def test_max(self):
l = [1, 2]
dl = make_deferrable(l)
self.assertEqual(2, sn.max(dl))
l.append(3)
self.assertEqual(3, sn.max(dl))
def test_max(self):
l = [1, 2]
dl = sn.defer(l)
self.assertEqual(2, sn.max(dl))
l.append(3)
self.assertEqual(3, sn.max(dl))
def __init__(self):
super().__init__()
self.descr = 'Test Cray LibSci on the GPU (dgemm with libsci alloc)'
self.build_system = 'SingleSource'
self.sourcesdir = None
self.sourcepath = ('$CRAY_LIBSCI_ACC_DIR/examples/examples/c_simple/'
'dgemm_simple.c')
self.sanity_patterns = sn.assert_found(r'(4096\s+){3}', self.stdout)
regex = r'(\s+\d+){3}\s+(?P<gpu_flops>\S+)\s+(?P<cpu_flops>\S+)\s+'
self.perf_patterns = {
'dgemm_gpu':
sn.max(sn.extractall(regex, self.stdout, 'gpu_flops', float)),
'dgemm_cpu':
sn.max(sn.extractall(regex, self.stdout, 'cpu_flops', float)),
}
self.reference = {
'daint:gpu': {
'dgemm_gpu': (2264.0, -0.05, None, 'GFLop/s'),
'dgemm_cpu': (45.0, -0.05, None, 'GFLop/s'),
},
'dom:gpu': {
'dgemm_gpu': (2264.0, -0.05, None, 'GFLop/s'),
'dgemm_cpu': (45.0, -0.05, None, 'GFLop/s'),
},
}
def test_max():
l = [1, 2]
dl = sn.defer(l)
assert 2 == sn.max(dl)
l.append(3)
assert 3 == sn.max(dl)
def set_perf_patterns(self):
'''Set performance patterns.'''
self.perf_patterns = {
'latency': sn.max(sn.extractall(
r'\[\S+\] \[gpu \d+\] Kernel launch latency: '
r'(?P<latency>\S+) us', self.stdout, 'latency', float))
}
def __init__(self):
super().__init__()
self.maintainers = ['JG']
self.valid_systems += ['eiger:mc', 'pilatus:mc']
self.time_limit = '5m'
self.sourcepath = 'eatmemory_mpi.c'
self.tags.add('mem')
self.executable_opts = ['100%']
self.sanity_patterns = sn.assert_found(r'(oom-kill)|(Killed)',
self.stderr)
# {{{ perf
regex = (r'^Eating \d+ MB\/mpi \*\d+mpi = -\d+ MB memory from \/proc\/'
r'meminfo: total: \d+ GB, free: \d+ GB, avail: \d+ GB, using:'
r' (\d+) GB')
self.perf_patterns = {
'max_cn_memory':
sn.getattr(self, 'reference_meminfo'),
'max_allocated_memory':
sn.max(sn.extractall(regex, self.stdout, 1, int)),
}
no_limit = (0, None, None, 'GB')
self.reference = {
'*': {
'max_cn_memory':
no_limit,
'max_allocated_memory':
(sn.getattr(self, 'reference_meminfo'), -0.05, None, 'GB'),
}
}
def __init__(self, **kwargs):
super().__init__('Monch', **kwargs)
self.tags = {'monch_acceptance'}
self.valid_systems = ['monch:compute']
self.valid_prog_environs = ['PrgEnv-gnu']
self.num_tasks = 1
self.num_tasks_per_node = 1
self.num_tasks_per_core = 1
self.num_cpus_per_task = 20
self.num_tasks_per_socket = 10
self.use_multithreading = False
self.cflags = '-O3 -I$EBROOTOPENBLAS/include'
self.ldflags = '-L$EBROOTOPENBLAS/lib -lopenblas -lpthread -lgfortran'
self.variables = {
'OMP_NUM_THREADS': str(self.num_cpus_per_task),
'MV2_ENABLE_AFFINITY': '0'
}
self.perf_patterns = {
'perf': sn.max(
sn.extractall(r'Run\s\d\s+:\s+(?P<gflops>\S+)\s\S+',
self.stdout, "gflops", float)
)
}
self.reference = {
'monch:compute': {
'perf': (350, -0.1, None)
}
}
def __init__(self):
self.valid_systems = ['cannon:local-gpu','cannon:gpu_test','fasse:fasse_gpu','test:gpu']
self.descr = 'GPU burn test'
self.valid_prog_environs = ['gpu']
self.executable_opts = ['-d', '40']
self.build_system = 'Make'
self.build_system.makefile = 'makefile.cuda'
self.executable = './gpu_burn.x'
patt = (r'^\s*\[[^\]]*\]\s*GPU\s+\d+\(\S*\):\s+(?P<perf>\S*)\s+GF\/s'
r'\s+(?P<temp>\S*)\s+Celsius')
self.perf_patterns = {
'perf': sn.min(sn.extractall(patt, self.stdout, 'perf', float)),
'temp': sn.max(sn.extractall(patt, self.stdout, 'temp', float)),
