#!/usr/bin/env expect
############################################################################
# Purpose: Test of Slurm functionality
#          Test --mem-per-gpu option
############################################################################
# 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 exit_code   0
set file_in     "test$test_id.input"

proc run_gpu_per_job { mem_per_gpu } {
	global exit_code file_in number srun test_id

	set mem_size 0
	set srun_pid [spawn $srun --gpus=1 --mem-per-gpu=$mem_per_gpu -J "test$test_id" -t1 ./$file_in]
	expect {
		-re "TRES=cpu=($number),mem=($number)M" {
			set mem_size $expect_out(2,string)
			exp_continue
		}
		timeout {
			log_error "srun not responding"
			slow_kill $srun_pid
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$mem_size != $mem_per_gpu} {
		log_error "srun --mem-per-gpu failure ($mem_size != $mem_per_gpu)"
		set exit_code 1
	}
}

proc run_gpu_per_node { mem_per_gpu } {
	global exit_code file_in number srun test_id

	set mem_size 0
	set srun_pid [spawn $srun --gpus-per-node=1 -N1 --mem-per-gpu=$mem_per_gpu -J "test$test_id" -t1 ./$file_in]
	expect {
		-re "TRES=cpu=($number),mem=($number)M" {
			set mem_size $expect_out(2,string)
			exp_continue
		}
		timeout {
			log_error "srun not responding"
			slow_kill $srun_pid
			set exit_code 1
		}
		eof {
			wait
		}
	}
	if {$mem_size != $mem_per_gpu} {
		log_error "srun --mem-per-gpu failure ($mem_size != $mem_per_gpu)"
		set exit_code 1
	}
}

proc run_gpu_per_task { mem_per_gpu gpu_cnt } {
	global exit_code file_in number srun test_id

	set mem_size 0
	set srun_pid [spawn $srun --gpus-per-task=$gpu_cnt -n1 --mem-per-gpu=$mem_per_gpu -J "test$test_id" -t1 ./$file_in]
	expect {
		-re "TRES=cpu=($number),mem=($number)M" {
			set mem_size $expect_out(2,string)
			exp_continue
		}
		timeout {
			log_error "srun not responding"
			slow_kill $srun_pid
			set exit_code 1
		}
		eof {
			wait
		}
	}
	set mem_target [expr $mem_per_gpu * $gpu_cnt]
	if {$mem_size != $mem_target} {
		log_error "srun --mem-per-gpu failure ($mem_size != $mem_target)"
		set exit_code 1
	}
}

proc run_gpu_check_mem { srun_opts mem_target node_target } {
	global exit_code file_in number srun test_name

	set mem_size 0
	set node_count 0

	set output [run_command_output -fail "$srun $srun_opts -J $test_name -t1 ./$file_in"]
	regexp "NumNodes=($number)"                $output - node_count
	regexp "TRES=cpu=($number),mem=($number)M" $output - - mem_size

	if {$node_count < $node_target} {
		log_error "srun --mem-per-gpu failure, bad node count ($node_count < $node_target)"
		set exit_code 1
	}
	if {$mem_size != $mem_target} {
		log_error "srun $srun_opts failure ($mem_size != $mem_target)"
		set exit_code 1
	}
}

if {![check_config_select "cons_tres"]} {
        skip "This test is only compatible with select/cons_tres"
}
if {![param_contains [get_config_param "SelectTypeParameters"] "*MEMORY"]} {
        skip "This test requires memory allocation management"
}

set nb_nodes  2
set gpu_cnt   [get_highest_gres_count $nb_nodes "gpu"]
if {$gpu_cnt < 2} {
	skip "This test requires 2 or more GPUs on $nb_nodes nodes of the default partition"
}

set nodes [get_nodes_by_request "--gres=gpu:$gpu_cnt -t1 -N $nb_nodes"]
if { [llength $nodes] != $nb_nodes } {
	skip "This test need to be able to submit jobs with at least --gres=gpu:$gpu_cnt to $nb_nodes nodes"
}

# Get the node with the maximum number of GPUs
dict for {node gpus} [get_gres_count "gpu" [join $nodes ,]] {
	if {$gpus >= $gpu_cnt} {
		set node_name $node
		set gpu_cnt   $gpus
	}
}
set node_memory [get_node_param $node_name "RealMemory"]

log_debug "GPU count is $gpu_cnt"
log_debug "Memory Size is $node_memory"
log_debug "Node count used $nb_nodes"

