my $job_api_token;
my $no_clear_tmp;
my $resume_stash;
+my $cgroup_root = "/sys/fs/cgroup";
my $docker_bin = "docker.io";
+my $docker_run_args = "";
GetOptions('force-unlock' => \$force_unlock,
'git-dir=s' => \$git_dir,
'job=s' => \$jobspec,
'job-api-token=s' => \$job_api_token,
'no-clear-tmp' => \$no_clear_tmp,
'resume-stash=s' => \$resume_stash,
+ 'cgroup-root=s' => \$cgroup_root,
'docker-bin=s' => \$docker_bin,
+ 'docker-run-args=s' => \$docker_run_args,
);
if (defined $job_api_token) {
# If this job requires a Docker image, install that.
my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem, $dockeruserarg);
if ($docker_locator = $Job->{docker_image_locator}) {
+ Log (undef, "Install docker image $docker_locator");
($docker_stream, $docker_hash) = find_docker_image($docker_locator);
if (!$docker_hash)
{
croak("No Docker image hash found from locator $docker_locator");
}
+ 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_pid == 0)
{
srun (["srun", "--nodelist=" . join(',', @node)],
- ["/bin/sh", "-ec", $docker_install_script]);
+ ["/bin/bash", "-o", "pipefail", "-ec", $docker_install_script]);
exit ($?);
}
while (1)
}
if ($? != 0)
{
- croak("Installing Docker image from $docker_locator exited "
- .exit_status_s($?));
+ Log(undef, "Installing Docker image from $docker_locator exited " . exit_status_s($?));
+ exit(EX_RETRY_UNLOCKED);
}
# Determine whether this version of Docker supports memory+swap limits.
}
srun(["srun", "--nodelist=" . $node[0]],
["/bin/sh", "-ec",
- "a=`$docker_bin run $try_user_arg $docker_hash id --user` && " .
+ "a=`$docker_bin run $docker_run_args $try_user_arg $docker_hash id --user` && " .
" test \$a -ne 0"],
{fork => 1});
if ($? == 0) {
$ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
$ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
$ENV{"HOME"} = $ENV{"TASK_WORK"};
- $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
$ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
$ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
$ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
+ my $keep_mnt = $ENV{"TASK_WORK"}.".keep";
+
$ENV{"GZIP"} = "-n";
my @srunargs = (
my $stdbuf = " stdbuf --output=0 --error=0 ";
- my $arv_file_cache = "--file-cache=" . (256 * 1024 * 1024);
+ my $arv_file_cache = "";
if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
$arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
}
my $command =
- "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
- ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
- ."&& cd $ENV{CRUNCH_TMP} "
+ "if [ -e \Q$ENV{TASK_WORK}\E ]; then rm -rf \Q$ENV{TASK_WORK}\E; fi; "
+ ."mkdir -p \Q$ENV{CRUNCH_TMP}\E \Q$ENV{JOB_WORK}\E \Q$ENV{TASK_WORK}\E \Q$keep_mnt\E "
+ ."&& cd \Q$ENV{CRUNCH_TMP}\E "
# These environment variables get used explicitly later in
# $command. No tool is expected to read these values directly.
.q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
.q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
- $command .= "&& exec arv-mount --by-pdh --crunchstat-interval=10 --allow-other $arv_file_cache $ENV{TASK_KEEPMOUNT} --exec ";
+
+ $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";
+ $ENV{TASK_KEEPMOUNT_TMP} = "$keep_mnt/tmp";
+
if ($docker_hash)
{
my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
- $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
- $command .= "$docker_bin run --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
+ $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
+ $command .= "$docker_bin run $docker_run_args --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
# We only set memory limits if Docker lets us limit both memory and swap.
# Memory limits alone have been supported longer, but subprocesses tend
# to get SIGKILL if they exceed that without any swap limit set.
$command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
$command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
- # Currently, we make arv-mount's mount point appear at /keep
- # inside the container (instead of using the same path as the
- # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
- # crunch scripts and utilities must not rely on this. They must
- # use $TASK_KEEPMOUNT.
+ # Currently, we make the "by_pdh" directory in arv-mount's mount
+ # point appear at /keep inside the container (instead of using
+ # the same path as the host like we do with CRUNCH_SRC and
+ # CRUNCH_INSTALL). However, crunch scripts and utilities must
+ # not rely on this. They must use $TASK_KEEPMOUNT.
$command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
$ENV{TASK_KEEPMOUNT} = "/keep";
+ # Ditto TASK_KEEPMOUNT_TMP, as /keep_tmp.
