
  if ($num_failed == 0) { exit(0); }
  else { # we failed.
    if (@logfiles == 1) {
      if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/$jobstart/g; }
      print STDERR "queue.pl: job failed with status $status, log is in $logfile
  ";
      if ($logfile =~ m/JOB/) {
        print STDERR "queue.pl: probably you forgot to put JOB=1:\$nj in your script.
  ";
      }
    } else {
      if (defined $jobname) { $logfile =~ s/\$SGE_TASK_ID/*/g; }
      $numjobs = 1 + $jobend - $jobstart;
      print STDERR "queue.pl: $num_failed / $numjobs failed, log is in $logfile
  ";
    }
    exit(1);
  }