@@ -16,79 +16,76 @@ def to_seconds(str):
1616 datetime .strptime ('00:00:00' , '%H:%M:%S' )).total_seconds ()
1717
1818
19- @rfm .parameterized_test (
20- ['serial' , 'gpu' , 24 , 12 , 1 , 1 ],
21- ['serial' , 'mc' , 72 , 36 , 1 , 1 ],
22- ['openmp' , 'gpu' , 24 , 3 , 1 , 4 ],
23- ['openmp' , 'mc' , 72 , 9 , 1 , 4 ],
24- ['mpi' , 'gpu' , 24 , 4 , 3 , 1 ],
25- ['mpi' , 'mc' , 72 , 12 , 3 , 1 ],
26- ['mpi+openmp' , 'gpu' , 24 , 3 , 2 , 2 ],
27- ['mpi+openmp' , 'mc' , 72 , 6 , 3 , 2 ]
28- )
19+ @rfm .simple_test
2920class GREASYCheck (rfm .RegressionTest ):
30- def __init__ (self , variant , partition , num_greasy_tasks , nworkes_per_node ,
31- nranks_per_worker , ncpus_per_worker ):
32- self .valid_systems = ['daint:' + partition , 'dom:' + partition ]
21+ configuration = parameter ([('serial' , 'gpu' , 24 , 12 , 1 , 1 ),
22+ ('serial' , 'mc' , 72 , 36 , 1 , 1 ),
23+ ('openmp' , 'gpu' , 24 , 3 , 1 , 4 ),
24+ ('openmp' , 'mc' , 72 , 9 , 1 , 4 ),
25+ ('mpi' , 'gpu' , 24 , 4 , 3 , 1 ),
26+ ('mpi' , 'mc' , 72 , 12 , 3 , 1 ),
27+ ('mpi+openmp' , 'gpu' , 24 , 3 , 2 , 2 ),
28+ ('mpi+openmp' , 'mc' , 72 , 6 , 3 , 2 )])
29+ variant = variable (str )
30+ partition = variable (str )
31+ num_greasy_tasks = variable (int )
32+ workers_per_node = variable (int )
33+ ranks_per_worker = variable (int )
34+ cpus_per_worker = variable (int )
35+ valid_prog_environs = ['PrgEnv-gnu' ]
36+ sourcepath = 'tasks_mpi_openmp.c'
37+ build_system = 'SingleSource'
38+ executable = 'tasks_mpi_openmp.x'
39+ tasks_file = variable (str , value = 'tasks.txt' )
40+ greasy_logfile = variable (str , value = 'greasy.log' )
41+ nnodes = variable (int , value = 2 )
3342
34- self .valid_prog_environs = ['PrgEnv-gnu' ]
35- self .sourcepath = 'tasks_mpi_openmp.c'
36- self .build_system = 'SingleSource'
43+ # sleep enough time to distinguish if the files are running in parallel
44+ # or not
45+ sleep_time = variable (int , value = 60 )
46+ use_multithreading = False
47+ modules = ['GREASY' ]
48+ maintainers = ['VH' , 'SK' ]
49+ tags = {'production' }
3750
38- # sleep enough time to distinguish if the files are running in parallel
39- # or not
40- self .sleep_time = 60
51+ @run_after ('init' )
52+ def unpack_configuration_parameter (self ):
53+ self .variant , self .partition = self .configuration [0 :2 ]
54+ self .num_greasy_tasks , self .workers_per_node = self .configuration [2 :4 ]
55+ self .ranks_per_worker , self .cpus_per_worker = self .configuration [4 :6 ]
56+
57+ @run_after ('init' )
58+ def set_valid_systems (self ):
59+ self .valid_systems = [f'daint:{ self .partition } ,
60+ f'dom:{ self .partition } ]
61+
62+ @run_before ('compile' )
63+ def setup_build_system (self ):
4164 self .build_system .cflags = [f'-DSLEEP_TIME={ self .sleep_time :d} ]
42- self .variant = variant
43- if variant == 'openmp' :
65+ if self .variant == 'openmp' :
4466 self .build_system .cflags += ['-fopenmp' ]
45- elif variant == 'mpi' :
67+ elif self . variant == 'mpi' :
4668 self .build_system .cflags += ['-D_MPI' ]
47- elif variant == 'mpi+openmp' :
69+ elif self . variant == 'mpi+openmp' :
4870 self .build_system .cflags += ['-fopenmp' , '-D_MPI' ]
4971
50- self . executable = 'tasks_mpi_openmp.x'
51- self . tasks_file = 'tasks.txt'
72+ @ run_before ( 'run' )
73+ def setup_greasy_run ( self ):
5274 self .executable_opts = [self .tasks_file ]
53- self .greasy_logfile = 'greasy.log'
5475 self .keep_files = [self .tasks_file , self .greasy_logfile ]
55- nnodes = 2
56- self .use_multithreading = False
57- self .num_greasy_tasks = num_greasy_tasks
58- self .nworkes_per_node = nworkes_per_node
59- self .nranks_per_worker = nranks_per_worker
60- self .num_tasks_per_node = nranks_per_worker * nworkes_per_node
61- self .num_tasks = self .num_tasks_per_node * nnodes
62- self .num_cpus_per_task = ncpus_per_worker
63- self .sanity_patterns = self .eval_sanity ()
76+ self .num_tasks_per_node = self .ranks_per_worker * self .workers_per_node
