my @jobstep_tomerge = ();
my $jobstep_tomerge_level = 0;
my $squeue_checked = 0;
+my $sinfo_checked = 0;
my $latest_refresh = scalar time;
Log (undef, "docker image hash is $docker_hash");
$docker_stream =~ s/^\.//;
my $docker_install_script = qq{
-if $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
+if $docker_bin images -q --no-trunc --all | grep -xF \Q$docker_hash\E >/dev/null; then
exit 0
fi
declare -a exit_codes=("\${PIPESTATUS[@]}")
}
}
+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
{
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 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)/) {
$jobstep[$jobstepidx]->{tempfail} = 1;
return $s;
}
-
sub srun_sync
{
my $srunargs = shift;
if (!$busy || ($latest_refresh + 2 < scalar time)) {
check_refresh_wanted();
check_squeue();
+ check_sinfo();
}
if (!$busy) {
select(undef, undef, undef, 0.1);