def write_archive(path, name):
archive = _otf2.Archive_Open(path, name, _otf2.FILEMODE_WRITE, 1024 * 1024,
4 * 1024 * 1024, _otf2.SUBSTRATE_POSIX,
_otf2.COMPRESSION_NONE)
flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(archive, flush_callbacks, None)
_otf2.Archive_SetSerialCollectiveCallbacks(archive)
write_events(archive)
write_definitions(archive)
write_global_definitions(archive)
_otf2.Archive_Close(archive)
def __init__(self,
archive_path,
archive_name="traces",
file_mode=FileMode.WRITE,
chunk_size_events=1024 * 1024,
chunk_size_definitions=4 * 1024 * 1024,
file_substrate=FileSubstrate.POSIX,
compression=Compression.NONE,
pre_flush=_default_pre_flush,
post_flush=None,
flush_data=None,
collective_callbacks=None,
collective_data=None,
global_comm_context=None,
local_comm_context=None,
memory_callbacks=None,
memory_data=None,
timer_resolution=None,
definitions=None):
self._file_mode = file_mode
self._first_timestamp = None
self._last_timestamp = None
self._event_writers = dict()
self._handle = _otf2.Archive_Open(archive_path, archive_name,
file_mode, chunk_size_events,
chunk_size_definitions,
file_substrate, compression)
# We MUST keep those flush callbacks, they cannot be GC'ed...
self._flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(self._handle, self._flush_callbacks,
flush_data)
if collective_callbacks is None:
_otf2.Archive_SetSerialCollectiveCallbacks(self._handle)
else:
_otf2.Archive_SetCollectiveCallbacks(self._handle,
collective_callbacks,
collective_data,
global_comm_context,
local_comm_context)
if memory_callbacks is not None:
_otf2.Archive_SetMemoryCallbacks(self._handle, memory_callbacks,
memory_data)
_otf2.Archive_OpenDefFiles(self._handle)
_otf2.Archive_OpenEvtFiles(self._handle)
if definitions is None:
definitions = DefinitionRegistry()
# Reserve id 0 for empty string
definitions.strings.create("")
if timer_resolution is None:
if definitions.clock_properties is None:
definitions.clock_properties = ClockProperties(
None, timer_resolution=1)
else:
if definitions.clock_properties is None:
definitions.clock_properties = ClockProperties(
None, timer_resolution=timer_resolution)
else:
definitions.clock_properties.timer_resolution = timer_resolution
self._global_definition_registry = definitions
def test(self):
archive_path = "otf2_AnchorFile_trace/trace"
archive_prefix = "archive"
archive_name = archive_prefix + ".otf2"
# Information stored in the trace. */
chunk_size_events = 1024 * 1024
chunk_size_defs = 4 * 1024 * 1024
substrate = _otf2.SUBSTRATE_POSIX
compression = _otf2.COMPRESSION_NONE
num_locations = 123
num_definitions = 456
machine_name = "TestMachine"
creator = "TestCreator"
description = "TestDescription"
# We will write only two properties. The third one will be deleted,
# that means it won't be written to anchor file. */
invalid_property_names = [
"", # empty
"foo", # no "::" at all
"::foo", # empty first component 1
"::foo::bar", # empty first component 2
"foo::::", # empty middle component 1
"::::foo", # empty middle component 2
"foo::", # empty last component 1
"foo::bar::", # empty last component 2
"foo::ba ::baz", # invalid character 1
"foo::bar::ba ", # invalid character 2
# NOTE: None is not an invalid property name but an invalid argument
]
string_property_names = [
"NamespaceA::Property1", "NamespaceA::Property2",
"NamespaceA::Property3"
]
string_property_values = ["Valid", "OK", "true"]
bool_property_names = [
"NamespaceB::BoolProperty1", "NamespaceB::BoolProperty2"
]
bool_property_values = [True, False]
property_name_to_delete = "Whatever::Property"
arbitrary_property_value = "Whatever::Value"
value_to_delete_property = ""
archive = _otf2.Archive_Open(archive_path, archive_prefix,
_otf2.FILEMODE_WRITE, chunk_size_events,
chunk_size_defs, substrate, compression)
flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(archive, flush_callbacks, None)
