#!/usr/bin/env expect
############################################################################
# Purpose: Test of Slurm functionality
#          Test allocating sub-sets of GRES to job steps
#
# Output:  "TEST: #.#" followed by "SUCCESS" if test was successful, OR
#          "FAILURE: ..." otherwise with an explanation of the failure, OR
#          anything else indicates a failure mode that must be investigated.
############################################################################
# Copyright (C) 2018 SchedMD LLC
# Written by Morris Jette
#
# This file is part of Slurm, a resource management program.
# For details, see <https://slurm.schedmd.com/>.
# Please also read the included file: DISCLAIMER.
#
# Slurm is free software; you can redistribute it and/or modify it under
# the terms of the GNU General Public License as published by the Free
# Software Foundation; either version 2 of the License, or (at your option)
# any later version.
#
# Slurm is distributed in the hope that it will be useful, but WITHOUT ANY
# WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
# FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
# details.
#
# You should have received a copy of the GNU General Public License along
# with Slurm; if not, write to the Free Software Foundation, Inc.,
# 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA.
############################################################################
source ./globals

set test_id        "39.17"
set exit_code      0
set file_in1       "test$test_id.input1"
set file_in2       "test$test_id.input2"
set file_out       "test$test_id.output"
set number_commas  "\[0-9_,\]+"

print_header $test_id

if {[test_cons_tres]} {
	send_user "\nValid configuration, using select/cons_tres\n"
} else {
	send_user "\nWARNING: This test is only compatible with select/cons_tres\n"
	exit 0
}
if {[test_front_end]} {
	send_user "\nWARNING: This test is incompatible with front-end systems\n"
	exit $exit_code
}

set def_part_name [default_partition]
set nb_nodes [get_node_cnt_in_part $def_part_name]
send_user "\nDefault partition node count is $nb_nodes\n"
if {$nb_nodes > 1} {
	set nb_nodes 2
}
set gpu_cnt [get_gpu_count $nb_nodes]
if {$gpu_cnt < 0} {
	send_user "\nFAILURE: Error getting GPU count\n"
	exit 1
}
if {$gpu_cnt < 2} {
	send_user "\nWARNING: This test requires 2 or more GPUs in the default partition\n"
	exit 0
}

set constrain_devices [test_constrain_devices]
if {$constrain_devices} {
	log_info "Devices files are constrained by cgroups\n"
} else {
	log_info "Devices files are NOT constrained by cgroups\n"
}

get_node_config
set cpus_per_node [expr $sockets_per_node * $cpus_per_socket]

send_user "GPUs per node is $gpu_cnt\n"
send_user "Sockets per node is $sockets_per_node\n"
send_user "CPUs per socket is $cpus_per_socket\n"
send_user "CPUs per node is $cpus_per_node\n"

if {$cpus_per_node < 3} {
	send_user "\nWARNING: This test requires 3 or more CPUs per node in the default partition\n"
	exit 0
}

#
# Test --gpus-per-node option by job step
#
send_user "\n\nTEST 1: --gpus-per-node option\n"

make_bash_script $file_in1 "
	$scontrol -dd show job \${SLURM_JOBID}
	$srun -n1 --gpus-per-node=1 -n1 --mem=0 ./$file_in2 &
	$srun -n1 --gpus-per-node=1 -n1 --mem=0 ./$file_in2 &
	$srun -n1 --gpus-per-node=1 -n1 --mem=0 ./$file_in2 &
	wait
	exit 0"

make_bash_script $file_in2 "
	echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
	sleep 3
	if \[ \$SLURM_STEP_ID -eq 2 \]; then
		$squeue -s --name=test$test_id
	fi
	exit 0"

exec $bin_rm -f $file_out
set job_id 0
set timeout $max_job_delay
spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=2 -N1 -n3 -t1 -o $file_out -J "test$test_id" ./$file_in1
expect {
	-re "Submitted batch job ($number)" {
		set job_id $expect_out(1,string)
		exp_continue
	}
	timeout {
		send_user "\nFAILURE: sbatch not responding\n"
		set exit_code 1
	}
	eof {
		wait
	}
}
if {$job_id == 0} {
	send_user "\nFAILURE: job not submitted\n"
	exit 1
}

