blob: 3da06efd89ef8abfc787b9ff3df490b0ef7599dc [file] [log] [blame]
#!/usr/bin/env expect
############################################################################
# Purpose: Test of Slurm functionality
# Test some valid combinations of srun --gpu 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 skipped ""
set file_in1 "$test_dir/input1"
set file_in2 "$test_dir/input2"
set number_commas "\[0-9_,\]+"
set one_task_pc 0
proc _test_gpus_cnt {cmd expected_gpus_times_tasks expected_job_gpus} {
set match 0
set job_gpus 0
set output [run_command_output $cmd]
if [regexp {JOB_GRES=(\S+)} $output {} gres_string] {
set gres_dict [count_gres $gres_string]
if [dict exists $gres_dict "gpu"] {
set job_gpus [dict get $gres_dict "gpu"]
}
}
foreach {- cuda_string} [regexp -all -inline {CUDA_VISIBLE_DEVICES:([0-9_,]+)} $output] {
incr match [cuda_count $cuda_string]
}
subtest {$match == $expected_gpus_times_tasks} "$cmd tasks*gpus result=$match, expected=$expected_gpus_times_tasks"
subtest {$job_gpus == $expected_job_gpus} "$cmd JOB_GRES result=$job_gpus, expected=$expected_job_gpus"
}
proc test_gpus_cnt {srun_opts expected_gpus_times_tasks expected_job_gpus} {
global file_in1 srun number_commas bin_cat test_name
set expected_failure false
if {$expected_gpus_times_tasks == 0 && $expected_job_gpus == 0} {
set expected_failure true
}
_test_gpus_cnt "$srun $srun_opts -t1 -J $test_name -l $file_in1" $expected_gpus_times_tasks $expected_job_gpus
if {$expected_failure} {
set jobid [submit_job -xfail -subtest "$srun_opts -t1 -J sbatch_$test_name --wrap='$srun -l $file_in1'"]
} else {
set jobid [submit_job -subtest "$srun_opts -t1 -J sbatch_$test_name --wrap='$srun -l $file_in1'"]
if {$jobid != 0} {
if {[wait_for_job -pollinterval 0.1 $jobid "DONE"] != 0} {
fail "Job did not complete within timeout"
}
if {[wait_for_file -pollinterval 0.1 "./slurm-$jobid.out"]} {
fail "No output created within timeout"
}
_test_gpus_cnt "$bin_cat ./slurm-$jobid.out" $expected_gpus_times_tasks $expected_job_gpus
run_command "rm ./slurm-$jobid.out"
} else {
subskip -count 2 "Skipping #GPUs * #NTASKS and #GPUS subtests since job submit failed"
}
}
}
if {![check_config_select "cons_tres"]} {
skip "This test is only compatible with select/cons_tres"
}
set nb_nodes [get_partition_param [default_partition] "TotalNodes"]
if {$nb_nodes > 1} {
set nb_nodes 2
}
set gpu_cnt [get_highest_gres_count $nb_nodes "gpu"]
if {$gpu_cnt < 1} {
skip "This test requires 1 or more GPUs on $nb_nodes nodes of the default partition"
}
# Determine what the selecttype param is
if {[param_contains [get_config_param "SelectTypeParameters"] "CR_ONE_TASK_PER_CORE"]} {
set one_task_pc 1
}
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 threads_per_core [get_node_param $node_name "ThreadsPerCore"]
set cores_per_socket [get_node_param $node_name "CoresPerSocket"]
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 $cpus_per_socket"
if {$gpu_cnt > $cpus_per_node} {
set gpu_cnt $cpus_per_node
}
set tot_gpus $gpu_cnt
if {$nb_nodes > 1} {
incr tot_gpus $gpu_cnt
}
set gpus_per_node $gpu_cnt
if {$gpus_per_node > 1 && $sockets_per_node > 1} {
set gpus_per_socket [expr $gpus_per_node / $sockets_per_node]
} else {
set gpus_per_socket $gpus_per_node
}
#
# Build input script files
#
make_bash_script $file_in1 "echo HOST:\$SLURMD_NODENAME CUDA_VISIBLE_DEVICES:\$CUDA_VISIBLE_DEVICES
if \[ \$SLURM_PROCID -eq 0 \]; then
$scontrol -dd show job \$SLURM_JOB_ID | grep JOB_GRES=