# Set machine name, creator and description (optional)
_otf2.Archive_SetMachineName(archive, machine_name)
_otf2.Archive_SetCreator(archive, creator)
_otf2.Archive_SetDescription(archive, description)
_otf2.Archive_SetSerialCollectiveCallbacks(archive)
self.archive_set_numbers(archive, num_locations, num_definitions)
for invalid_property_name in invalid_property_names:
with self.assertRaises(_otf2.Error) as cm:
_otf2.Archive_SetProperty(archive, invalid_property_name,
"value", True)
self.assertEqual(
cm.exception.code, _otf2.ERROR_PROPERTY_NAME_INVALID,
'Incorrect return type for invalid property name: "{}"'.format(
invalid_property_name))
_otf2.Archive_SetProperty(archive, property_name_to_delete,
value_to_delete_property, False)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
arbitrary_property_value, False)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
value_to_delete_property, True)
_otf2.Archive_SetProperty(archive, string_property_names[0],
arbitrary_property_value, False)
_otf2.Archive_SetProperty(archive, string_property_names[1],
string_property_values[1], False)
_otf2.Archive_SetProperty(archive, string_property_names[0],
string_property_values[0], True)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
arbitrary_property_value, False)
_otf2.Archive_SetProperty(archive, string_property_names[2],
string_property_values[2], False)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
value_to_delete_property, True)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
arbitrary_property_value, False)
_otf2.Archive_SetProperty(archive, property_name_to_delete,
value_to_delete_property, True)
_otf2.Archive_SetBoolProperty(archive, bool_property_names[0],
bool_property_values[0], False)
_otf2.Archive_SetBoolProperty(archive, bool_property_names[1],
bool_property_values[1], False)
# Close archive. */
_otf2.Archive_Close(archive)
# Open new archive. */
full_path = os.path.join(archive_path, archive_name)
reader = _otf2.Reader_Open(full_path)
_otf2.Reader_SetSerialCollectiveCallbacks(reader)
version_read = _otf2.Reader_GetVersion(reader)
version_expected = (_otf2.VERSION_MAJOR, _otf2.VERSION_MINOR,
_otf2.VERSION_BUGFIX)
self.assertEqual(version_read, version_expected, "Version mismatch")
num_locations_read = _otf2.Reader_GetNumberOfLocations(reader)
self.assertEqual(num_locations, num_locations_read,
"Number of locations does not match")
num_definitions_read = _otf2.Reader_GetNumberOfGlobalDefinitions(
reader, )
self.assertEqual(num_definitions, num_definitions_read,
"Number of definitions does not match")
(chunk_size_events_read,
chunk_size_defs_read) = _otf2.Reader_GetChunkSize(reader)
self.assertEqual(chunk_size_events, chunk_size_events_read,
"Event chunk size does not match")
self.assertEqual(chunk_size_defs, chunk_size_defs_read,
"Definition chunk size does not match")
substrate_read = _otf2.Reader_GetFileSubstrate(reader)
self.assertEqual(substrate, substrate_read,
"File substrate does not match")
compression_read = _otf2.Reader_GetCompression(reader)
self.assertEqual(compression, compression_read,
"Compression does not match")
machine_name_read = _otf2.Reader_GetMachineName(reader)
self.assertEqual(machine_name, machine_name_read,
"Machine name does not match")
creator_read = _otf2.Reader_GetCreator(reader)
self.assertEqual(creator, creator_read, "Creator does not match")
description_read = _otf2.Reader_GetDescription(reader)
self.assertEqual(description, description_read,
"Description does not match")
# Get additional meta-information (a.k.a. properties) */
property_names_read = _otf2.Reader_GetPropertyNames(reader)
self.assertEqual(len(property_names_read),
len(string_property_names) + len(bool_property_names),
"Number of properties does not match")
for (property_name_read, property_name_written, property_value_written) \
in zip(property_names_read, string_property_names + bool_property_names,
string_property_values + bool_property_values):
