NVIDIA Validation Suite User Guide (PDF) - vR352 (older) - Last updated May 7, 2015 - Send Feedback

Overview

The NVIDIA Validation Suite (NVVS) is the system administrator and cluster manager's tool for detecting and troubleshooting common problems affecting NVIDIA® Tesla™ GPUs in a high performance computing environments. NVVS focuses on software and system configuration issues, diagnostics, topological concerns, and relative performance.

NVVS Goals

The NVIDIA Validation Suite is designed to:
  1. Provide a system-level tool, in production environments, to assess cluster readiness levels before a workload is deployed.
  2. Facilitate multiple run modes:
    • Interactive via an administrator or user in plain text.
    • Scripted via another tool with easily parseable output.
  3. Provide multiple test timeframes to facilitate different preparedness or failure conditions:
    • Quick tests to use as a readiness metric
    • Medium tests to use as an epilogue on failure
    • Long tests to be run by an administrator as post-mortem
  4. Integrate the following concepts into a single tool to discover deployment, system software and hardware configuration issues, basic diagnostics, integration issues, and relative system performance.
    • Deployment and Software Issues
      • NVML library access and versioning
      • CUDA library access and versioning
      • Software conflicts
    • Hardware Issues and Diagnostics
      • PCIe interface checks
      • Framebuffer and memory checks
      • Compute engine checks
    • Integration Issues
      • PCIe replay counter checks
      • Topological limitations
      • Permissions, driver, and cgroups checks
      • Basic power and thermal constraint checks
    • Stress Checks
      • Power and thermal stress
      • Throughput stress
      • Constant relative system performance
      • Maximum relative system performance
  5. Provide troubleshooting help
  6. Easily integrate into Cluster Scheduler and Cluster Management applications
  7. Reduce downtime and failed GPU jobs

Beyond the Scope of the NVIDIA Validation Suite

NVVS is not designed to:
  1. Provide comprehensive hardware diagnostics
  2. Actively fix problems
  3. Replace the field diagnosis tools. Please refer to http://docs.nvidia.com/deploy/hw-field-diag/index.html for that process.
  4. Facilitate any RMA process. Please refer to http://docs.nvidia.com/deploy/rma-process/index.html for those procedures.

Dependencies

  • This version of the NVIDIA Validation Suite is designed to support the target environments supported by the GPU Deployment Kit (GDK) from whence it came. Please refer to the GDK README for version information. The GDK and supporting documentation can be located at https://developer.nvidia.com/gpu-deployment-kit.
  • NVVS requires the standard C++ runtime library with GLIBCXX of at least version 3.4.5 or greater.

Supported Products

The NVIDIA Validation Suite supports Tesla GPUs running on 64-bit Linux (bare metal) operating systems. NVIDIA® Tesla™ Line:

  • All Kepler architecture GPUs
  • All Maxwell architecture GPUs

Using NVVS

The various command line options of NVVS are designed to control general execution parameters, whereas detailed changes to execution behavior are contained within the configuration files detailed in the next chapter.

Command line options

The various options for NVVS are as follows:

Short option Long option Description
-a --appendLog When generating a debug logfile, do not overwrite the contents of a current log. Used in conjuction with the -d and -l options.
-c --config Specify the configuration file to be used. The default is /etc/nvidia-validation-suite/nvvs.conf
--configless Run NVVS in a configless mode. Executes a "long" test on all supported GPUs.
-d --debugLevel Specify the debug level for the output log. The range is 0 to 5 with 5 being the most verbose. Used in conjunction with the -l flag.
-g --listGpus List the GPUs available and exit. This will only list GPUs that are supported by NVVS.
-l --debugLogFile Specify the logfile for debug information. This will produce an encrypted log file intended to be returned to NVIDIA for post-run analysis after an error.
--quiet No console output given. See logs and return code for errors.
-p --pluginpath Specify a custom path for the NVVS plugins.
-s --scriptable Produce output in a colon-separated, more script-friendly and parseable format.
--specifiedtest Run a specific test in a configless mode. Multiple word tests should be in quotes.
--statsonfail Output statistic logs only if a test failure is encountered.
-t --listTests List the tests available to be executed through NVVS and exit. This will list only the readily loadable tests given the current path and library conditions.
-v --verbose Enable verbose reporting.
  --version Displays the version information and exits.
-h --help Display usage information and exit.