if {[wait_for_job $job_id "DONE"] != 0} {
	send_user "\nFAILURE: job $job_id did not complete\n"
	cancel_job $job_id
	exit 1
}

if {[wait_for_file $file_out] != 0} {
	send_user "\nFAILURE: no output file\n"
	exit 1
}
set match 0
spawn $bin_cat $file_out
expect {
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)" {
		incr match
		exp_continue
	}
	eof {
		wait
	}
}
if {$match != 3} {
	send_user "\nFAILURE: bad CUDA information about job ($match != 3)\n"
	set exit_code 1
}

log_user 0
set match 0
spawn $bin_grep "step creation temporarily disabled" $file_out
expect {
	-re "step creation temporarily disabled" {
		incr match
		exp_continue
	}
	eof {
		wait
	}
}
log_user 1
if {$match != 1} {
	send_user "\nFAILURE: did not delay step for available GPU\n"
	set exit_code 1
}

#
# Test --gpus-per-task option by job step
# Note that job is using --gpus-per-node, while step is using --gpus-per-task
#
send_user "\n\nTEST 2: --gpus-per-node for job and multiple simultaneous steps using --gpus-per-task option\n"

make_bash_script $file_in1 "
	$scontrol -dd show job \${SLURM_JOBID}
	$srun -n1 --gpus-per-task=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
	$srun -n1 --gpus-per-task=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
	wait
	exit 0"

make_bash_script $file_in2 "
	echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
	sleep 3
	if \[ \$SLURM_STEP_ID -eq 1 \]; then
		$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
	fi
	exit 0"

exec $bin_rm -f $file_out
set job_id 0
set timeout $max_job_delay
spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=2 -N1 -n2 -t1 -o $file_out -J "test$test_id" ./$file_in1
expect {
	-re "Submitted batch job ($number)" {
		set job_id $expect_out(1,string)
		exp_continue
	}
	timeout {
		send_user "\nFAILURE: sbatch not responding\n"
		set exit_code 1
	}
	eof {
		wait
	}
}
if {$job_id == 0} {
	send_user "\nFAILURE: job not submitted\n"
	exit 1
}

if {[wait_for_job $job_id "DONE"] != 0} {
	send_user "\nFAILURE: job $job_id did not complete\n"
	cancel_job $job_id
	exit 1
}

if {[wait_for_file $file_out] != 0} {
	send_user "\nFAILURE: no output file\n"
	exit 1
}
set bad_cuda_cnt 0
set bad_cuda_val 0
set last_cuda_val -1
spawn $bin_cat $file_out
expect {
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)," {
		incr bad_cuda_cnt
		exp_continue
	}
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)" {
		set val $expect_out(2,string)
		if {$constrain_devices} {
			if {$val != 0} {
				log_error "Expected CUDA_VISIBLE_DEVICES:0 with cgroup-constrained devices"
				set exit_code 1
			}
		} elseif {$last_cuda_val == -1} {
			set last_cuda_val $val
		} elseif {$last_cuda_val == $val} {
			incr bad_cuda_val
		}
		exp_continue
	}
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:" {
		incr bad_cuda_cnt
		exp_continue
	}
	eof {
		wait
	}
}
if {$bad_cuda_cnt != 0} {
	send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
	set exit_code 1
} elseif {$bad_cuda_val != 0} {
	send_user "\nFAILURE: Duplicate CUDA values for parallel steps 2 steps\n"
	set exit_code 1
}

#
# Test --gpus (per job or step) option by job step
#
send_user "\n\nTEST 3: --gpus-per-node for job and multiple simultaneous steps using --gpus (per step) option\n"

make_bash_script $file_in1 "
	$scontrol -dd show job \${SLURM_JOBID}
	$srun -n1 --gpus=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
	$srun -n1 --gpus=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
	wait
	exit 0"

make_bash_script $file_in2 "
	echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
	sleep 3
	if \[ \$SLURM_STEP_ID -eq 1 \]; then
		$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
	fi
	exit 0"