# First, this test writes #number_of_string_properties properties of type
# string to the anchor file. Afterwards, #number_of_bool_properties are
# written to the anchor file.
# Property names are store in upper case. If everything works correctly,
# _otf2.Reader_GetProperty will return our property names converted from
# mixed case to upper case. Therefore, we also have to convert property
# names to upper case.
converted_name = property_name_written.upper()
self.assertEqual(property_name_read, converted_name,
"Property name does not match")
if isinstance(property_value_written, str):
property_value_read = _otf2.Reader_GetProperty(
reader, property_name_read)
elif isinstance(property_value_written, bool):
property_value_read = _otf2.Reader_GetBoolProperty(
reader, property_name_read)
else:
self.fail("Invalid written property type")
self.assertEqual(property_value_written, property_value_read,
"Property value does not match")
_otf2.Reader_Close(reader)
def write_trace(self):
archive_path = "%s/%s" % (self.tmp_dirname, self.archive_name)
archive = _otf2.Archive_Open(archive_path, "trace_files",
_otf2.FILEMODE_WRITE, 256 * 1024,
_otf2.UNDEFINED_UINT64,
_otf2.SUBSTRATE_POSIX,
_otf2.COMPRESSION_NONE)
flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(archive,
flushCallbacks=flush_callbacks,
flushData=None)
_otf2.Archive_SetSerialCollectiveCallbacks(archive)
_otf2.Archive_SetDescription(
archive, description="OTF2 trace for integrity test.")
_otf2.Archive_SetCreator(archive, creator="OTF2_Integrity")
_otf2.Archive_SetDefChunkSize(archive, 4 * 1024 * 1024)
global_def_writer = _otf2.Archive_GetGlobalDefWriter(archive)
_otf2.GlobalDefWriter_WriteString(global_def_writer, 0, "Process")
for location in self.locations:
_otf2.GlobalDefWriter_WriteLocation(global_def_writer, location, 0,
_otf2.LOCATION_TYPE_CPU_THREAD,
self.num_events * 6, 0)
_otf2.GlobalDefWriter_WriteString(global_def_writer, 1, "MPI_Send")
_otf2.GlobalDefWriter_WriteString(global_def_writer, 2,
"Send an MPI message")
_otf2.GlobalDefWriter_WriteString(global_def_writer, 3, "MPI_Recv")
_otf2.GlobalDefWriter_WriteString(global_def_writer, 4,
"Receive an MPI message")
_otf2.GlobalDefWriter_WriteRegion(global_def_writer, 0, 1, 1, 2,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, 0, 0, 0)
_otf2.GlobalDefWriter_WriteRegion(global_def_writer, 1, 3, 3, 4,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, 0, 0, 0)
_otf2.Archive_CloseGlobalDefWriter(archive, global_def_writer)
_otf2.Archive_OpenDefFiles(archive)
for location in self.locations:
def_writer = _otf2.Archive_GetDefWriter(archive, location)
_otf2.Archive_CloseDefWriter(archive, def_writer)
_otf2.Archive_CloseDefFiles(archive)
_otf2.Archive_OpenEvtFiles(archive)
for _ in range(self.num_events):
for location in self.locations:
evt_writer = _otf2.Archive_GetEvtWriter(archive, location)
ts = get_time()
_otf2.EvtWriter_Enter(evt_writer, None, ts, 0)
_otf2.EvtWriter_MpiSend(evt_writer, None, ts, location, 0, 42,
1024)
_otf2.EvtWriter_Leave(evt_writer, None, get_time(), 0)
_otf2.EvtWriter_Enter(evt_writer, None, get_time(), 1)
ts = get_time()
_otf2.EvtWriter_MpiRecv(evt_writer, None, ts, location, 0, 42,
1024)
_otf2.EvtWriter_Leave(evt_writer, None, ts, 1)
for location in self.locations:
evt_writer = _otf2.Archive_GetEvtWriter(archive, location)
_otf2.Archive_CloseEvtWriter(archive, evt_writer)
_otf2.Archive_CloseEvtFiles(archive)
_otf2.Archive_Close(archive)
def main():
archive_name = "TestTrace"
# Array used to store metric values */
metric_values = (_otf2.MetricValue * NUM_OF_METRICS)()