fi
exit 0"
make_bash_script $file_in2 "echo HOST:\$SLURMD_NODENAME CUDA_VISIBLE_DEVICES:\$CUDA_VISIBLE_DEVICES
$scontrol show node \$SLURM_JOB_NODENAME | grep Gres
exit 0"
#
# Test --gpus options using a subset of GPUs actually available on the node
#
log_info "TEST: --gpus option"
set match 0
set timeout $max_job_delay
if {$tot_gpus > 1} {
set use_gpus_per_job [expr $tot_gpus - 1]
} else {
set use_gpus_per_job $tot_gpus
}
# Every node requires at least 1 GPU
if {$use_gpus_per_job < $nb_nodes} {
set nb_nodes $use_gpus_per_job
}
test_gpus_cnt "--cpus-per-gpu=1 --gpus=$use_gpus_per_job --nodes=$nb_nodes" $use_gpus_per_job $use_gpus_per_job
if {$use_gpus_per_job > 2} {
log_info "TEST: --gpus option, part 2"
test_gpus_cnt "--cpus-per-gpu=1 --gpus=$use_gpus_per_job --nodes=$nb_nodes" $use_gpus_per_job $use_gpus_per_job
}
#
# Test --gpus-per-node options using a subset of GPUs actually available on the node
#
log_info "TEST: --gpus-per-node option"
set match 0
if {$gpus_per_node > 1} {
set use_gpus_per_node [expr $gpus_per_node - 1]
} else {
set use_gpus_per_node $gpus_per_node
}
test_gpus_cnt "--cpus-per-gpu=1 --gpus-per-node=$use_gpus_per_node --nodes=$nb_nodes" [expr $use_gpus_per_node * $nb_nodes] [expr $use_gpus_per_node * $nb_nodes]
#
# Test --gpus-per-socket options using a subset of GPUs actually available on the node
#
log_info "TEST: --gpus-per-socket option"
set sockets_with_gpus [get_gpu_socket_count $gpu_cnt $sockets_per_node]
set match 0
set node_list [get_nodes_by_request "--gpus-per-socket=1 --sockets-per-node=$sockets_with_gpus --nodes=$nb_nodes"]
if { [llength $node_list] != $nb_nodes } {
lappend skipped "This test need to be able to submit jobs with at least --gpus-per-socket=1 --sockets-per-node=$sockets_with_gpus --nodes=$node_list"
} else {
spawn $srun --gpus-per-socket=1 --sockets-per-node=$sockets_with_gpus --ntasks-per-socket=1 --exclusive --nodelist=[join $node_list ","] -t1 -J "$test_name" -l $file_in1
expect {
-re "CUDA_VISIBLE_DEVICES:($number_commas)" {
incr match [cuda_count $expect_out(1,string)]
exp_continue
}
timeout {
fail "srun not responding"
}
eof {
wait
}
}
set gpus_per_node [get_gres_count "gpu" [join $node_list ","]]
set expected_gpus 0
dict for {node gpus} $gpus_per_node {
incr expected_gpus $gpus
}
if {$match < $expected_gpus} {
fail "srun --gpus-per-socket failure ($match < $expected_gpus)"
}
}
#
# Test --gpus-per-task options using a subset of GPUs actually available on the node
#
log_info "TEST: --gpus-per-task option"
set match 0
if {$gpu_cnt > 1} {
set use_gpus_per_node [expr $gpu_cnt - 1]
} else {
set use_gpus_per_node $gpu_cnt
}
set total_cores [expr $cores_per_socket * $sockets_per_node]
if {$one_task_pc && $cpus_per_node > $total_cores} {
set ntasks_per_core [expr $cpus_per_node / $total_cores]
set extra_opt "--ntasks-per-core=$ntasks_per_core"
} else {
set extra_opt "-t1"
}
test_gpus_cnt "--cpus-per-gpu=1 --gpus-per-task=1 -N1 --ntasks=$use_gpus_per_node $extra_opt" $use_gpus_per_node $use_gpus_per_node
#
# Test --gpus-per-task option without task count
#
log_info "TEST: --gpus-per-task option, part 2 (implicit task count)"
if {$cpus_per_node > 1} {
set use_tasks_per_node 2
} else {
set use_tasks_per_node 1
}
if {$gpu_cnt > 1} {
set use_gpus_per_node [expr ($gpu_cnt - 1) / $use_tasks_per_node]
} else {
set use_gpus_per_node [expr $gpu_cnt / $use_tasks_per_node]
}
if {$use_gpus_per_node == 0} {
set use_tasks_per_node 1
set use_gpus_per_node $gpu_cnt
