def agent_worker_function(handle, groupId, groupInfo, status_handle):
NUM_ITERATIONS = 5
count = 0
while True:
dcgm_agent.dcgmUpdateAllFields(handle, 1)
## Get the current configuration for the group
config_values = dcgm_agent.dcgmConfigGet(handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE, groupInfo.count, status_handle)
## Since this is a group operation, Check for the status codes if any of the property failed
helper_investigate_status(status_handle)
dcgm_agent.dcgmStatusClear(status_handle)
## Display current configuration for the group
for x in xrange(0,groupInfo.count):
print "GPU Id : %d" % (config_values[x].gpuId)
print "Ecc Mode : %s" % (convert_value_to_string(config_values[x].mEccMode))
print "Auto Boost : %s" % (convert_value_to_string(config_values[x].mPerfState.autoBoost))
print "Sync Boost : %s" % (convert_value_to_string(config_values[x].mPerfState.autoBoost))
print "Mem Clock : %s" % (convert_value_to_string(config_values[x].mPerfState.minVPState.memClk))
print "SM Clock : %s" % (convert_value_to_string(config_values[x].mPerfState.minVPState.procClk))
print "Power Limit : %s" % (convert_value_to_string(config_values[x].mPowerLimit.val))
print "Compute Mode: %s" % (convert_value_to_string(config_values[x].mComputeMode))
print "\n"
count = count + 1
if count == NUM_ITERATIONS:
break
sleep(2)
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 Get(self, configType):
status = pydcgm.DcgmStatus()
gpuIds = self._dcgmGroup.GetGpuIds()
configList = dcgm_agent.dcgmConfigGet(self._dcgmHandle.handle, self._groupId, configType, len(gpuIds), status.handle)
#Throw specific errors before return error
status.ThrowExceptionOnErrors()
return configList
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
autoBoost_set = 1
print "configure autobost"
assert attributesForDevices[
0].vpStates.count > 0, "Can't find clocks for the device"
total_clocks = attributesForDevices[0].vpStates.count
proc_clk_set = attributesForDevices[0].vpStates.vpState[total_clocks /
2].procClk
mem_clk_set = attributesForDevices[0].vpStates.vpState[total_clocks / 2].memClk
## Always Switch the ecc mode
ecc_set = 1
groupInfo = dcgm_agent.dcgmGroupGetInfo(handle, groupId,
dcgm_structs.c_dcgmGroupInfo_version2)
config_values = dcgm_agent.dcgmConfigGet(
handle, groupId, dcgm_structs.DCGM_CONFIG_CURRENT_STATE, groupInfo.count,
0)
assert len(config_values) > 0, "Failed to work with NULL status handle"
eccmodeOnGroupExisting = config_values[0].mEccMode
if eccmodeOnGroupExisting == 0:
ecc_set = 1
else:
ecc_set = 0
syncboost_set = 1
compute_set = dcgm_structs.DCGM_CONFIG_COMPUTEMODE_DEFAULT
config_values = dcgm_structs.c_dcgmDeviceConfig_v1()
config_values.mEccMode = ecc_set
config_values.mPerfState.syncBoost = syncboost_set
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)