# If you add new metrics please adapt also the array of metric types
# and the routine to generate random metric values. */
# Array of metric types */
metric_types = [
_otf2.TYPE_UINT64, _otf2.TYPE_DOUBLE, _otf2.TYPE_INT64,
_otf2.TYPE_DOUBLE, _otf2.TYPE_INT64
]
# Initialize accumulated metric */
metric_values[0].unsigned_int = 0
metric_values[1].floating_point = 12.5
metric_values[2].signed_int = 1
metric_values[3].floating_point = 10.0
metric_values[4].unsigned_int = 1
# Distribute metric members to its classes */
metric_members_of_class_1 = [METRIC_A, METRIC_B]
metric_members_of_class_2 = [METRIC_C]
metric_members_of_class_3 = [METRIC_D]
metric_members_of_class_4 = [METRIC_E]
# Array of offsets needed to get correct type or values of a metric in corresponding arrays. */
metric_classes_offsets = [0, 2, 3, 4, 4]
# The collection of metric classes of instances for which we write metric events */
class MetricWriteParameter:
def __init__(self, id_, offset, count, record_for_location):
self.id = id_
self.metric_types = metric_types[offset:offset + count]
self.metric_values = metric_values[offset:offset + count]
self.record_for_location = record_for_location
# FIXME missing class 4
metric_write_parameters = [
MetricWriteParameter(METRIC_CLASS_1, metric_classes_offsets[0],
len(metric_members_of_class_1),
record_for_all_locations),
MetricWriteParameter(METRIC_CLASS_2, metric_classes_offsets[1],
len(metric_members_of_class_2),
record_for_all_locations),
MetricWriteParameter(METRIC_CLASS_3, metric_classes_offsets[2],
len(metric_members_of_class_3),
record_for_all_locations),
MetricWriteParameter(METRIC_INSTANCE_1, metric_classes_offsets[3],
len(metric_members_of_class_4),
record_for_location_0)
]
def create_new_metric_values():
metric_values[0].unsigned_int += 1
metric_values[1].floating_point += 12.5
metric_values[2].unsigned_int = -metric_values[2].unsigned_int
metric_values[3].floating_point += 10.0
metric_values[4].unsigned_int += 1
def write_metrics(new, metric_time, metric_rank, metric_thread):
if new:
create_new_metric_values()
for metric_write_parameter in metric_write_parameters:
if metric_write_parameter.record_for_location(
metric_rank, metric_thread):
_otf2.EvtWriter_Metric(evt_writer, None, metric_time,
metric_write_parameter.id,
metric_write_parameter.metric_types,
metric_write_parameter.metric_values)
# Calculate number of MPI ranks and OpenMP threads. */
num_ranks = 1
num_threads = 1
if otf2_MPI:
num_ranks = otf2_LOCATIONS
archive_path = "otf2_trace_gen_trace_mpi_{}_{}{}".format(
num_ranks, otf2_EVENTS, "_attr" if otf2_ATTRIBUTES else "")
elif otf2_OPENMP:
num_threads = otf2_LOCATIONS
archive_path = "otf2_trace_gen_trace_omp_{}_{}{}".format(
num_threads, otf2_EVENTS, "_attr" if otf2_ATTRIBUTES else "")
elif otf2_HYBRID:
if (otf2_LOCATIONS % 4) == 0:
num_threads = 4
elif (otf2_LOCATIONS % 3) == 0:
num_threads = 3
elif (otf2_LOCATIONS % 2) == 0:
num_threads = 2
else:
# FIXME shadow
for num_threads in range(5, otf2_LOCATIONS):
if (otf2_LOCATIONS % num_threads) == 0:
break
num_ranks = otf2_LOCATIONS // num_threads
archive_path = "otf2_trace_gen_trace_mpi_omp_{}x{}_{}{}".format(
num_ranks, num_threads, otf2_EVENTS,
"_attr" if otf2_ATTRIBUTES else "")
else:
raise RuntimeError("No paradigm set")
print("\nGenerating an OTF2 test archive")
print("with {} rank(s) and {} thread(s) each ({} locations total)".format(
num_ranks, num_threads, otf2_LOCATIONS))
print("to {}/{}.otf2\n".format(archive_path, archive_name))
if os.path.isdir(archive_path):
for i in itertools.count():
archive_path_backup = "{}_bk_{}".format(archive_path, i)
if not os.path.exists(archive_path_backup):
os.rename(archive_path, archive_path_backup)
