Merge branch '10497-cwl-job-logs' closes #10497

[arvados.git] / sdk / cli / bin / crunch-job

diff --git a/sdk/cli/bin/crunch-job b/sdk/cli/bin/crunch-job
index cdb5729a7bcdba93e8f0eb3af08c261a45bcc36e..40c9cf325cf11f11d077b6783f0b6fd46b47c74d 100755 (executable)
--- a/sdk/cli/bin/crunch-job
+++ b/sdk/cli/bin/crunch-job
@@ -355,6 +355,7 @@ my @jobstep_done = ();
 my @jobstep_tomerge = ();
 my $jobstep_tomerge_level = 0;
 my $squeue_checked = 0;
+my $sinfo_checked = 0;
 my $latest_refresh = scalar time;
 
 
@@ -863,9 +864,9 @@ for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
         ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
         ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP "
         .q{&& declare -a VOLUMES=() }
-        .q{&& if which crunchrunner >/dev/null ; then VOLUMES+=("--volume=$(which crunchrunner):/usr/local/bin/crunchrunner") ; fi }
-        .q{&& if test -f /etc/ssl/certs/ca-certificates.crt ; then VOLUMES+=("--volume=/etc/ssl/certs/ca-certificates.crt:/etc/arvados/ca-certificates.crt") ; }
-        .q{elif test -f /etc/pki/tls/certs/ca-bundle.crt ; then VOLUMES+=("--volume=/etc/pki/tls/certs/ca-bundle.crt:/etc/arvados/ca-certificates.crt") ; fi };
+        .q{&& if which crunchrunner >/dev/null ; then VOLUMES+=("--volume=$(which crunchrunner):/usr/local/bin/crunchrunner:ro") ; fi }
+        .q{&& if test -f /etc/ssl/certs/ca-certificates.crt ; then VOLUMES+=("--volume=/etc/ssl/certs/ca-certificates.crt:/etc/arvados/ca-certificates.crt:ro") ; }
+        .q{elif test -f /etc/pki/tls/certs/ca-bundle.crt ; then VOLUMES+=("--volume=/etc/pki/tls/certs/ca-bundle.crt:/etc/arvados/ca-certificates.crt:ro") ; fi };
 
     $command .= "&& exec arv-mount --read-write --mount-by-pdh=by_pdh --mount-tmp=tmp --crunchstat-interval=10 --allow-other $arv_file_cache \Q$keep_mnt\E --exec ";
     $ENV{TASK_KEEPMOUNT} = "$keep_mnt/by_pdh";
@@ -1401,6 +1402,37 @@ sub check_squeue
   }
 }
 
+sub check_sinfo
+{
+  # If a node fails in a multi-node "srun" call during job setup, the call
+  # may hang instead of exiting with a nonzero code.  This function checks
+  # "sinfo" for the health of the nodes that were allocated and ensures that
+  # they are all still in the "alloc" state.  If a node that is allocated to
+  # this job is not in "alloc" state, then set please_freeze.
+  #
+  # This is only called from srun_sync() for node configuration.  If a
+  # node fails doing actual work, there are other recovery mechanisms.
+
+  # Do not call `sinfo` more than once every 15 seconds.
+  return if $sinfo_checked > time - 15;
+  $sinfo_checked = time;
+
+  # The output format "%t" means output node states.
+  my @sinfo = `sinfo --nodes=\Q$ENV{SLURM_NODELIST}\E --noheader -o "%t"`;
+  if ($? != 0)
+  {
+    Log(undef, "warning: sinfo exit status $? ($!)");
+    return;
+  }
+  chop @sinfo;
+
+  foreach (@sinfo)
+  {
+    if ($_ != "alloc" && $_ != "alloc*") {
+      $main::please_freeze = 1;
+    }
+  }
+}
 
 sub release_allocation
 {
@@ -1477,18 +1509,24 @@ sub preprocess_stderr
     my $line = $1;
     substr $jobstep[$jobstepidx]->{stderr}, 0, 1+length($line), "";
     Log ($jobstepidx, "stderr $line");
-    if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
-      # whoa.
+    if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/i) {
+      # If the allocation is revoked, we can't possibly continue, so mark all
+      # nodes as failed.  This will cause the overall exit code to be
+      # EX_RETRY_UNLOCKED instead of failure so that crunch_dispatch can re-run
+      # this job.
       $main::please_freeze = 1;
+      foreach my $st (@slot) {
+        $st->{node}->{fail_count}++;
+      }
     }
-    elsif ($line =~ /srun: error: (Node failure on|Aborting, .*\bio error\b)/) {
+    elsif ($line =~ /srun: error: .*?\b(Node failure on|Aborting, .*?\bio error\b)/i) {
       $jobstep[$jobstepidx]->{tempfail} = 1;
       if (defined($job_slot_index)) {
         $slot[$job_slot_index]->{node}->{fail_count}++;
         ban_node_by_slot($job_slot_index);
       }
     }
-    elsif ($line =~ /srun: error: (Unable to create job step|.*: Communication connection failure)/) {
+    elsif ($line =~ /srun: error: (Unable to create job step|.*?: Communication connection failure)/i) {
       $jobstep[$jobstepidx]->{tempfail} = 1;
       ban_node_by_slot($job_slot_index) if (defined($job_slot_index));
     }
@@ -1906,7 +1944,6 @@ sub freezeunquote
   return $s;
 }
 
-
 sub srun_sync
 {
   my $srunargs = shift;
@@ -1961,6 +1998,7 @@ sub srun_sync
     if (!$busy || ($latest_refresh + 2 < scalar time)) {
       check_refresh_wanted();
       check_squeue();
+      check_sinfo();
     }
     if (!$busy) {
       select(undef, undef, undef, 0.1);