}
self.reference = {
'cannon:local-gpu': {
'perf': (6200, -0.10, None, 'Gflop/s per gpu'),
},
'cannon:gpu_test': {
'perf': (6200, -0.10, None, 'Gflop/s per gpu'),
},
'test:gpu': {
'perf': (4115, None, None, 'Gflop/s per gpu'),
},
'*': {
'perf': (4115, None, None, 'Gflop/s per gpu'),
},
'*': {'temp': (0, None, None, 'degC')}
}
def __init__(self, linkage, **kwargs):
super().__init__('scalapack_performance_compile_run_', linkage,
**kwargs)
# FIXME:
# Currently, this test case is only aimed for the monch acceptance,
# yet it could be interesting to extend it to other systems.
# NB: The test case is very small, but larger cases did not succeed!
self.tags |= {'monch_acceptance'}
self.sourcepath = 'scalapack_performance_compile_run.f'
self.valid_systems = ['monch:compute']
self.valid_prog_environs = ['PrgEnv-gnu']
self.num_tasks = 64
self.num_tasks_per_node = 16
self.sanity_patterns = sn.assert_found(r'Run', self.stdout)
self.perf_patterns = {
'perf':
sn.max(
sn.extractall(r'GFLOPS/s:\s+(?P<gflops>\S+)', self.stdout,
'gflops', float))
}
self.reference = {'monch:compute': {'perf': (24., -0.1, None)}}
def mpip_perf_patterns(obj, reg):
'''More perf_patterns for the tool
.. code-block::
-----------------------------------
@--- MPI Time (seconds) -----------
-----------------------------------
Task AppTime MPITime MPI%
0 8.6 0.121 1.40 <-- min
1 8.6 0.157 1.82
2 8.6 5.92 68.84 <-- max
* 25.8 6.2 24.02 <---
=> NonMPI= AppTime - MPITime
Typical performance reporting:
.. code-block::
* mpip_avg_app_time: 8.6 s (= 25.8/3mpi)
* mpip_avg_mpi_time: 2.07 s (= 6.2/3mpi)
* %mpip_avg_mpi_time: 24.02 %
* %max/%min
* %mpip_avg_non_mpi_time: 75.98 %
'''
# rpt = os.path.join(obj.stagedir, obj.rpt_file_txt)
rpt = sn.extractsingle(r'^mpiP: Storing mpiP output in \[(?P<rpt>.*)\]',
obj.stdout, 'rpt', str)
regex_star = r'^\s+\*\s+(?P<appt>\S+)\s+(?P<mpit>\S+)\s+(?P<pct>\S+)$'
regex_minmax = (r'^\s+(?P<mpirk>\S+)\s+(?P<appt>\S+)\s+(?P<mpit>\S+)\s+'
r'(?P<pct>\S+)$')
if reg == 1:
# mpip_avg_mpi_time
result = sn.round(
sn.extractsingle(regex_star, rpt, 'mpit', float) / obj.num_tasks,
2)
elif reg == 2:
# mpip_avg_app_time
result = sn.round(
sn.extractsingle(regex_star, rpt, 'appt', float) / obj.num_tasks,
2)
elif reg == 3:
# %mpip_avg_mpi_time
result = sn.extractsingle(regex_star, rpt, 'pct', float)
elif reg == 4:
# %nonmpi
mpi_pct = sn.extractsingle(regex_star, rpt, 'pct', float)
result = sn.round(100 - mpi_pct, 2)
elif reg == 5:
# %mpip_avg_mpi_time_max
result = sn.max(sn.extractall(regex_minmax, rpt, 'pct', float))
elif reg == 6:
# %mpip_avg_mpi_time_min
result = sn.min(sn.extractall(regex_minmax, rpt, 'pct', float))
else:
raise ValueError('unknown region id in mpip_perf_patterns')
return result
def set_performance_patterns(self):
self.perf_patterns = {
'average_latency': sn.max(sn.extractall(
r'^\s*\[[^\]]*\]\s* On device \d+, '
r'the chase took on average (\d+) '
r'cycles per node jump.', self.stdout, 1, int)
),
}
def ru_maxrss_rk0(obj):
'''Reports the ``maximum resident set size``
'''
maxrss_rk0 = sn.max(
sn.extractall(
r'^METRIC\s+0\s+.*ru_maxrss\" <2>; UINT64; (?P<rss>\d+)\)',
obj.rpt_otf2, 'rss', int))
return maxrss_rk0
def stress_diff(ostream, ostream_ref):
''' Return the difference between obtained and reference stress tensor components'''
stress = get_stress(ostream)
stress_ref = get_stress(ostream_ref)
return sn.max(
sn.abs(stress_ref[i][j] - stress[i][j]) for i in range(2)
for j in range(2))
def average_D2D_latency(self):
'''Extract the average D2D latency.