break
# Create new archive handle. */
archive = _otf2.Archive_Open(archive_path, archive_name,
_otf2.FILEMODE_WRITE, 1024 * 1024,
4 * 1024 * 1024, _otf2.SUBSTRATE_POSIX,
_otf2.COMPRESSION_NONE)
flush_callbacks = _otf2.FlushCallbacks(pre_flush=pre_flush,
post_flush=post_flush)
_otf2.Archive_SetFlushCallbacks(archive, flush_callbacks, None)
# Set description, and creator. */
_otf2.Archive_SetSerialCollectiveCallbacks(archive)
_otf2.Archive_SetDescription(
archive, "Simple OTF2 trace archive to valid reader components")
_otf2.Archive_SetCreator(archive, "otf2_trace_gen")
# Generate location IDs. Just to have non-consecutive location IDs. */
locations = []
mpi_ranks = []
master_threads = []
for rank in range(num_ranks):
for thread in range(num_threads):
locations.append((rank << 16) + thread)
mpi_ranks.append(rank)
master_threads.append(rank << 16)
_otf2.Archive_OpenEvtFiles(archive)
# Create local event writer. */
# Create an attribute list. */
attributes = None
if otf2_ATTRIBUTES:
attributes = _otf2.AttributeList_New()
t = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(archive, master_threads[rank])
# Enter main. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t, REGION_MAIN)
for event in range(otf2_EVENTS):
t = get_time()
# FIXME - this makes no sense
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# Write enter MPI_Send. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_SEND)
_otf2.EvtWriter_MpiSend(evt_writer, None, t,
(rank + 1) % num_ranks, rank, 0, 1024)
t = get_time()
# Write leave MPI_Send. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_SEND)
# Immediately enter the MPI_Recv */
t = get_time()
# Write enter MPI_Recv. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_RECV)
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiRecv(evt_writer, None, t,
(rank + num_ranks - 1) % num_ranks,
rank, 0, 1024)
# Write enter MPI_Recv. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_RECV)
# MPI Barrier. */
t = get_time()
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_BARRIER)
_otf2.EvtWriter_MpiCollectiveBegin(evt_writer, None, t)
# End all started MPI_Barrier */
t = get_time()
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiCollectiveEnd(evt_writer, None, t,
_otf2.COLLECTIVE_OP_BARRIER,
rank, _otf2.UNDEFINED_UINT32,
0, 0)
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_BARRIER)
# If in OpenMP or Hybrid mode. */
if not otf2_MPI:
t_enter = get_time()
t_leave = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# At this point the master creates a team of several parallel threads */
_otf2.EvtWriter_OmpFork(evt_writer, attributes, t_enter,
num_threads)
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
# Enter OpenMP parallel loop (outer region). */
write_metrics(True, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_PARALLEL)
# Enter OpenMP parallel loop (inner region). */
write_metrics(False, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_LOOP)
# Enter OpenMP implicit barrier. */
write_metrics(False, t_enter, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t_enter,
REGION_OPENMP_IMPLICIT_BARRIER)
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
# Leave OpenMP implicit barrier. */
write_metrics(True, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_IMPLICIT_BARRIER)
# Leave OpenMP parallel loop (inner region). */
write_metrics(False, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_LOOP)
# Leave OpenMP parallel loop (outer region). */
write_metrics(False, t_leave, rank, thread)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t_leave,
REGION_OPENMP_PARALLEL)