#
# Build input script file
#
exec $bin_rm -f $file_in
make_bash_script $file_in "echo HOST:\$SLURMD_NODENAME CUDA_VISIBLE_DEVICES:\$CUDA_VISIBLE_DEVICES
$scontrol show job \$SLURM_JOB_ID
exit 0"

#
# Run test job with global GPU count
# Increase mem_per_gpu value 10x on each iteration
#
for {set inx 12} {$inx <= $node_memory} {set inx [expr $inx * 10]} {
	run_gpu_per_job $inx
	if {$exit_code != 0} {
		break
	}
}

#
# Run test job with gpus-per-node count
# Increase mem_per_gpu value 10x on each iteration
#
for {set inx 12} {$inx <= $node_memory} {set inx [expr $inx * 10]} {
	run_gpu_per_node $inx
	if {$exit_code != 0} {
		break
	}
}

#
# Run test job with gpus-per-task count and one GPU
# Increase mem_per_gpu value 10x on each iteration
#
for {set inx 12} {$inx <= $node_memory} {set inx [expr $inx * 10]} {
	run_gpu_per_task $inx 1
	if {$exit_code != 0} {
		break
	}
}

#
# Run test job with gpus-per-task count and two GPUs
# Increase mem_per_gpu value 10x on each iteration
#
if {$gpu_cnt > 1} {
	for {set inx 13} {$inx <= [expr $node_memory / 2]} \
	    {set inx [expr $inx * 10]} {
		run_gpu_per_task $inx 2
		if {$exit_code != 0} {
			break
		}
	}
}

#
# Test heterogeneous GPU allocation (gpu_cnt GPUs on one node, 1 GPU on another node)
#
if {$gpu_cnt > 1 && $nb_nodes > 1} {
	set gpu_target [expr $gpu_cnt + 1]
	set mem_spec 13
	set node_target 2
	set mem_target [expr $mem_spec * $gpu_target]
	run_gpu_check_mem "--gpus=$gpu_target --mem-per-gpu=$mem_spec" $mem_target $node_target
}

#
# Run test with --gpus=2 and mem_per_gpu value that pushed job to 2 nodes
#
if {$gpu_cnt > 1 && $nb_nodes > 1} {
	set mem_spec [expr $node_memory / $gpu_cnt + 1]
	set node_target 2
	set mem_target [expr $mem_spec * $gpu_cnt]
	run_gpu_check_mem "--gpus=$gpu_cnt --mem-per-gpu=$mem_spec" $mem_target $node_target
}

log_info "Testing --mem-per-gpu with --exclusive and --gres=gpu:1"
for {set inx 12} {$inx <= [expr $node_memory / $gpu_cnt]} {set inx [expr $inx * 10]} {
	run_gpu_check_mem "--gres=gpu:1 --mem-per-gpu=$inx --exclusive -w $node_name" [expr $gpu_cnt * $inx] 1
	if {$exit_code != 0} {
		break
	}
}

log_info "Testing --mem-per-gpu with --exclusie and --gpus=1"
for {set inx 12} {$inx <= [expr $node_memory / $gpu_cnt]} {set inx [expr $inx * 10]} {
	run_gpu_check_mem "--gpus=1 --mem-per-gpu=$inx --exclusive -w $node_name" [expr $gpu_cnt * $inx] 1
	if {$exit_code != 0} {
		break
	}
}

log_info "Testing --mem-per-gpu with --exclusie and --gpus-per-task=1"
for {set inx 12} {$inx <= [expr $node_memory / $gpu_cnt]} {set inx [expr $inx * 10]} {
	run_gpu_check_mem "--gpus-per-task=1 --ntasks-per-node=1 --mem-per-gpu=$inx --exclusive -w $node_name" [expr $gpu_cnt * $inx] 1
	if {$exit_code != 0} {
		break
	}
}

log_info "Testing --mem-per-gpu with --exclusie and --gpus-per-socket=1"
for {set inx 12} {$inx <= [expr $node_memory / $gpu_cnt]} {set inx [expr $inx * 10]} {
	run_gpu_check_mem "--gpus-per-socket=1 --sockets-per-node=1 --mem-per-gpu=$inx --exclusive -w $node_name" [expr $gpu_cnt * $inx] 1
	if {$exit_code != 0} {
		break
	}
}

if {$exit_code == 0} {
	exec $bin_rm -f $file_in
} else {
	fail "Test failed due to previous errors (\$exit_code = $exit_code)"
}