The pChase code returns a table with the cummulative latency for all
D2D list traversals, and the last column of this table has the max
values for each device.
'''
return sn.max(
sn.extractall(r'^\s*\[[^\]]*\]\s*GPU\s*\d+\s+(\s*\d+.\s+)+',
self.stdout, 1, int))
def __init__(self):
self.valid_systems = [
'daint:gpu', 'dom:gpu', 'arolla:cn', 'tsa:cn', 'ault:amdv100',
'ault:intelv100', 'ault:amda100', 'ault:amdvega'
]
self.descr = 'GPU burn test'
self.valid_prog_environs = ['PrgEnv-gnu']
self.exclusive_access = True
self.executable_opts = ['-d', '40']
self.build_system = 'Make'
self.executable = './gpu_burn.x'
self.num_tasks = 0
self.num_tasks_per_node = 1
self.sanity_patterns = self.assert_num_tasks()
patt = (r'^\s*\[[^\]]*\]\s*GPU\s+\d+\(\S*\):\s+(?P<perf>\S*)\s+GF\/s'
r'\s+(?P<temp>\S*)\s+Celsius')
self.perf_patterns = {
'perf': sn.min(sn.extractall(patt, self.stdout, 'perf', float)),
'temp': sn.max(sn.extractall(patt, self.stdout, 'temp', float)),
}
self.reference = {
'dom:gpu': {
'perf': (4115, -0.10, None, 'Gflop/s'),
},
'daint:gpu': {
'perf': (4115, -0.10, None, 'Gflop/s'),
},
'arolla:cn': {
'perf': (5861, -0.10, None, 'Gflop/s'),
},
'tsa:cn': {
'perf': (5861, -0.10, None, 'Gflop/s'),
},
'ault:amda100': {
'perf': (15000, -0.10, None, 'Gflop/s'),
},
'ault:amdv100': {
'perf': (5500, -0.10, None, 'Gflop/s'),
},
'ault:intelv100': {
'perf': (5500, -0.10, None, 'Gflop/s'),
},
'ault:amdvega': {
'perf': (3450, -0.10, None, 'Gflop/s'),
},
'*': {
'temp': (0, None, None, 'degC')
}
}
self.maintainers = ['AJ', 'TM']
self.tags = {'diagnostic', 'benchmark', 'craype'}
def __init__(self):
super().__init__()
self.valid_systems = (
self.single_device_systems + self.multi_device_systems
)
self.perf_patterns = {
'average_latency': sn.max(sn.extractall(
r'^\s*\[[^\]]*\]\s* On device \d+, '
r'the chase took on average (\d+) '
r'cycles per node jump.', self.stdout, 1, int)
),
}
def set_perf_patterns(self):
'''Extract the minimum performance and maximum temperature recorded.
The performance and temperature data are reported in Gflops/s and
deg. Celsius respectively.