Usage Examples

To display the list of GPUs available on the system.

user@hostname
$ nvvs -g
			
NVIDIA Validation Suite (version 352.00)

Supported GPUs available:
	[0000:01:00.0] -- Tesla K40c
	[0000:05:00.0] -- Tesla K20c
	[0000:06:00.0] -- Tesla K20c
		

An example "quick" test (explained later) using a custom configuration file.

user@hostname
$ nvvs -c Tesla_K40c_quick.conf
			
NVIDIA Validation Suite (version 352.00)

	Software
		Blacklist ......................................... PASS
		NVML Library ...................................... PASS
		CUDA Main Library ................................. PASS
		CUDA Toolkit Libraries ............................ PASS
		Permissions and OS-related Blocks ................. PASS
		Persistence Mode .................................. PASS
		Environmental Variables ........................... PASS			
		

To output an encrypted debug file at the highest debug level to send to NVIDIA for analysis after a problem.

user@hostname
$ nvvs -c Tesla_K40c_medium.conf -d 5 -l debug.log

NVIDIA Validation Suite (version 352.00)

	Software
		Blacklist ......................................... PASS
		NVML Library ...................................... PASS
		CUDA Main Library ................................. PASS
		CUDA Toolkit Libraries ............................ PASS
		Permissions and OS-related Blocks ................. PASS
		Persistence Mode .................................. PASS
		Environmental Variables ........................... PASS
	Hardware
		Memory GPU0 ....................................... PASS
	Integration
		PCIe .............................................. FAIL
           *** GPU 0 is running at PCI link width 8X, which is below the minimum allowed link width of 16X (parameter: 
min_pci_width)"			
		

The output file, debug.log would then be returned to NVIDIA.

Configuration file

The NVVS configuration file is a YAML-formatted (e.g. human-readable JSON) text file with three main stanzas controlling the various tests and their execution.

The general format of a configuration file consists of:

%YAML 1.2
---
      
globals:
  key1: value
  key2: value
      
test_suite_name:
- test_class_name1:
    test_name1:
      key1: value
      key2: value
      subtests:
        subtest_name1:
          key1: value
          key2: value
    test_name2:
      key1: value
      key2: value
-test_class_name2:
    test_name3:
      key1: value
      key2: value
      
gpus:
- gpuset: name
  properties:
    key1: value
    key2: value
  tests:
    name: test_suite_name
    

There are three distinct sections: globals, test_suite_name, and gpus each with its own subsection of parameters and as is with any YAML document, indentation is important thus if errors are generated from your own configuration files please refer to this example for indentation reference.

Global parameters

Keyword Value Type Description
logfile String The prefix for all detailed test data able to be used for post-processing.
logfile_type String Can be json, text, or binary. Used in conjunction with the logfile global parameter. Default is JSON.
scriptable Boolean Accepts true, or false. Produces a script-friendly, colon-separated output and is identical to the -s command line parameter.
serial_override Boolean Accepts true, or false. Some tests are designed to run in parallel if multiple GPUs are given. This parameter overrides that behavior serializing execution across all tests.
require_persistence_mode Boolean Accepts true, or false. Persistence mode is a prerequisite for some tests, this global overrides that requirement and should only be used if it is not possible to activate persistence mode on your system.

GPU parameters

The gpus stanza may consist of one or more gpusets which will each match zero or more GPUs on the system based on their properties(a match of zero will produce an error).

GPUs are matched based on the following criteria with their configuration file keywords in parenthesis:
  • Name of the GPU, i.e. Tesla K40c (name)
  • Brand of the GPU, i.e. Tesla (brand)
  • A comma separated list of indexes (index)
  • The GPU UUID (uuid)
  • or the PCIe Bus ID (busid)

The matching rules are based off of exclusion. First, the list of supported GPUs is taken and if no properties tag is given then all GPUs will be used in the test. Because a UUID or PCIe Bus ID can only match a single GPU, if those properties are given then only that GPU will be used if found. The remaining properties, index, brand, and name work in an "AND" fashion such that, if specified, the result must match at least one GPU on the system for a test to be performed.

For example, if name is set to "Tesla K40c" and index is set to "0" NVVS will error if index 0 is not a Tesla K40c. By specifying both brand and index a user may limit a test to specific "Tesla" cards for example. In this version of NVVS, all matching GPUs must be homogeneous.

The second identifier for a gpuset is tests. This parameter specifies either the suite of tests that a user wishes to run or the test itself.

