Merge branch '10004-check-sinfo' closes #10004

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

diff --git a/sdk/cli/bin/crunch-job b/sdk/cli/bin/crunch-job
index 39238b0fc649d400a380e397c8f0520fbc75b476..e0aff312cce8ec4737f0285e5f768d4cb5028b50 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;
 
 
@@ -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
 {
@@ -1906,7 +1938,6 @@ sub freezeunquote
   return $s;
 }
 
-
 sub srun_sync
 {
   my $srunargs = shift;
@@ -1961,6 +1992,7 @@ sub srun_sync
     if (!$busy || ($latest_refresh + 2 < scalar time)) {
       check_refresh_wanted();
       check_squeue();
+      check_sinfo();
     }
     if (!$busy) {
       select(undef, undef, undef, 0.1);