'''
patt = (r'^\s*\[[^\]]*\]\s*GPU\s+\d+\(\S*\):\s+(?P<perf>\S*)\s+GF\/s'
r'\s+(?P<temp>\S*)\s+Celsius')
self.perf_patterns = {
'perf': sn.min(sn.extractall(patt, self.stdout, 'perf', float)),
'temp': sn.max(sn.extractall(patt, self.stdout, 'temp', float)),
}
def vtune_momentumAndEnergyIAD(self):
'''
sphexa::sph::computeMomentumAndEnergyIADImpl<...> sqpatch.exe 40.919s
sphexa::sph::computeMomentumAndEnergyIADImpl<...> sqpatch.exe 38.994s
sphexa::sph::computeMomentumAndEnergyIADImpl<...> sqpatch.exe 40.245s
sphexa::sph::computeMomentumAndEnergyIADImpl<...> sqpatch.exe 39.487s
'''
# ^[sphexa::|MPI|[Others].*\s+(?P<sec>\S+)s$'
regex1 = r'^\s+CPU Time: (?P<sec>\S+)s'
result1 = sn.max(sn.extractall(regex1, self.stdout, 'sec', float))
regex2 = r'^sphexa::sph::computeMomentumAndEnergyIADImpl.*\s+(?P<x>\S+)s$'
result2 = sn.max(sn.extractall(regex2, self.stdout, 'x', float))
print("vtune_cput=", result1)
print("vtune_energ=", result2)
print("vtune_cput/24=", result1 / 24)
print("vtune_energ/24=", result2 / 24)
# print("t=", result1/result2)
# print("c=", self.num_tasks)
# print("t=", (result1/result2) / self.num_tasks)
# t= 5.208910219363269 / 24 = 0.2170379258068029
# vtune_momentumAndEnergyIAD: 5.2089 %
return 0
def set_mpip_perf_patterns(self):
'''More perf_patterns for the tool
.. code-block::
-----------------------------------
@--- MPI Time (seconds) -----------
-----------------------------------
Task AppTime MPITime MPI%
0 8.6 0.121 1.40 <-- min
1 8.6 0.157 1.82
2 8.6 5.92 68.84 <-- max
* 25.8 6.2 24.02 <---
=> NonMPI= AppTime - MPITime
Typical performance reporting:
.. code-block::
* mpip_avg_app_time: 8.6 s (= 25.8/3mpi)
* mpip_avg_mpi_time: 2.07 s (= 6.2/3mpi)
* %mpip_avg_mpi_time: 24.02 %
* %mpip_avg_non_mpi_time: 75.98 %
'''
regex_star = r'^\s+\*\s+(?P<appt>\S+)\s+(?P<mpit>\S+)\s+(?P<pct>\S+)$'
app_t = sn.extractsingle(regex_star, self.rpt, 'appt', float)
mpi_t = sn.extractsingle(regex_star, self.rpt, 'mpit', float)
mpi_pct = sn.extractsingle(regex_star, self.rpt, 'pct', float)
nonmpi_pct = sn.round(100 - mpi_pct, 2)
# min/max
regex = (r'^\s+(?P<mpirk>\S+)\s+(?P<appt>\S+)\s+(?P<mpit>\S+)\s+'
r'(?P<pct>\S+)$')
mpi_pct_max = sn.max(sn.extractall(regex, self.rpt, 'pct', float))
mpi_pct_min = sn.min(sn.extractall(regex, self.rpt, 'pct', float))
perf_pattern = {
'mpip_avg_app_time': sn.round(app_t / self.num_tasks, 2),
'mpip_avg_mpi_time': sn.round(mpi_t / self.num_tasks, 2),
'%mpip_avg_mpi_time': mpi_pct,
'%mpip_avg_mpi_time_max': mpi_pct_max,
'%mpip_avg_mpi_time_min': mpi_pct_min,
'%mpip_avg_non_mpi_time': nonmpi_pct,
}
if self.perf_patterns:
self.perf_patterns = {**self.perf_patterns, **perf_pattern}
else:
self.perf_patterns = perf_pattern
def __init__(self, kernel_version):
self.valid_systems = [
'cannon:local-gpu', 'cannon:gpu_test', 'fasse:fasse_gpu',
'test:gpu'
]
self.valid_prog_environs = ['gpu']
self.build_system = 'Make'
self.executable = './kernel_latency.x'
if kernel_version == 'sync':
self.build_system.cppflags = ['-D SYNCKERNEL=1']
else:
self.build_system.cppflags = ['-D SYNCKERNEL=0']
self.perf_patterns = {
'latency':
sn.max(
sn.extractall(