+ $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT_TMP}:/keep_tmp\E ";
+ $ENV{TASK_KEEPMOUNT_TMP} = "/keep_tmp";
+
# TASK_WORK is almost exactly like a docker data volume: it
# starts out empty, is writable, and persists until no
# containers use it any more. We don't use --volumes-from to
}
} else {
# Non-docker run
- $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
+ $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -poll=10000 ";
$command .= $stdbuf;
$command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
}
check_refresh_wanted();
check_squeue();
update_progress_stats();
- select (undef, undef, undef, 0.1);
}
elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
{
update_progress_stats();
}
+ if (!$gotsome) {
+ select (undef, undef, undef, 0.1);
+ }
$working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
$_->{node}->{hold_count} < 4 } @slot);
if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
if (!defined $task_success) {
# task did not indicate one way or the other --> fail
Log($jobstepid, sprintf(
- "ERROR: Task process exited %d, but never updated its task record to indicate success and record its output.",
+ "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
exit_status_s($childstatus)));
$Jobstep->{'arvados_task'}->{success} = 0;
$Jobstep->{'arvados_task'}->save;
# squeue check interval (15s) this should make the squeue check an
# infrequent event.
my $silent_procs = 0;
- for my $jobstep (values %proc)
+ for my $procinfo (values %proc)
{
+ my $jobstep = $jobstep[$procinfo->{jobstep}];
if ($jobstep->{stderr_at} < $last_squeue_check)
{
$silent_procs++;
return if $silent_procs == 0;
# use killem() on procs whose killtime is reached
- while (my ($pid, $jobstep) = each %proc)
+ while (my ($pid, $procinfo) = each %proc)
{
- if (exists $jobstep->{killtime}
- && $jobstep->{killtime} <= time
+ my $jobstep = $jobstep[$procinfo->{jobstep}];
+ if (exists $procinfo->{killtime}
+ && $procinfo->{killtime} <= time
&& $jobstep->{stderr_at} < $last_squeue_check)
{
my $sincewhen = "";
if ($jobstep->{stderr_at}) {
$sincewhen = " in last " . (time - $jobstep->{stderr_at}) . "s";
}
- Log($jobstep->{jobstep}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
+ Log($procinfo->{jobstep}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
killem ($pid);
}
}
}
# Check for child procs >60s old and not mentioned by squeue.
- while (my ($pid, $jobstep) = each %proc)
+ while (my ($pid, $procinfo) = each %proc)
{
- if ($jobstep->{time} < time - 60
- && $jobstep->{jobstepname}
- && !exists $ok{$jobstep->{jobstepname}}
- && !exists $jobstep->{killtime})
+ if ($procinfo->{time} < time - 60
+ && $procinfo->{jobstepname}
+ && !exists $ok{$procinfo->{jobstepname}}
+ && !exists $procinfo->{killtime})
{
# According to slurm, this task has ended (successfully or not)
# -- but our srun child hasn't exited. First we must wait (30
# terminated, we'll conclude some slurm communication
# error/delay has caused the task to die without notifying srun,
# and we'll kill srun ourselves.
- $jobstep->{killtime} = time + 30;
- Log($jobstep->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
+ $procinfo->{killtime} = time + 30;
+ Log($procinfo->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
}
}
}
foreach my $job (keys %reader)
{
my $buf;
- while (0 < sysread ($reader{$job}, $buf, 8192))
+ if (0 < sysread ($reader{$job}, $buf, 65536))
{
print STDERR $buf if $ENV{CRUNCH_DEBUG};
$jobstep[$job]->{stderr_at} = time;
$jobstep[$job]->{stderr} .= $buf;
+
+ # Consume everything up to the last \n
preprocess_stderr ($job);
+
if (length ($jobstep[$job]->{stderr}) > 16384)
{
- substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
+ # If we get a lot of stderr without a newline, chop off the
+ # front to avoid letting our buffer grow indefinitely.
+ substr ($jobstep[$job]->{stderr},
+ 0, length($jobstep[$job]->{stderr}) - 8192) = "";
}
$gotsome = 1;
}
# whoa.
$main::please_freeze = 1;
}
- elsif ($line =~ /srun: error: Node failure on/) {
+ elsif ($line =~ /srun: error: (Node failure on|Aborting, .*\bio error\b)/) {
my $job_slot_index = $jobstep[$job]->{slotindex};
$slot[$job_slot_index]->{node}->{fail_count}++;
$jobstep[$job]->{tempfail} = 1;
$jobstep[$job]->{tempfail} = 1;
ban_node_by_slot($jobstep[$job]->{slotindex});
}
- elsif ($line =~ /arvados\.errors\.Keep/) {
+ elsif ($line =~ /\bKeep(Read|Write|Request)Error:/) {
$jobstep[$job]->{tempfail} = 1;
}
}
projects / arvados.git / blobdiff