}
test_gpus_cnt "--gpus-per-task=1 --ntasks-per-node=$use_tasks_per_node -N $nb_nodes" [expr $use_tasks_per_node * $nb_nodes] [expr $use_tasks_per_node * $nb_nodes]
#
# Test --gpus-per-task option without task count
#
log_info "TEST: --gpus-per-task option, part 3 (implicit task count)"
if {$gpu_cnt > 1} {
set use_gpus_per_node [expr $gpu_cnt - 1]
} else {
set use_gpus_per_node $gpu_cnt
}
test_gpus_cnt "--gpus-per-task=$use_gpus_per_node -N $nb_nodes" [expr $nb_nodes * $use_gpus_per_node] [expr $nb_nodes * $use_gpus_per_node]
if {$sockets_per_node <= $gpu_cnt} {
set target_gpus $sockets_per_node
} else {
set target_gpus $gpu_cnt
}
#
# Test --cpus-per-task with different gpu options (Bug 9937)
#
log_info "TEST: --cpus-per-task option with different gpu options(Bug 9937)"
test_gpus_cnt "--gres=gpu:1 --ntasks-per-core=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--gpus=1 --ntasks-per-core=1 -c $cpus_per_node" 1 1
test_gpus_cnt "-n1 --gpus-per-task=1 --ntasks-per-core=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--sockets-per-node=1 --gpus-per-socket=1 --ntasks-per-core=1 -c $cpus_per_node" $target_gpus $target_gpus
# Because we're doing one sockets worth of cpus and if there are two sockets we'll get two gpus.
test_gpus_cnt "--gpus-per-socket=1 --sockets-per-node=1 --threads-per-core=1 -c [expr $cpus_per_node / $threads_per_core]" $sockets_per_node $target_gpus
if {$threads_per_core == 4} {
log_info "Testing high thread count ThreadsPerCore = 4"
test_gpus_cnt "--gpus-per-socket=1 --sockets-per-node=1 --threads-per-core=2 -c [expr $cpus_per_node / 2]" $sockets_per_node $sockets_per_node
}
# Negative tests - srun is expected to fail
log_info "Negative tests - srun is expected to fail here"
test_gpus_cnt "--sockets-per-node=1 --gpus-per-socket=1 --ntasks-per-core=1 -c [expr $cpus_per_node * 2]" 0 0
test_gpus_cnt "--gres=gpu:1 -c [expr $cpus_per_node * 2]" 0 0
test_gpus_cnt "--gpus=1 -c [expr $cpus_per_node * 2]" 0 0
if {$threads_per_core > 1} {
test_gpus_cnt "--gpus=1 --threads-per-core=1 -c $cpus_per_node" 0 0
} else {
test_gpus_cnt "--gpus=1 --threads-per-core=1 -c [expr $cpus_per_node]" 1 1
}
# End of tests from Bug 9937
test_gpus_cnt "--gres=gpu:1 --ntasks-per-node=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--gres=gpu:1 --ntasks-per-node=1 -c $cpus_per_socket" 1 1
test_gpus_cnt "--gres=gpu:1 --ntasks-per-node=1 -c 1" 1 1
test_gpus_cnt "--gres=gpu:1 --ntasks-per-socket=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--gres=gpu:1 --ntasks-per-socket=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--gpus=1 --ntasks-per-node=1 -c $cpus_per_node" 1 1
test_gpus_cnt "--gpus=1 --ntasks-per-socket=1 -c $cpus_per_node" 1 1
test_gpus_cnt "-n1 --gpus-per-task=1 --ntasks-per-node=1 -c $cpus_per_node" 1 1
test_gpus_cnt "-n1 --gpus-per-task=1 --ntasks-per-socket=1 -c $cpus_per_node" 1 1
# Fatals slurmd w/out -n. Bug 10197
test_gpus_cnt "--sockets-per-node=$sockets_per_node --gpus-per-socket=1 --ntasks-per-core=1 -c 1 -n$sockets_per_node" [expr $sockets_per_node * $sockets_per_node] $sockets_per_node
test_gpus_cnt "--sockets-per-node=1 --gpus-per-socket=1 --ntasks-per-core=1 -c $cpus_per_socket" 1 1
set target_gpus [expr min($sockets_per_node, $gpu_cnt)]
test_gpus_cnt "--sockets-per-node=$sockets_per_node --gpus-per-socket=1 --ntasks-per-core=1 -c $cpus_per_node" $target_gpus $target_gpus
if {[llength $skipped] != 0} {
skip "Some tests skipped:\n[join $skipped \n]"
}