r'\[\S+\] \[gpu \d+\] Kernel launch latency: '
r'(?P<latency>\S+) us', self.stdout, 'latency', float))
}
self.sys_reference = {
'sync': {
'cannon:local-gpu': {
'latency': (6.0, None, 0.10, 'us')
},
'cannon:gpu_test': {
'latency': (4.0, None, 0.10, 'us')
},
'*': {
'latency': (15.1, None, None, 'us')
},
},
'async': {
'cannon:local-gpu': {
'latency': (6.0, None, 0.10, 'us')
},
'cannon:gpu_test': {
'latency': (4.0, None, 0.10, 'us')
},
'*': {
'latency': (2.2, None, None, 'us')
},
},
}
self.reference = self.sys_reference[kernel_version]
def forces_diff(ostream, ostream_ref):
''' Return the difference between obtained and reference atomic forces'''
forces = get_forces(ostream)
forces_ref = get_forces(ostream_ref)
na = 0
for e in forces:
na += 1
na_ref = 0
for e in forces_ref:
na_ref += 1
sn.assert_eq(na, na_ref,
msg='Wrong length of forces array: {0} != {1}').evaluate()
return sn.max(
sn.abs(forces[i][j] - forces_ref[i][j]) for i in range(na)
for j in range(2))
def pw_perf_patterns(obj):
'''Reports hardware counter values from the tool
.. code-block::
collector time time (%) PAPI_REF_CYC PAPI_L2_DCM
--------------------------------------------------------------------------
computeMomentumAndEnergyIAD 0.6816 100.00 1770550470 2438527
^^^^^^^
'''
regex = r'^computeMomentumAndEnergyIAD\s+\S+\s+\S+\s+\S+\s+(?P<hwc>\d+)$'
hwc_min = sn.min(sn.extractall(regex, obj.stderr, 'hwc', int))
hwc_avg = sn.round(sn.avg(sn.extractall(regex, obj.stderr, 'hwc', int)), 1)
hwc_max = sn.max(sn.extractall(regex, obj.stderr, 'hwc', int))
res_d = {
'papiwrap_hwc_min': hwc_min,
'papiwrap_hwc_avg': hwc_avg,
'papiwrap_hwc_max': hwc_max,
}
return res_d
def __init__(self, linkage):
super().__init__(linkage)
self.tags |= {'monch_acceptance'}
self.sourcepath = 'scalapack_performance_compile_run.f'
self.valid_systems = ['monch:compute']
self.valid_prog_environs = ['PrgEnv-gnu']
self.num_tasks = 64
self.num_tasks_per_node = 16
self.sanity_patterns = sn.assert_found(r'Run', self.stdout)
self.perf_patterns = {
'perf': sn.max(
sn.extractall(r'GFLOPS/s:\s+(?P<gflops>\S+)',
self.stdout, 'gflops', float)
)
}
self.reference = {
'monch:compute': {
'perf': (24., -0.1, None)
}
}
def density_SQ_INSTS_SALU(self):
regex = self.set_regex('density')
rpt = os.path.join(self.stagedir,
self.metric_file.replace(".txt", ".csv"))
return sn.round(sn.max(sn.extractall(regex, rpt, 'm4', int)), 0)
def __init__(self, kernel_version):
super().__init__()
# List known partitions here so as to avoid specifying them every time
# with --system
self.valid_systems = ['daint:gpu', 'dom:gpu', 'kesch:cn']
self.num_tasks = 0
self.num_tasks_per_node = 1
self.sourcepath = 'kernel_latency.cu'
self.build_system = 'SingleSource'
self.build_system.cxxflags = ['-std=c++11']
if self.current_system.name in {'dom', 'daint'}:
self.num_gpus_per_node = 1
gpu_arch = '60'
self.modules = ['craype-accel-nvidia60']
self.valid_prog_environs = [
'PrgEnv-cray', 'PrgEnv-pgi', 'PrgEnv-gnu'
]
elif self.current_system.name == 'kesch':
self.num_gpus_per_node = 16
self.valid_prog_environs = ['PrgEnv-cray', 'PrgEnv-pgi']
self.modules = ['craype-accel-nvidia35']