At present the following suites are available:

  • Quick -- meant as a pre-run sanity check to ensure that the GPUs are ready for a job. Currently runs the Deployment tests described in the next chapter.
  • Medium -- meant as a quick, post-error check to make sure that nothing very obvious such as ECC enablement or double-bit errors have occurred on a GPU. Currently runs the Deployment, Memory, and PCIe/Bandwidth tests.
  • Long -- meant as a more extensive check to find potential power and/or performance problems within a cluster. Currently runs an extensive test that involves Deployment, Memory, PCI/Bandwidth, Power, and Performance.

An individual test can also be specified. Currently the keywords are: Memory, Targeted Performance, Targeted Power, and PCIe. Please see the "custom" section in the next subchapter to configure and tweak the parameters when this method is used.

Test parameters

The format of the NVVS configuration file is designed for extensibility. Each test suite above can be customized in a number of ways described in detail in the following chapter for each test. Individual tests belong to a specific class of functionality which, when wanting to customize specific parameters, must also be specified.

The classes and the respective tests they perform are as follows:

Class name Tests Brief description
Software Deployment Checks for various runtime libraries, persistence mode, permissions, environmental variables, and blacklisted drivers.
Hardware Memory Allocate as much memory as possible and perform read/write operations.
Integration PCIe Test host to GPU, GPU to host, and P2P (if possible) bandwidth.
Performance Targeted Performance Sustain a specific targeted performance level for a given amount of time.
Targeted Power Sustain a specific targeted power level for a given amount of time.

Some tests also have subtests that can be enabled by using the subtests keyword and then hierarchically adding the subtest parameters desired beneath. An example would be the PCIe Bandwidth test which may have a section that looks similar to this:

long:
- integration:
    pcie:
      test_unpinned: false
      subtests:
        h2d_d2h_single_pinned:
          min_bandwidth: 20
          min_pci_width: 16
        

When only a specific test is given in the GPU set portion of the configuration file, both the suite and class of the test are custom. For example:

%YAML 1.2 
---

globals:
  logfile: nvvs.log

custom:
- custom:
    targeted performance:
      test_duration: 60

gpus:
- gpuset: all_K40c
  properties:
    name: Tesla K40c
  tests: 
    - name: targeted performance
        

Overview of Plugins

The NVIDIA Validation Suite consists of a series of plugins that are each designed to accomplish a different goal.

Deployment Plugin

The deployment plugin's purpose is to verify the compute environment is ready to run Cuda applications and is able to load the NVML library.

Preconditions

  • LD_LIBRARY_PATH must include the path to the cuda libraries, which for version X.Y of Cuda is normally /usr/local/cuda-X.Y/lib64, which can be set by running export LD_LIBRARY_PATH=/usr/local/cuda-X.Y/lib64
  • The linux nouveau driver must not be running, and should be blacklisted since it will conflict with the nvidia driver

Configuration Parameters

None at this time.

Stat Outputs

None at this time.

Failure

The plugin will fail if:
  • The corresponding device nodes for the target GPU(s) are being blocked by the operating system (e.g. cgroups) or exist without r/w permissions for the current user.
  • The NVML library libnvidia-ml.so cannot be loaded
  • The Cuda runtime libraries cannot be loaded
  • The nouveau driver is found to be loaded
  • Any pages are pending retirement on the target GPU(s)

GPU Bandwidth Plugin

The GPU bandwidth plugin's purpose is to measure the bandwidth and latency to and from the GPUs and the host.

Preconditions

  • This plugin must be run as root or be granted the ability to set application clocks on supporting hardware (K10 or higher). To allow non-root to change application clocks on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change application clocks with nvidia-smi -i deviceIndex -acp 0
  • If the GPUs the plugin is being run on are K80 or higher, the plugin must be run as root or be granted the ability to disable auto boosted clocks. Non-root is allowed to set auto boost by default on K80 or higher. To enable non-root setting auto boost on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change auto boosted clocks with nvidia-smi -i deviceIndex --auto-boost-permission=0

Sub Tests

The plugin consists of several self-tests that each measure a different aspect of bandwidth or latency. Each subtest has either a pinned/unpinned pair or a p2p enabled/p2p disabled pair of identical tests. Pinned/unpinned tests use either pinned or unpinned memory when copying data between the host and the GPUs.

Each sub test is represented with a tag that is used both for specifying configuration parameters for the sub test and for outputting stats for the sub test. P2p enabled/p2p disabled tests enable or disable GPUs on the same card talking to each other directly rather than through the PCIe bus.