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
# All threads have finished execution of statements in parallel region.
# Now they synchronize and terminates. Only the master continues execution. */
t_leave = get_time()
_otf2.EvtWriter_OmpJoin(evt_writer, attributes, t_leave)
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
t = get_time()
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Enter(evt_writer, attributes, t,
REGION_MPI_BARRIER)
_otf2.EvtWriter_MpiCollectiveBegin(evt_writer, None, t)
# End all started MPI_Barrier */
t = get_time()
if not otf2_OPENMP:
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, master_threads[rank])
_otf2.EvtWriter_MpiCollectiveEnd(evt_writer, None, t,
_otf2.COLLECTIVE_OP_BARRIER,
rank, _otf2.UNDEFINED_UINT32,
0, 0)
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t,
REGION_MPI_BARRIER)
t = get_time()
for rank in range(num_ranks):
evt_writer = _otf2.Archive_GetEvtWriter(archive, master_threads[rank])
# Leave main. */
write_metrics(True, t, rank, 0)
add_attributes(attributes)
_otf2.EvtWriter_Leave(evt_writer, attributes, t, REGION_MAIN)
# last timestamp */
t = get_time()
number_of_events_per_location = []
for rank in range(num_ranks):
for thread in range(num_threads):
evt_writer = _otf2.Archive_GetEvtWriter(
archive, locations[num_threads * rank + thread])
n = _otf2.EvtWriter_GetNumberOfEvents(evt_writer)
number_of_events_per_location.append(n)
_otf2.Archive_CloseEvtWriter(archive, evt_writer)
_otf2.Archive_CloseEvtFiles(archive)
# Write the global and local definitions. */
glob_def_writer = _otf2.Archive_GetGlobalDefWriter(archive)
# Write the timerange of the trace */
_otf2.GlobalDefWriter_WriteClockProperties(glob_def_writer, 1, 0, t)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, STRING_EMPTY, "")
string = STRING_EMPTY + 1
# Write paradigm definitions */
if otf2_MPI or otf2_HYBRID:
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI")
_otf2.GlobalDefWriter_WriteParadigm(glob_def_writer,
_otf2.PARADIGM_MPI, string,
_otf2.PARADIGM_CLASS_PROCESS)
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"MPI_COMM_${id}")
value = _otf2.AttributeValue()
value.stringRef = string
string += 1
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_COMM_NAME_TEMPLATE, _otf2.TYPE_STRING,
value)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"MPI_WIN_${id}")
value.stringRef = string
string += 1
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_MPI,
_otf2.PARADIGM_PROPERTY_RMA_WIN_NAME_TEMPLATE, _otf2.TYPE_STRING,
value)
(type_, value) = _otf2.AttributeValue_SetBoolean(_otf2.FALSE)
_otf2.GlobalDefWriter_WriteParadigmProperty(
glob_def_writer, _otf2.PARADIGM_OPENMP,
_otf2.PARADIGM_PROPERTY_RMA_ONLY, type_, value)
if otf2_OPENMP or otf2_HYBRID:
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "OpenMP")
_otf2.GlobalDefWriter_WriteParadigm(
glob_def_writer, _otf2.PARADIGM_OPENMP, string,
_otf2.PARADIGM_CLASS_THREAD_FORK_JOIN)
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"OMP Thread Team ${id}")
value = _otf2.AttributeValue()
value.stringRef = string
string += 1
# FIXME seems like a bad idea to write a hardcoded number 4 here
_otf2.GlobalDefWriter_WriteParadigmProperty(glob_def_writer,
_otf2.PARADIGM_OPENMP, 4,
_otf2.TYPE_STRING, value)
# Write system tree definitions. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Rack 0")
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string + 1, "Rack")
_otf2.GlobalDefWriter_WriteSystemTreeNode(glob_def_writer, 0, string,
string + 1,
_otf2.UNDEFINED_SYSTEM_TREE_NODE)
string += 2
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Blade")
string_blade = string
string += 1
for rank in range(num_ranks // 2):
name_buffer = "Blade #{}".format(rank)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name_buffer)
_otf2.GlobalDefWriter_WriteSystemTreeNode(glob_def_writer, rank + 1,
string, string_blade, 0)
string += 1
# Write location group and location definitions. */
for rank in range(num_ranks):
name_buffer = "Process {}".format(rank)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name_buffer)
_otf2.GlobalDefWriter_WriteLocationGroup(
glob_def_writer, rank, string, _otf2.LOCATION_GROUP_TYPE_PROCESS,
(rank // 2) + 1)
string += 1
for thread in range(num_threads):
name_buffer = "Thread {}.{}".format(rank, thread)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
name_buffer)
_otf2.GlobalDefWriter_WriteLocation(
glob_def_writer, locations[rank * num_threads + thread],
string, _otf2.LOCATION_TYPE_CPU_THREAD,
number_of_events_per_location[rank * num_threads + thread],
rank)
string += 1
# Write region definition for "main". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"main (mangled name)")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"main (demangled name)")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MAIN, string,
string - 1, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for an all-threads function. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"!$omp parallelfor")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_OPENMP_PARALLEL,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_OPENMP_LOOP,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for OpenMP implicit barrier. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"!$omp implicit barrier")
