def AddGpu(self, gpuId):
if self._IsGroupIdStatic():
raise pydcgm.DcgmException("Can't add a GPU to a static group")
ret = dcgm_agent.dcgmGroupAddDevice(self._dcgmHandle.handle,
self._groupId, gpuId)
dcgm_structs._dcgmCheckReturn(ret)
def test_dcgm_configure_ecc_mode(handle, gpuIds):
test_utils.skip_test("Skipping this test until bug 200377294 is fixed")
groupId = dcgm_agent.dcgmGroupCreate(handle, dcgm_structs.DCGM_GROUP_EMPTY,
"test1")
validDevice = -1
for x in gpuIds:
fvSupported = dcgm_agent_internal.dcgmGetLatestValuesForFields(
handle, x, [
dcgm_fields.DCGM_FI_DEV_ECC_CURRENT,
])
if (fvSupported[0].value.i64 != dcgmvalue.DCGM_INT64_NOT_SUPPORTED):
validDevice = x
break
if (validDevice == -1):
test_utils.skip_test(
"Can only run if at least one GPU with ECC is present")
ret = dcgm_agent.dcgmGroupAddDevice(handle, groupId, validDevice)
assert (ret == dcgm_structs.DCGM_ST_OK
), "Failed to add a device to the group %d. Return %d" % (
groupId.value, ret)
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId)
#Create a status handle
status_handle = dcgm_agent.dcgmStatusCreate()
## Get original ECC mode on the device
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"
eccmodeOnGroupExisting = config_values[0].mEccMode
if eccmodeOnGroupExisting == 0:
eccmodeOnGroupToSet = 1
else:
eccmodeOnGroupToSet = 0
#print eccmodeOnGroupExisting
#print eccmodeOnGroupToSet
## Toggle the ECC mode on the group
config_values = dcgm_structs.c_dcgmDeviceConfig_v1()
config_values.mEccMode = eccmodeOnGroupToSet
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 = dcgmvalue.DCGM_INT32_BLANK
config_values.mPowerLimit.val = dcgmvalue.DCGM_INT32_BLANK
#Clear the status handle to log the errors while setting the config
ret = dcgm_agent.dcgmStatusClear(status_handle)
assert ret == dcgm_structs.DCGM_ST_OK, "Failed to clear the status handle. Return %d" % ret
try:
ret = dcgm_agent.dcgmConfigSet(handle, groupId, config_values,
status_handle)
except dcgm_structs.DCGMError as e:
pass
errors = helper_get_status_list(status_handle)
if len(errors) > 0:
for error in errors:
if error.status == dcgm_structs.DCGM_ST_RESET_REQUIRED:
test_utils.skip_test(
"Skipping the test - Unable to reset the Gpu, FieldId - %d, Return - %d"
% (error.fieldId, error.status))
else:
test_utils.skip_test(
"Skipping the test - Unable to set the ECC mode. FieldId - %d, Return %d"
% (error.fieldId, error.status))
#Sleep after reset
time.sleep(2)
#Clear the status handle to log the errors while setting the config
ret = dcgm_agent.dcgmStatusClear(status_handle)
assert ret == dcgm_structs.DCGM_ST_OK, "Failed to clear the status handle. Return %d" % ret
#Get the current configuration
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"
fvs = dcgm_agent_internal.dcgmGetLatestValuesForFields(
handle, validDevice, [
dcgm_fields.DCGM_FI_DEV_ECC_PENDING,
dcgm_fields.DCGM_FI_DEV_ECC_CURRENT
])
if fvs[0].value.i64 != fvs[1].value.i64:
logger.warning(
"Pending ECC %d != Current ECC %d for gpuId %d. Box probably needs a reboot"
% (fvs[0].value.i64, fvs[1].value.i64, validDevice))
else:
assert config_values[0].mEccMode == (eccmodeOnGroupToSet), "ECC mode %d different from the set value %d" % \
(config_values[0].mEccMode, eccmodeOnGroupToSet)
])
if (fvSupported[0].value.i64 != dcgmvalue.DCGM_INT64_NOT_SUPPORTED):