exec $bin_rm -f $file_out
set job_id 0
set timeout $max_job_delay
spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=2 -N1 -n2 -t1 -o $file_out -J "test$test_id" ./$file_in1
expect {
	-re "Submitted batch job ($number)" {
		set job_id $expect_out(1,string)
		exp_continue
	}
	timeout {
		send_user "\nFAILURE: sbatch not responding\n"
		set exit_code 1
	}
	eof {
		wait
	}
}
if {$job_id == 0} {
	send_user "\nFAILURE: job not submitted\n"
	exit 1
}

if {[wait_for_job $job_id "DONE"] != 0} {
	send_user "\nFAILURE: job $job_id did not complete\n"
	cancel_job $job_id
	exit 1
}

if {[wait_for_file $file_out] != 0} {
	send_user "\nFAILURE: no output file\n"
	exit 1
}
set bad_cuda_cnt 0
set bad_cuda_val 0
set last_cuda_val -1
spawn $bin_cat $file_out
expect {
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)," {
		incr bad_cuda_cnt
		exp_continue
	}
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)" {
		set val $expect_out(2,string)
		if {$constrain_devices} {
			if {$val != 0} {
				log_error "Expected CUDA_VISIBLE_DEVICES:0 with cgroup-constrained devices"
				set exit_code 1
			}
		} elseif {$last_cuda_val == -1} {
			set last_cuda_val $val
		} elseif {$last_cuda_val == $val} {
			incr bad_cuda_val
		}
		exp_continue
	}
	-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:" {
		incr bad_cuda_cnt
		exp_continue
	}
	eof {
		wait
	}
}
if {$bad_cuda_cnt != 0} {
	send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
	set exit_code 1
} elseif {$bad_cuda_val != 0} {
	send_user "\nFAILURE: Duplicate CUDA values for parallel steps 2 steps\n"
	set exit_code 1
}

#
# Test --gpus (per job or step) option by job step
#
if {$gpu_cnt > 2} {
	send_user "\n\nTEST 4: --gpus-per-node for job and multiple simultaneous steps using --gpus (per step) option\n"

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -n1 --gpus=2 --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n1 --gpus=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 1 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=3 -N1 -n3 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set bad_cuda_cnt 0
	set bad_cuda_val 0
	set last_cuda_val -1
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number)" {
			set val $expect_out(2,string)
			set val2 $expect_out(3,string)
			if {$constrain_devices} {
				if {$val != 0 || $val2 != 1} {
					log_error "Expected CUDA_VISIBLE_DEVICES:0,1 with cgroup-constrained devices"
					set exit_code 1
				}
			} elseif {$last_cuda_val == -1} {
				set last_cuda_val $val
			} elseif {$last_cuda_val == $val} {
				incr bad_cuda_val
			}
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)" {
			set val $expect_out(2,string)
			if {$constrain_devices} {
				if {$val != 0} {
					log_error "Expected CUDA_VISIBLE_DEVICES:0 with cgroup-constrained devices"
					set exit_code 1
				}
			} elseif {$last_cuda_val == -1} {
				set last_cuda_val $val
			} elseif {$last_cuda_val == $val} {
				incr bad_cuda_val
			}
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:" {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
		set exit_code 1
	} elseif {$bad_cuda_val != 0} {
		send_user "\nFAILURE: Duplicate CUDA values for parallel steps 2 steps\n"
		set exit_code 1
	}
}

#
# Test --gpus-per-task option by job step
#
if {$gpu_cnt > 2} {
	send_user "\n\nTEST 5: --gpus-per-node for job and multiple simultaneous steps using --gpus-per-task option\n"

	set job_gpus 3
	set step_gpus 2

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 2 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N1 -n3 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set match 0
	set bad_cuda_cnt 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number)" {
			if {$step_gpus == 3} {
				incr match
			} else {
				incr bad_cuda_cnt
			}
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number)" {
			if {$step_gpus == 2} {
				incr match
			} else {
				incr bad_cuda_cnt
			}
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:" {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 3 steps\n"
		set exit_code 1
	} elseif {$match != 3} {
		send_user "\nFAILURE: Bad CUDA information for $match of 3 steps\n"
		set exit_code 1
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 2} {
		send_user "\nFAILURE: did not delay steps for available GPU\n"
		set exit_code 1
	}
}

