def helper_dcgm_config_powerbudget(handle, gpuIds):
handleObj = pydcgm.DcgmHandle(handle=handle)
systemObj = handleObj.GetSystem()
groupObj = systemObj.GetEmptyGroup("test1")
## Add first GPU to the group
groupObj.AddGpu(gpuIds[0])
gpuIds = groupObj.GetGpuIds() #Only reference GPUs we are testing against
## Get Min and Max Power limit on the group
attributes = dcgm_agent.dcgmGetDeviceAttributes(handle, gpuIds[0])
## Verify that power is supported on the GPUs in the group
if dcgmvalue.DCGM_INT32_IS_BLANK(attributes.powerLimits.maxPowerLimit):
test_utils.skip_test("Needs Power limit to be supported on the GPU")
powerLimit = int((attributes.powerLimits.maxPowerLimit +
attributes.powerLimits.minPowerLimit) / 2)
config_values = dcgm_structs.c_dcgmDeviceConfig_v1()
config_values.mEccMode = dcgmvalue.DCGM_INT32_BLANK
config_values.mPerfState.syncBoost = dcgmvalue.DCGM_INT32_BLANK
config_values.mPerfState.targetClocks.memClock = dcgmvalue.DCGM_INT32_BLANK
config_values.mPerfState.targetClocks.smClock = dcgmvalue.DCGM_INT32_BLANK
config_values.mComputeMode = dcgmvalue.DCGM_INT32_BLANK
config_values.mPowerLimit.type = dcgm_structs.DCGM_CONFIG_POWER_BUDGET_GROUP
config_values.mPowerLimit.val = powerLimit * len(
gpuIds) #Assumes homogenous GPUs
groupObj.config.Set(config_values)
config_values = groupObj.config.Get(dcgm_structs.DCGM_CONFIG_CURRENT_STATE)
assert len(config_values
) > 0, "Failed to get configuration using groupObj.config.Get"
for x in range(0, len(gpuIds)):
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 == powerLimit, "The power limit value for gpuID %d is incorrect. Returned: %d Expected: %s" % (
x, config_values[x].mPowerLimit.val, powerLimit)
pass
def initialize_devices(handle, flags):
gpuIds = dcgm_agent.dcgmGetEntityGroupEntities(handle,
dcgm_fields.DCGM_FE_GPU,
flags)
switchIds = dcgm_agent.dcgmGetEntityGroupEntities(
handle, dcgm_fields.DCGM_FE_SWITCH, flags)
i = 0
for gpuId in gpuIds:
attributes = dcgm_agent.dcgmGetDeviceAttributes(handle, gpuId)
gpuObj = Entity(gpuId,
entityType=dcgm_fields.DCGM_FE_GPU,
uuid=attributes.identifiers.uuid,
bdf=attributes.identifiers.pciBusId)
g_gpus.append(gpuObj)
i = i + 1
i = 0
for switchId in switchIds:
switchObj = Entity(switchId, entityType=dcgm_fields.DCGM_FE_SWITCH)
g_switches.append(switchObj)
i = i + 1
def GetGpuAttributes(self, gpuId):
return dcgm_agent.dcgmGetDeviceAttributes(self._dcgmHandle.handle,
gpuId)
sys.exit(1)
print "Number of valid devices: %d" % len(validDevices)
groupId = dcgm_agent.dcgmGroupCreate(handle, dcgm_structs.DCGM_GROUP_EMPTY,
"test1")
statusHandle = dcgm_agent.dcgmStatusCreate()
for device in validDevices:
ret = dcgm_agent.dcgmGroupAddDevice(handle, groupId, device)
assert (ret == dcgm_structs.DCGM_ST_OK)
## Get attributes for all the devices
attributesForDevices = list()
for device in validDevices:
attributes = dcgm_agent.dcgmGetDeviceAttributes(handle, device)
attributesForDevices.append(attributes)
assert len(
attributesForDevices) != 0, "Can't get attributes for all the devices"
device0_name = attributesForDevices[0].identifiers.deviceName
for attribute in attributesForDevices:
if attribute.identifiers.deviceName != device0_name:
print "Can only run test if all the GPUs are same"
sys.exit(1)
powerLimit_set = dcgmvalue.DCGM_INT32_BLANK
fvSupported = dcgm_agent_internal.dcgmGetLatestValuesForFields(
handle, x, [
dcgm_fields.DCGM_FI_DEV_POWER_MGMT_LIMIT,