validDevices.append(x)
if (len(validDevices) == 0):
print "Can only run if at least one GPU with ECC is present"
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"
def test_dcgm_vgpu_configure_ecc_mode(handle, gpuIds):
test_utils.skip_test("Skipping this test until bug 200377294 is fixed")
groupId = dcgm_agent.dcgmGroupCreate(handle, dcgm_structs.DCGM_GROUP_EMPTY,
"test1")
validDevice = -1
for x in gpuIds:
fvSupported = dcgm_agent_internal.dcgmGetLatestValuesForFields(
handle, x, [dcgm_fields.DCGM_FI_DEV_RETIRED_DBE])
if (fvSupported[0].value.i64 != dcgmvalue.DCGM_INT64_NOT_SUPPORTED):
validDevice = x
break
if (validDevice == -1):
test_utils.skip_test(
"Can only run if at least one GPU with ECC is present")
ret = dcgm_agent.dcgmGroupAddDevice(handle, groupId, validDevice)
assert (
ret == dcgm_structs.DCGM_ST_OK
), "Failed to add a device to the group %d. Return %d" % (groupId, ret)
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId)
#Create a status handle
status_handle = dcgm_agent.dcgmStatusCreate()
## Get original ECC mode on the device
vgpu_config_values = dcgm_agent_internal.dcgmVgpuConfigGet(
handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE,
groupInfo.count, status_handle)
assert len(
vgpu_config_values) > 0, "Failed to work with NULL status handle"
eccmodeOnGroupExisting = vgpu_config_values[0].mEccMode
if eccmodeOnGroupExisting == 0:
eccmodeOnGroupToSet = 1
else:
eccmodeOnGroupToSet = 0
#print eccmodeOnGroupExisting
#print eccmodeOnGroupToSet
## Toggle the ECC mode on the group
vgpu_config_values = dcgm_structs.c_dcgmDeviceVgpuConfig_v1()
vgpu_config_values.mEccMode = eccmodeOnGroupToSet
vgpu_config_values.mComputeMode = dcgmvalue.DCGM_INT32_BLANK
vgpu_config_values.mPowerLimit.type = dcgmvalue.DCGM_INT32_BLANK
vgpu_config_values.mPowerLimit.val = dcgmvalue.DCGM_INT32_BLANK
#Clear the status handle to log the errors while setting the config
ret = dcgm_agent.dcgmStatusClear(status_handle)
assert ret == dcgm_structs.DCGM_ST_OK, "Failed to clear the status handle. Return %d" % ret
try:
ret = dcgm_agent_internal.dcgmVgpuConfigSet(handle, groupId,
vgpu_config_values,
status_handle)
except dcgm_structs.DCGMError as e:
pass
errors = helper_get_status_list(status_handle)
if len(errors) > 0:
for error in errors:
if error.status == dcgm_structs.DCGM_ST_RESET_REQUIRED:
test_utils.skip_test(
"Skipping the test - Unable to reset the Gpu, FieldId - %d, Return - %d"
% (error.fieldId, error.status))
else:
test_utils.skip_test(
"Skipping the test - Unable to set the ECC mode. FieldId - %d, Return %d"
% (error.fieldId, error.status))
#Sleep after reset and then apply update for it to occur
time.sleep(2)
dcgm_agent.dcgmUpdateAllFields(handle, 1)
#Clear the status handle to log the errors while setting the config
ret = dcgm_agent.dcgmStatusClear(status_handle)
assert ret == dcgm_structs.DCGM_ST_OK, "Failed to clear the status handle. Return %d" % ret
#Get the current configuration
config_values = dcgm_agent_internal.dcgmVgpuConfigGet(
handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE,
groupInfo.count, status_handle)
assert len(config_values
) > 0, "Failed to get configuration using dcgmiVgpuConfigGet"
assert config_values[0].mEccMode == (
eccmodeOnGroupToSet), "ECC mode different from the set value"