testsuite/expect/test39.8

#!/usr/bin/env expect
############################################################################
# Purpose: Test of Slurm functionality
#          Test --gpu-bind options
############################################################################
# Copyright (C) SchedMD LLC.
#
# 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 file_in         "$test_dir/input"
set number_commas   "\[0-9_,\]+"

if {![check_config_select "cons_tres"]} {
	skip "This test is only compatible with select/cons_tres"
}
if {![param_contains [get_config_param "TaskPlugin"] "*cgroup"] || [get_config_param "ConstrainDevices"] ne "yes"} {
	skip "This test is only compatible with TaskPlugin=cgroup,... in slurm.conf and ConstrainDevices=yes in cgroup.conf"
}

set gpu_cnt [get_highest_gres_count 1 "gpu"]
if {$gpu_cnt < 2} {
	skip "This test requires 2 or more GPUs on 1 node of the default partition"
}

set node_name [get_nodes_by_request "--gres=gpu:1 -n1 -t1"]
if { [llength $node_name] != 1 } {
	skip "This test need to be able to submit jobs with at least --gres=gpu:1"
}
set cpus_per_node     [get_node_param $node_name "CPUTot"]
set sockets_per_node  [get_node_param $node_name "Sockets"]
set cpus_per_socket   [expr $cpus_per_node / $sockets_per_node]

log_debug "GPU count is $gpu_cnt"
log_debug "Sockets per node is $sockets_per_node"
log_debug "CPUs per socket is $cpus_per_socket"
log_debug "CPUs per node is $cpus_per_node"

if {$cpus_per_node < 2} {
	skip "This test requires 2 or more CPUs per node in the default partition"
}

#
# Build input script file
#
make_bash_script $file_in "echo HOST:\$SLURMD_NODENAME CUDA_VISIBLE_DEVICES:\$CUDA_VISIBLE_DEVICES
if \[ \$SLURM_PROCID -eq 0 \]; then
   $scontrol -dd show job \$SLURM_JOB_ID
fi
exit 0"

set timeout $max_job_delay

if {$sockets_per_node < 2} {
	log_warn "Tests ${test_id}.0 and ${test_id}.1 require 2 or more sockets per node in the default partition"
} else {
	log_info "TEST ${test_id}.0"
	spawn $srun --cpus-per-gpu=1 --gpus-per-socket=1 --sockets-per-node=2 -n2 --accel-bind=g -J $test_name -t1 $file_in
	expect {
		timeout {
			fail "srun not responding"
		}
		eof {
			wait
		}
	}

	#
	# Test of --gpu-bind=closest
	#
	log_info "TEST ${test_id}.1"
	spawn $srun --cpus-per-gpu=1 --gpus-per-socket=1 --sockets-per-node=2 -n2 --gpu-bind=closest -J $test_name -t1 $file_in
	expect {
		timeout {
			fail "srun not responding"
		}
		eof {
			wait
		}
	}
}

#
# Test of --gpu-bind=map_gpu
# Note that if the task count exceeds the provided map_gpu, the map will be cycled over for additional tasks
#
log_info "TEST ${test_id}.2"
if {$gpu_cnt < 4} {
	set map_gpu "map_gpu:1,0"
	set match_goal 2
	set tasks_per_node 2
} else {
	set map_gpu "map_gpu:1,0,3,2"
	set match_goal 4
	set tasks_per_node 4
}

#
# If $gpu_cnt > $cpus_per_node then assume there is no DefCpusPerGpu configured
#
set matches 0
if {$gpu_cnt > $cpus_per_node} {
	spawn $srun --gpus-per-node=$gpu_cnt --ntasks=$tasks_per_node -N1 -l --gpu-bind=$map_gpu -J $test_name -l -t1 $file_in
} else {
	spawn $srun --cpus-per-gpu=1 --gpus-per-node=$gpu_cnt --ntasks=$tasks_per_node -N1 -l --gpu-bind=$map_gpu -J $test_name -l -t1 $file_in
}
expect {
	-re "($number): HOST:($re_word_str) CUDA_VISIBLE_DEVICES:($number)" {
		if       {$expect_out(1,string) == 0 && $expect_out(3,string) == 0} {
			incr matches
		} elseif {$expect_out(1,string) == 1 && $expect_out(3,string) == 0} {
			incr matches
		} elseif {$expect_out(1,string) == 2 && $expect_out(3,string) == 0} {
			incr matches
		} elseif {$expect_out(1,string) == 3 && $expect_out(3,string) == 0} {
			incr matches
		}
		exp_continue
	}
	timeout {
		fail "srun not responding"
	}
	eof {
		wait
	}
}
subtest {$matches == $match_goal} "Verify --gpu-bind=$map_gpu is respected" "$matches != $match_goal"

#
# Test of --gpu-bind=mask_gpu
# Note that if the task count exceeds the provided mask_gpu, the mask will be cycled over for additional tasks
#
log_info "TEST ${test_id}.3"
if {$gpu_cnt < 4} {
	set mask_gpu "mask_gpu:0x3,0x2"
} else {
	set mask_gpu "mask_gpu:0x3,0x2,0x5,0xF"
}

set matches 0
#
# If $gpu_cnt > $cpus_per_node then assume there is no DefCpusPerGpu configured
#
if {$gpu_cnt > $cpus_per_node} {
	spawn $srun --gpus-per-node=$gpu_cnt --ntasks=$tasks_per_node -N1 -l --gpu-bind=$mask_gpu -J $test_name -l -t1 $file_in
} else {
	spawn $srun --cpus-per-gpu=1 --gpus-per-node=$gpu_cnt --ntasks=$tasks_per_node -N1 -l --gpu-bind=$mask_gpu -J $test_name -l -t1 $file_in
}
expect {
	-re "($number): HOST:($re_word_str) CUDA_VISIBLE_DEVICES:($number_commas)" {
		if       {$expect_out(1,string) == 0 && $expect_out(3,string) eq "0,1"} {
			incr matches
		} elseif {$expect_out(1,string) == 1 && $expect_out(3,string) eq "0"} {
			incr matches
		} elseif {$expect_out(1,string) == 2 && $expect_out(3,string) eq "0,1"} {
			incr matches
		} elseif {$expect_out(1,string) == 3 && $expect_out(3,string) eq "0,1,2,3"} {
			incr matches
		}
		exp_continue
	}
	timeout {
		fail "srun not responding"
	}
	eof {
		wait
	}
}
subtest {$matches == $match_goal} "Verify --gpu-bind=$mask_gpu is respected" "$matches != $match_goal"