_otf2.GlobalDefWriter_WriteRegion(
glob_def_writer, REGION_OPENMP_IMPLICIT_BARRIER, string, string,
STRING_EMPTY, _otf2.REGION_ROLE_UNKNOWN, _otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0, 0)
string += 1
# Write region definition for "MPI_Send". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Send")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_SEND, string,
string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for "MPI_Recv". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Recv")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_RECV, string,
string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Write region definition for "MPI_Barrier". */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_Barrier")
_otf2.GlobalDefWriter_WriteRegion(glob_def_writer, REGION_MPI_BARRIER,
string, string, STRING_EMPTY,
_otf2.REGION_ROLE_UNKNOWN,
_otf2.PARADIGM_UNKNOWN,
_otf2.REGION_FLAG_NONE, STRING_EMPTY, 0,
0)
string += 1
# Define the list of locations which are mpi ranks */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_LOCATIONS,
STRING_EMPTY,
_otf2.GROUP_TYPE_COMM_LOCATIONS,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
master_threads)
# Define the MPI group for MPI_COMM_WORLD */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_WORLD,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_GROUP,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
mpi_ranks)
# Define the MPI group for COMM_SELF like communicators */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_SELF,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_SELF,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
[])
# Define a group that includes all locations. It will be used by metric instance. */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "All locations")
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_ALL_LOCATIONS,
string, _otf2.GROUP_TYPE_LOCATIONS,
_otf2.PARADIGM_UNKNOWN,
_otf2.GROUP_FLAG_NONE, locations)
string += 1
# Define MPI_COMM_WORD */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"My MPI_COMM_WORLD communicator")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_WORLD,
string, GROUP_MPI_COMM_WORLD,
_otf2.UNDEFINED_COMM)
string += 1
# Define MPI_COMM_SELF */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
"My MPI_COMM_SELF communicator")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_SELF,
string, GROUP_MPI_COMM_SELF,
_otf2.UNDEFINED_COMM)
string += 1
# Write member definition of first metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_A")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "FLOPS")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_A,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_PAPI,
_otf2.METRIC_ACCUMULATED_START,
metric_types[0],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of second metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_B")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "s")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_B,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[1],
_otf2.BASE_DECIMAL, -6, string)
string += 1
# Write member definition of third metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_C")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "apples")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_C,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_RELATIVE_LAST,
metric_types[2],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of fourth metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_D")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "pears")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_D,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[3],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Write member definition of fifth metric */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Metric_E")
string += 1
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "cherries")
_otf2.GlobalDefWriter_WriteMetricMember(glob_def_writer, METRIC_E,
string - 1, STRING_EMPTY,
_otf2.METRIC_TYPE_OTHER,
_otf2.METRIC_ABSOLUTE_POINT,
metric_types[4],
_otf2.BASE_DECIMAL, 0, string)
string += 1
# Define metric classes and instances */
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_1,
metric_members_of_class_1,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_2,
metric_members_of_class_2,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_3,
metric_members_of_class_3,
_otf2.METRIC_SYNCHRONOUS_STRICT,
_otf2.RECORDER_KIND_CPU)
_otf2.GlobalDefWriter_WriteMetricClass(glob_def_writer, METRIC_CLASS_4,
metric_members_of_class_4,
_otf2.METRIC_ASYNCHRONOUS,
_otf2.RECORDER_KIND_ABSTRACT)