#
# Test --gpus option by job step
#
if {$gpu_cnt >= 2 && $nb_nodes >= 2 && $cpus_per_node >= 3} {
	send_user "\n\nTEST 6: --gpus-per-node for job and multiple simultaneous steps using --gpus option\n"

	set job_gpus 2
	set step_gpus 2

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -n2 --gpus=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n2 --gpus=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n2 --gpus=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'HOST:'\$SLURMD_NODENAME 'NODE_ID:'\$SLURM_NODEID 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 2 -a \$SLURM_NODEID -eq 0 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N2 -n6 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set match 0
	set bad_cuda_cnt 0
	array set cuda_val {}
	spawn $bin_cat $file_out
	expect {
		-re "NODE_ID:($number) STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "NODE_ID:($number) STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number)" {
			incr match
			if {$expect_out(1,string) == 0} {
				set cuda_val($expect_out(2,string)) $expect_out(3,string)
			}
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:" {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 6 tasks\n"
		set exit_code 1
	} elseif {$match != 6} {
		send_user "\nFAILURE: Bad CUDA information for $match of 6 tasks\n"
		set exit_code 1
	}
	if {$exit_code == 0} {
		if {$constrain_devices} {
			if {$cuda_val(0) != 0 || $cuda_val(1) != 0 || $cuda_val(2) != 0} {
				log_error "Expected all steps with NODE_ID=0 to have CUDA_VISIBLE_DEVICES:0 with cgroup-constrained devices"
				set exit_code 1
			}
		} elseif {$cuda_val(0) == $cuda_val(1) && $cuda_val(1) == $cuda_val(2)} {
			send_user "\nFAILURE: Duplicated CUDA values for all 3 steps on node 0 of allocation\n"
			set exit_code 1
		}
		if {$cuda_val(0) != $cuda_val(1) && $cuda_val(1) != $cuda_val(2) && $cuda_val(2) != $cuda_val(0)} {
			send_user "\nFAILURE: Unique CUDA values for all 3 steps on node 0 of allocation\n"
			set exit_code 1
		}
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 1} {
		send_user "\nFAILURE: did not delay steps for available GPU\n"
		set exit_code 1
	}
}

#
# Test --gpus option by job step
#
if {$gpu_cnt >= 4 && $nb_nodes >= 2 && $cpus_per_node >= 3} {
	send_user "\n\nTEST 7: --gpus-per-node for job and multiple simultaneous steps using --gpus option\n"

	set job_gpus 4

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -n2 --gpus=6 --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n2 --gpus=7 --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n2 --gpus=8 --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'HOST:'\$SLURMD_NODENAME 'NODE_ID:'\$SLURM_NODEID 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 2 -a \$SLURM_NODEID -eq 0 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N2 -n6 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}

	set match 0
	set bad_cuda_cnt 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 3 tasks\n"
		set exit_code 1
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 2} {
		send_user "\nFAILURE: did not delay steps for available GPU ($match != 2)\n"
		set exit_code 1
	}
}

#
# Test --gpus-per-task option by job step
#
if {$gpu_cnt > 3 && $nb_nodes >= 2} {
	send_user "\n\nTEST 8: --gpus-per-node for job and multiple simultaneous steps using --gpus-per-task option\n"

	set job_gpus 4
	set step_gpus 2

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -n3 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -n3 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES ' '
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 1 -a \$SLURM_PROCID -eq 0 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N2 -n4 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set match2 0
	set match4 0
	set bad_cuda_cnt 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number),($number) " {
			incr match4
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES: " {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number) " {
			incr match2
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
		set exit_code 1
	} elseif {$match2 != 2} {
		send_user "\nFAILURE: Bad CUDA information for $match2 of 2 tasks\n"
		set exit_code 1
	} elseif {$match4 != 4} {
		send_user "\nFAILURE: Bad CUDA information for $match4 of 4 tasks\n"
		set exit_code 1
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 1} {
		send_user "\nFAILURE: did not delay steps for available GPU\n"
		set exit_code 1
	}
}