gpu_arch = '37'
else:
# Enable test when running on an unknown system
self.num_gpus_per_node = 1
self.valid_systems = ['*']
self.valid_prog_environs = ['*']
gpu_arch = None
if gpu_arch:
self.build_system.cxxflags += [
'-arch=compute_%s' % gpu_arch,
'-code=sm_%s' % gpu_arch
]
if kernel_version == 'sync':
self.build_system.cppflags = ['-D SYNCKERNEL=1']
else:
self.build_system.cppflags = ['-D SYNCKERNEL=0']
self.sanity_patterns = sn.all([
sn.assert_eq(
sn.count(sn.findall(r'\[\S+\] Found \d+ gpu\(s\)',
self.stdout)), self.num_tasks_assigned),
sn.assert_eq(
sn.count(
sn.findall(
r'\[\S+\] \[gpu \d+\] Kernel launch '
r'latency: \S+ us', self.stdout)),
self.num_tasks_assigned * self.num_gpus_per_node)
])
self.perf_patterns = {
'latency':
sn.max(
sn.extractall(
r'\[\S+\] \[gpu \d+\] Kernel launch latency: '
r'(?P<latency>\S+) us', self.stdout, 'latency', float))
}
self.sys_reference = {
'sync': {
'dom:gpu': {
'latency': (6.6, None, 0.10, 'us')
},
'daint:gpu': {
'latency': (6.6, None, 0.10, 'us')
},
'kesch:cn': {
'latency': (12.0, None, 0.10, 'us')
},
'*': {
'latency': (0.0, None, None, 'us')
}
},
'async': {
'dom:gpu': {
'latency': (2.2, None, 0.10, 'us')
},
'daint:gpu': {
'latency': (2.2, None, 0.10, 'us')
},
'kesch:cn': {
'latency': (5.7, None, 0.10, 'us')
},
'*': {
'latency': (0.0, None, None, 'us')
}
},
}
self.reference = self.sys_reference[kernel_version]
self.maintainers = ['TM']
self.tags = {'benchmark', 'diagnostic'}
def max_gpu_memory(self):
# Node name Usage Max mem Execution time
# ------------ ----------- ------------ --------------
# nid06681 38 % 2749 MiB 00:00:06
regex = r'^\s+nid\S+\s+\d+\s+%\s+(\d+)\s+MiB.*:'
return sn.max(sn.extractall(regex, self.stdout, 1, int))
def vtune_time(self):
'''Vtune creates 1 report per compute node. For example, a 48 mpi tasks job
(= 2 compute nodes when running with 24 c/cn) will create 2 directories:
* rpt.nid00001/rpt.nid00001.vtune
* rpt.nid00002/rpt.nid00002.vtune
Typical output (for each compute node) is:
.. code-block::
Elapsed Time: 14.866s
CPU Time: 319.177s /24 = 13.3
Effective Time: 308.218s /24 = 12.8
Idle: 0s
Poor: 19.725s
Ok: 119.570s
Ideal: 168.922s
Over: 0s
Spin Time: 10.959s /24 = 0.4
MPI Busy Wait Time: 10.795s
Other: 0.164s
Overhead Time: 0s
Total Thread Count: 25
Paused Time: 0s
'''
result_d = {}
# --- ranks per node
if self.num_tasks < self.num_tasks_per_node:
vtune_tasks_per_node = self.num_tasks
else:
vtune_tasks_per_node = self.num_tasks_per_node
# --- Elapsed Time (min, max)
regex = r'.*Elapsed Time: (?P<sec>\S+)s'
result = sn.extractall(regex, self.stdout, 'sec', float)
result_d['elapsed_min'] = sn.round(sn.min(result), 4)
result_d['elapsed_max'] = sn.round(sn.max(result), 4)
# --- CPU Time (max)
regex = r'^\s+CPU Time: (?P<sec>\S+)s'
result = sn.extractall(regex, self.stdout, 'sec', float)
result_d['elapsed_cput'] = sn.round(
sn.max(result) / vtune_tasks_per_node, 4)
# --- CPU Time: Effective Time (max)
regex = r'^\s+Effective Time: (?P<sec>\S+)s'
result = sn.extractall(regex, self.stdout, 'sec', float)
result_d['elapsed_cput_efft'] = sn.round(
sn.max(result) / vtune_tasks_per_node, 4)