# We can define an instance with differing recorder and scope.
# In this example metric values are valid for all locationa
# but are recorded on first location. */
_otf2.GlobalDefWriter_WriteMetricInstance(glob_def_writer,
METRIC_INSTANCE_1,
METRIC_CLASS_4, locations[0],
_otf2.SCOPE_GROUP,
GROUP_ALL_LOCATIONS)
# write attribute defs */
if otf2_ATTRIBUTES:
for a, attribute_def in enumerate(attribute_defs):
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string,
attribute_def.name)
_otf2.GlobalDefWriter_WriteAttribute(glob_def_writer, a, string,
STRING_EMPTY,
attribute_def.type)
string += 1
# create a sub communicator of MPI_COMM_WORLD and embed an hyper cube. */
num_cube_ranks = gpot(num_ranks)
cube_ranks = list(range(num_cube_ranks))
# Define the MPI group for the hyper cube */
_otf2.GlobalDefWriter_WriteGroup(glob_def_writer, GROUP_MPI_COMM_CUBE,
STRING_EMPTY, _otf2.GROUP_TYPE_COMM_GROUP,
_otf2.PARADIGM_MPI, _otf2.GROUP_FLAG_NONE,
cube_ranks)
# Define MPI_COMM_MPI_COMM_CUBE */
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "MPI_COMM_CUBE")
_otf2.GlobalDefWriter_WriteComm(glob_def_writer, MPI_COMM_MPI_COMM_CUBE,
string, GROUP_MPI_COMM_WORLD,
MPI_COMM_MPI_COMM_WORLD)
string += 1
num_cube_dims = 0
cube_dim = 1
while cube_dim < num_cube_ranks:
name = "Cube {}".format(cube_dim)
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, name)
_otf2.GlobalDefWriter_WriteCartDimension(glob_def_writer,
num_cube_dims, string, 2,
_otf2.CART_PERIODIC_FALSE)
string += 1
num_cube_dims += 1
cube_dim <<= 1
cube_dimensions = [
_otf2.CartDimensionRef(cube_dim) for cube_dim in range(num_cube_dims)
]
_otf2.GlobalDefWriter_WriteString(glob_def_writer, string, "Hyper Cube")
_otf2.GlobalDefWriter_WriteCartTopology(glob_def_writer, 0, string,
MPI_COMM_MPI_COMM_CUBE,
cube_dimensions)
string += 1
for rank in range(num_cube_ranks):
cube_coordinates = [
1 if rank & (2**cube_dim) else 0
for cube_dim in range(num_cube_dims)
]
_otf2.GlobalDefWriter_WriteCartCoordinate(glob_def_writer, 0, rank,
cube_coordinates)
_otf2.Archive_CloseGlobalDefWriter(archive, glob_def_writer)
# write local mappings for MPI communicators and metrics */
# write local mappings for metrics, this is an identity map, just to write
# something out
_otf2.Archive_OpenDefFiles(archive)
metric_map = _otf2.IdMap_Create(_otf2.ID_MAP_DENSE, NUM_OF_CLASSES)
for c in range(NUM_OF_CLASSES):
_otf2.IdMap_AddIdPair(metric_map, c, c)
for rank in range(num_ranks):
mpi_comm_map = _otf2.IdMap_Create(_otf2.ID_MAP_SPARSE, 2)
# Each location uses its rank as the communicator id which maps to the global 0 */
_otf2.IdMap_AddIdPair(mpi_comm_map, rank, MPI_COMM_MPI_COMM_WORLD)
_otf2.IdMap_AddIdPair(mpi_comm_map, rank + num_ranks,
MPI_COMM_MPI_COMM_SELF)
for thread in range(num_threads):
# Open a definition writer, so the file is created. */
def_writer = _otf2.Archive_GetDefWriter(
archive, locations[num_threads * rank + thread])
# Write metric mappings to local definitions. */
_otf2.DefWriter_WriteMappingTable(def_writer, _otf2.MAPPING_METRIC,
metric_map)
if otf2_MPI or otf2_HYBRID:
# Write MPI Comm mappings to local definitions. */
_otf2.DefWriter_WriteMappingTable(def_writer,
_otf2.MAPPING_COMM,
mpi_comm_map)
# Write clock offsets to local definitions. */
_otf2.DefWriter_WriteClockOffset(def_writer, 0, 0, 0.0)
_otf2.DefWriter_WriteClockOffset(def_writer, t, 0, 0.0)
_otf2.Archive_CloseDefWriter(archive, def_writer)
_otf2.IdMap_Free(mpi_comm_map)
_otf2.IdMap_Free(metric_map)
_otf2.Archive_CloseDefFiles(archive)
# Finalize. */
_otf2.Archive_Close(archive)
_otf2.AttributeList_Delete(attributes)
print("Done\n")