77+ self .num_tasks = self .num_tasks_per_node * self .nnodes
78+ self .num_cpus_per_task = self .cpus_per_worker
6479
65- # Reference value is system agnostic
66- # Adding 10 secs of slowdown per greasy tasks
67- # this is to compensate for whenever the systems are full and srun gets
68- # slightly slower
69- refperf = (
70- (self .sleep_time + 10 )* num_greasy_tasks / nworkes_per_node / nnodes
71- )
72- self .reference = {
73- '*' : {
74- 'time' : (refperf , None , 0.5 , 's' )
75- }
76- }
77- self .perf_patterns = {
78- 'time' : sn .extractsingle (r'Total time: (?P<perf>\S+)' ,
79- self .greasy_logfile ,
80- 'perf' , to_seconds )
81- }
80+ @run_before ('run' )
81+ def set_environment_variables (self ):
8282 # On SLURM there is no need to set OMP_NUM_THREADS if one defines
8383 # num_cpus_per_task, but adding for completeness and portability
8484 self .variables = {
8585 'OMP_NUM_THREADS' : str (self .num_cpus_per_task ),
86- 'GREASY_NWORKERS_PER_NODE' : str (nworkes_per_node ),
86+ 'GREASY_NWORKERS_PER_NODE' : str (self . workers_per_node ),
8787 'GREASY_LOGFILE' : self .greasy_logfile
8888 }
89- self .modules = ['GREASY' ]
90- self .maintainers = ['VH' , 'SK' ]
91- self .tags = {'production' }
9289
9390 @run_before ('run' )
9491 def generate_tasks_file (self ):
@@ -114,7 +111,7 @@ def daint_dom_gpu_specific_workaround(self):
114111 }
115112 }
116113 elif self .current_partition .fullname in ['daint:mc' ]:
117- if self . variant != 'serial' :
114+ if 'serial' not in self . variant :
118115 self .extra_resources = {
119116 'gres' : {
120117 'gres' : 'craynetwork:72'
@@ -133,17 +130,19 @@ def set_launcher(self):
133130 # make calls to srun
134131 self .job .launcher = getlauncher ('local' )()
135132
136- @sn . sanity_function
137- def eval_sanity (self ):
133+ @sanity_function
134+ def assert_success (self ):
138135 output_files = []
139136 output_files = [file for file in os .listdir (self .stagedir )
140137 if file .startswith ('output-' )]
141138 num_greasy_tasks = len (output_files )
142139 failure_msg = (f'Requested { self .num_greasy_tasks }
143140 f'executed only { num_greasy_tasks } )
144- sn .evaluate (sn .assert_eq (num_greasy_tasks , self .num_greasy_tasks ,
145- msg = failure_msg ))
146- num_tasks = sn .getattr (self , 'nranks_per_worker' )
141+ sn .evaluate (
142+ sn .assert_eq (num_greasy_tasks , self .num_greasy_tasks ,
143+ msg = failure_msg )
144+ )
145+ num_tasks = sn .getattr (self , 'ranks_per_worker' )
147146 num_cpus_per_task = sn .getattr (self , 'num_cpus_per_task' )
148147
149148 def tid (match ):
@@ -184,7 +183,7 @@ def num_ranks(match):
184183 lambda x : sn .assert_lt (
185184 rank (x ), num_ranks (x ),
186185 msg = (f'Rank id { rank (x )}
187- f'number of ranks { self .nranks_per_worker }
186+ f'number of ranks { self .ranks_per_worker }
188187 f'in output file' )
189188 ), result
190189 ),
@@ -217,7 +216,7 @@ def num_ranks(match):
217216 lambda x : sn .assert_eq (
218217 num_ranks (x ), num_tasks ,
219218 msg = (f'Number of ranks { num_ranks (x )}
220- f'equal to { self .nranks_per_worker }
219+ f'equal to { self .ranks_per_worker }
221220 f'output file { output_file } )
222221 ), result
223222 )
@@ -234,3 +233,24 @@ def num_ranks(match):
234233 ))
235234
236235 return True
236+
237+ @run_before ('performance' )
238+ def set_reference (self ):
239+ # Reference value is system agnostic
240+ # Adding 10 secs of slowdown per greasy tasks
241+ # this is to compensate for whenever the systems are full and srun gets
242+ # slightly slower
243+ refperf = (
244+ (self .sleep_time + 10 ) * self .num_greasy_tasks /
245+ self .workers_per_node / self .nnodes
246+ )
247+ self .reference = {
248+ '*' : {
249+ 'time' : (refperf , None , 0.5 , 's' )
250+ }
251+ }
252+
253+ @performance_function ('s' )
254+ def time (self ):
255+ return sn .extractsingle (r'Total time: (?P<perf>\S+)' ,
256+ self .greasy_logfile , 'perf' , to_seconds )
0 commit comments