# --- CPU Time: Spin Time (max)
regex = r'^\s+Spin Time: (?P<sec>\S+)s'
result = sn.extractall(regex, self.stdout, 'sec', float)
result_d['elapsed_cput_spint'] = sn.round(
sn.max(result) / vtune_tasks_per_node, 4)
# --- CPU Time: Spin Time: MPI Busy Wait (max)
if self.num_tasks > 1:
regex = r'\s+MPI Busy Wait Time: (?P<sec>\S+)s'
result = sn.extractall(regex, self.stdout, 'sec', float)
result_d['elapsed_cput_spint_mpit'] = sn.round(
sn.max(result) / vtune_tasks_per_node, 4)
else:
result_d['elapsed_cput_spint_mpit'] = 0
# TODO:
# 'vtune_momentumAndEnergyIAD':
# sphsintel.vtune_momentumAndEnergyIAD(self),
# '%vtune_srcf_lookupTables': self.vtune_pct_lookupTables,
# '%vtune_srcf_Octree': self.vtune_pct_Octree,
# '%vtune_srcf_momentumAndEnergyIAD':
# self.vtune_pct_momentumAndEnergyIAD,
# '%vtune_srcf_IAD': self.vtune_pct_IAD,
return result_d
def __init__(self):
super().__init__()
self.valid_systems = ['daint:gpu', 'dom:gpu', 'kesch:cn', 'tiger:gpu']
self.descr = 'GPU burn test'
self.valid_prog_environs = ['PrgEnv-gnu']
if self.current_system.name == 'kesch':
self.exclusive_access = True
self.modules = ['craype-accel-nvidia35']
# NOTE: The first option indicates the precision (-d for double)
# while the seconds is the time (in secs) to run the test.
# For multi-gpu nodes, we run the gpu burn test for more
# time to get reliable measurements.
self.executable_opts = ['-d', '40']
self.num_gpus_per_node = 16
gpu_arch = '37'
elif self.current_system.name in {'daint', 'dom', 'tiger'}:
self.modules = ['craype-accel-nvidia60']
self.executable_opts = ['-d', '20']
self.num_gpus_per_node = 1
gpu_arch = '60'
else:
self.num_gpus_per_node = 1
gpu_arch = None
self.sourcepath = 'gpu_burn.cu'
self.build_system = 'SingleSource'
if gpu_arch:
self.build_system.cxxflags = [
'-arch=compute_%s' % gpu_arch,
'-code=sm_%s' % gpu_arch
]
self.build_system.ldflags = ['-lcuda', '-lcublas', '-lnvidia-ml']
self.sanity_patterns = sn.assert_eq(
sn.count(sn.findall('OK', self.stdout)), self.num_tasks_assigned)
patt = r'GPU\s+\d+\(\S*\): (?P<perf>\S*) GF\/s (?P<temp>\S*) Celsius'
self.perf_patterns = {
'perf': sn.min(sn.extractall(patt, self.stdout, 'perf', float)),
'max_temp': sn.max(sn.extractall(patt, self.stdout, 'temp', float))
}
self.reference = {
'dom:gpu': {
'perf': (4115, -0.10, None, 'Gflop/s'),
'max_temp': (0, None, None, 'Celsius')
},
'daint:gpu': {
'perf': (4115, -0.10, None, 'Gflop/s'),
'max_temp': (0, None, None, 'Celsius')
},
'kesch:cn': {
'perf': (950, -0.10, None, 'Gflop/s'),
'max_temp': (0, None, None, 'Celsius')
},
'*': {
'perf': (0, None, None, 'Gflop/s'),
'max_temp': (0, None, None, 'Celsius')
}
}
self.num_tasks = 0
self.num_tasks_per_node = 1
self.maintainers = ['AJ', 'TM']
self.tags = {'diagnostic', 'benchmark', 'craype'}
def max_temp(self, nid=None):
'''Maximum temperature recorded.'''
return sn.max(self._extract_perf_metric('temp', nid))
def speedup(self):
regex = r'^\S+(f32|f64)\s+(\S+) ns\s+'
slowest = sn.max(sn.extractall(regex, self.stdout, 2, float))
fastest = sn.min(sn.extractall(regex, self.stdout, 2, float))
return sn.round(slowest / fastest, 3)