def main():
## Initilaize the DCGM Engine as manual operation mode. This implies that it's execution is
## controlled by the monitoring agent. The user has to periodically call APIs such as
## dcgmEnginePolicyTrigger and dcgmEngineUpdateAllFields which tells DCGM to wake up and
## perform data collection and operations needed for policy management.
with RunDCGM('127.0.0.1',
dcgm_structs.DCGM_OPERATION_MODE_MANUAL) as handle:
## Create a default group. (Default group is comprised of all the GPUs on the node)
## Let's call the group as "all_gpus_group". The method returns an opaque handle (groupId) to
## identify the newly created group.
groupId = dcgm_agent.dcgmGroupCreate(handle,
dcgm_structs.DCGM_GROUP_DEFAULT,
"all_gpus_group")
## Invoke method to get information on the newly created group
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId)
## Create reference to DCGM status handler which can be used to get the statuses for multiple
## operations on one or more devices present in the group
status_handle = dcgm_agent.dcgmStatusCreate()
## The worker function can be executed as a separate thread or as part of the main thread.
## Executed as a separate thread here
thread = Thread(target=agent_worker_function, args=(handle, groupId))
thread.start()
##########################################
# Any other useful work can be placed here
##########################################
thread.join()
print("Worker thread completed")
## Destroy the group
try:
dcgm_agent.dcgmGroupDestroy(handle, groupId)
except dcgm_structs.DCGMError as e:
print("Failed to remove the test group, error: %s" % e,
file=sys.stderr)
sys.exit(1)
## Destroy the status handle
try:
dcgm_agent.dcgmStatusDestroy(status_handle)
except dcgm_structs.DCGMError as e:
print("Failed to remove status handler, error: %s" % e,
file=sys.stderr)
sys.exit(1)
def helper_verify_power_value_standalone(handle, groupId, expected_power):
"""
Helper Method to verify power value
"""
groupInfo = dcgm_agent.dcgmGroupGetInfo(
handle, groupId, dcgm_structs.c_dcgmGroupInfo_version2)
status_handle = dcgm_agent.dcgmStatusCreate()
config_values = dcgm_agent.dcgmConfigGet(
handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE,
groupInfo.count, status_handle)
assert len(
config_values) > 0, "Failed to get configuration using dcgmConfigGet"
for x in range(0, groupInfo.count):
if (config_values[x].mPowerLimit.val !=
dcgmvalue.DCGM_INT32_NOT_SUPPORTED):
assert config_values[x].mPowerLimit.type == dcgm_structs.DCGM_CONFIG_POWER_CAP_INDIVIDUAL, \
"The power limit type for gpuId %d is incorrect. Returned: %d Expected :%d" \
% (x, config_values[x].mPowerLimit.type, dcgm_structs.DCGM_CONFIG_POWER_CAP_INDIVIDUAL)
assert config_values[x].mPowerLimit.val == expected_power, "The power limit value for gpuID %d is incorrect. Returned: %d Expected: %d" \
% (x, config_values[x].mPowerLimit.val, expected_power)
pass
ret = dcgm_agent.dcgmStatusDestroy(status_handle)
assert (ret == dcgm_structs.DCGM_ST_OK
), "Failed to remove status handler, error: %s" % ret
def helper_verify_config_values_standalone(handle, groupId, expected_power, expected_ecc, \
expected_proc_clock, expected_mem_clock, expected_compute_mode, \
expected_sync_boost, expected_auto_boost):
"""
Helper Method to verify all the values for the current configuration are as expected
"""
groupInfo = dcgm_agent.dcgmGroupGetInfo(
handle, groupId, dcgm_structs.c_dcgmGroupInfo_version2)
status_handle = dcgm_agent.dcgmStatusCreate()
config_values = dcgm_agent.dcgmConfigGet(
handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE,
groupInfo.count, status_handle)
assert len(
config_values) > 0, "Failed to get configuration using dcgmConfigGet"
for x in xrange(0, groupInfo.count):
assert config_values[x].mPowerLimit.type == dcgm_structs.DCGM_CONFIG_POWER_CAP_INDIVIDUAL, \
"The power limit type for gpuId %d is incorrect. Returned: %d Expected :%d" \
% (x, config_values[x].mPowerLimit.type, dcgm_structs.DCGM_CONFIG_POWER_CAP_INDIVIDUAL)
assert config_values[x].mPowerLimit.val == expected_power, "The power limit value for gpuID %d is incorrect. Returned: %d Expected: %d" \
% (x, config_values[x].mPowerLimit.val, expected_power)
assert config_values[x].mPerfState.syncBoost == expected_sync_boost, "The syncboost value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values[x].mPerfState.syncBoost, expected_sync_boost)
assert config_values[x].mPerfState.autoBoost == expected_auto_boost, "The autoboost value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values[x].mPerfState.autoBoost, expected_auto_boost)
assert config_values[x].mPerfState.minVPState.memClk == expected_mem_clock, "The min mem clock value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values.mPerfState.minVPState.memClk , expected_mem_clock)
assert config_values[x].mPerfState.minVPState.procClk == expected_proc_clock, "The min proc clock value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values[x].mPerfState.minVPState.procClk , expected_proc_clock)
assert config_values[x].mComputeMode == expected_compute_mode, "The compute mode value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values[x].mComputeMode, expected_compute_mode)
assert config_values[x].mEccMode == expected_ecc, "The ecc mode value for gpuID %d is incorrect."\
" Returned: %d Expected: %d" \
% (x, config_values[x].mEccMode, expected_ecc)
pass
ret = dcgm_agent.dcgmStatusDestroy(status_handle)
assert (ret == dcgm_structs.DCGM_ST_OK
), "Failed to remove status handler, error: %s" % ret
groupId = dcgm_agent.dcgmGroupCreate(handle, dcgm_structs.DCGM_GROUP_DEFAULT, "all_gpus_group")
## Invoke method to get information on the newly created group
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId)
## Create reference to DCGM status handler which can be used to get the statuses for multiple
## operations on one or more devices present in the group
status_handle = dcgm_agent.dcgmStatusCreate()
## The worker function can be executed as a separate thread or as part of the main thread.
## Executed as a separate thread here
thread = Thread(target = agent_worker_function, args = (handle, groupId, groupInfo, status_handle))
thread.start()
##########################################
# Any other useful work can be placed here
##########################################
thread.join()
print "Worker thread completed"
## Destroy the group
ret = dcgm_agent.dcgmGroupDestroy(handle, groupId)
assert(ret == dcgm_structs.DCGM_ST_OK), "Failed to remove the test group, error: %s" % ret
## Destroy the status handle
ret = dcgm_agent.dcgmStatusDestroy(status_handle)
assert(ret == dcgm_structs.DCGM_ST_OK), "Failed to remove status handler, error: %s" % ret
def __del__(self):
dcgm_agent.dcgmStatusDestroy(self.handle)