#
# Test --gpus-per-task option by job step
#
if {$gpu_cnt > 3 && $nb_nodes >= 2} {
	send_user "\n\nTEST 9: --gpus-per-node for job and multiple simultaneous steps using --gpus-per-task option\n"

	set job_gpus 4
	set step_gpus 2

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -N1 -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -N1 -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -N1 -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -N1 -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -N1 -n1 --gpus-per-task=$step_gpus --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES ' '
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 1 -a \$SLURM_PROCID -eq 0 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N2 -n5 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set match2 0
	set bad_cuda_cnt 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number) " {
			incr match2
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES: " {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
		set exit_code 1
	} elseif {$match2 != 5} {
		send_user "\nFAILURE: Bad CUDA information for $match2 of 5 tasks\n"
		set exit_code 1
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 1} {
		send_user "\nFAILURE: did not delay steps for available GPU\n"
		set exit_code 1
	}
}

#
# Test --gpus-per-task option by job step
#
if {$gpu_cnt > 3 && $nb_nodes >= 2} {
	send_user "\n\nTEST 10: --gpus-per-node for job and multiple simultaneous steps using --gpus-per-task option\n"

	set job_gpus 3
	set step_gpus 2

	make_bash_script $file_in1 "
		$scontrol -dd show job \${SLURM_JOBID}
		$srun -N2 -n4 --ntasks-per-node=2 --gpus-per-task=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
		$srun -N2 -n4 --ntasks-per-node=2 --gpus-per-task=1 --gpus-per-node=0 --mem=0 ./$file_in2 &
		wait
		exit 0"

	make_bash_script $file_in2 "
		echo 'STEP_ID:'\$SLURM_STEP_ID 'CUDA_VISIBLE_DEVICES:'\$CUDA_VISIBLE_DEVICES ' '
		sleep 3
		if \[ \$SLURM_STEP_ID -eq 1 -a \$SLURM_PROCID -eq 0 \]; then
			$scontrol show step \$SLURM_JOB_ID.\$SLURM_STEP_ID
		fi
		exit 0"

	exec $bin_rm -f $file_out
	set job_id 0
	set timeout $max_job_delay
	spawn $sbatch --gres=craynetwork:0 --cpus-per-gpu=1 --gpus-per-node=$job_gpus -N2 -n5 -t1 -o $file_out -J "test$test_id" ./$file_in1
	expect {
		-re "Submitted batch job ($number)" {
			set job_id $expect_out(1,string)
			exp_continue
		}
		timeout {
			send_user "\nFAILURE: sbatch not responding\n"
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$job_id == 0} {
		send_user "\nFAILURE: job not submitted\n"
		exit 1
	}

	if {[wait_for_job $job_id "DONE"] != 0} {
		send_user "\nFAILURE: job $job_id did not complete\n"
		cancel_job $job_id
		exit 1
	}

	if {[wait_for_file $file_out] != 0} {
		send_user "\nFAILURE: no output file\n"
		exit 1
	}
	set match2 0
	set bad_cuda_cnt 0
	spawn $bin_cat $file_out
	expect {
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number)," {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:($number),($number) " {
			incr match2
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES:NoDevFiles " {
			incr bad_cuda_cnt
			exp_continue
		}
		-re "STEP_ID:($number) CUDA_VISIBLE_DEVICES: " {
			incr bad_cuda_cnt
			exp_continue
		}
		eof {
			wait
		}
	}
	if {$bad_cuda_cnt != 0} {
		send_user "\nFAILURE: Bad CUDA count for $bad_cuda_cnt of 2 steps\n"
		set exit_code 1
	} elseif {$match2 != 8} {
		send_user "\nFAILURE: Bad CUDA information for $match2 of 8 tasks\n"
		set exit_code 1
	}

	log_user 0
	set match 0
	spawn $bin_grep "step creation temporarily disabled" $file_out
	expect {
		-re "step creation temporarily disabled" {
			incr match
			exp_continue
		}
		eof {
			wait
		}
	}
	log_user 1
	if {$match != 1} {
		send_user "\nFAILURE: did not delay steps for available GPU\n"
		set exit_code 1
	}
}

if {$exit_code == 0} {
	exec $bin_rm -f $file_in1 $file_in2 $file_out
	send_user "\nSUCCESS\n"
}
exit $exit_code