Sub Test Tag Pinned/Unpinned

P2P Enabled/P2P Disabled
Description
h2d_d2h_single_pinned Pinned Device <-> Host Bandwidth, one GPU at a time
h2d_d2h_single_unpinned Unpinned Device <-> Host Bandwidth, one GPU at a time
h2d_d2h_concurrent_pinned Pinned Device <-> Host Bandwidth, all GPUs concurrently
h2d_d2h_concurrent_unpinned Unpinned Device <-> Host Bandwidth, all GPUs concurrently
h2d_d2h_latency_pinned Pinned Device <-> Host Latency, one GPU at a time
h2d_d2h_latency_unpinned Unpinned Device <-> Host Latency, one GPU at a time
p2p_bw_p2p_enabled P2P Enabled Device <-> Device bandwidth one GPU pair at a time
p2p_bw_p2p_disabled P2P Disabled Device <-> Device bandwidth one GPU pair at a time
p2p_bw_concurrent_p2p_enabled P2P Enabled Device <-> Device bandwidth, concurrently, focusing on bandwidth between GPUs between GPUs likely to be directly connected to each other -> for each (index / 2) and (index / 2)+1
p2p_bw_concurrent_p2p_disabled P2P Disabled Device <-> Device bandwidth, concurrently, focusing on bandwidth between GPUs between GPUs likely to be directly connected to each other -> for each (index / 2) and (index / 2)+1
1d_exch_bw_p2p_enabled P2P Enabled Device <-> Device bandwidth, concurrently, focusing on bandwidth between gpus, every GPU either sending to the gpu with the index higher than itself (l2r) or to the gpu with the index lower than itself (r2l)
1d_exch_bw_p2p_disabled P2P Disabled Device <-> Device bandwidth, concurrently, focusing on bandwidth between gpus, every GPU either sending to the gpu with the index higher than itself (l2r) or to the gpu with the index lower than itself (r2l)
p2p_latency_p2p_enabled P2P Enabled Device <-> Device Latency, one GPU pair at a time
p2p_latency_p2p_disabled P2P Disabled Device <-> Device Latency, one GPU pair at a time

Configuration Parameters- Global

Parameter Name Type Default Value Range Description
test_pinned Bool True True/False Include subtests that test using pinned memory.
test_unpinned Bool True True/False Include subtests that test using unpinned memory.
test_p2p_on Bool True True/False Include subtests that require peer to peer (P2P) memory transfers between cards to occur.
test_p2p_off Bool True True/False Include subtests that do not require peer to peer (P2P) memory transfers between cards to occur.
max_pcie_replays Float 80.0 1.0 - 1000000.0 Maximum number of PCIe replays to allow per GPU for the duration of this plugin. This is based on an expected replay rate <8 per minute for PCIe Gen 3.0, assuming this plugin will run for less than a minute and allowing 10x as many replays before failure.

Configuration Parameters- Sub Test

Parameter Name Default (Range) Affected Sub Tests Description
min_bandwidth Null

(0.0 - 100.0)
h2d_d2h_single_pinned, h2d_d2h_single_unpinned, h2d_d2h_concurrent_pinned, h2d_d2h_concurrent_unpinned Minimum bandwidth in GB/s that must be reached for this sub-test to pass.
max_latency 100,000.0

(0.0 - 1,000,000.0)
h2d_d2h_latency_pinned, h2d_d2h_latency_unpinned Latency in microseconds that cannot be exceeded for this sub-test to pass.
min_pci_generation 1.0

(1.0 - 3.0)
h2d_d2h_single_pinned, h2d_d2h_single_unpinned Minimum allowed PCI generation that the GPU must be at or exceed for this sub-test to pass.
min_pci_width 1.0

(1.0 - 16.0)
h2d_d2h_single_pinned, h2d_d2h_single_unpinned Minimum allowed PCI width that the GPU must be at or exceed for this sub-test to pass. For example, 16x = 16.0.

Stat Outputs - Global

Stat Name Stat Scope Type Description
pcie_replay_count GPU Float The per second reading of PCIe replays that have occurred since the start of the GPU Bandwidth plugin.

Stat Outputs -Sub Test

Stats for the GPU Bandwidth test are also output on a test by test basis, using the sub test name as the group name key. The following stats sections are organized by sub test.

h2d_d2h_single_pinned/h2d_d2h_single_unpinned

Stat Name Type Description
N_h2d Float Average bandwidth from host to device for device N
N_d2h Float Average bandwidth from device to host for device N
N_bidir Float Average bandwidth from device to host and host to device at the same time for device N

h2d_d2h_concurrent_pinned/h2d_d2h_concurrent_unpinned

Stat Name Type Description
N_h2d Float Average bandwidth from host to device for device N
N_d2h Float Average bandwidth from device to host for device N
N_bidir Float Average bandwidth from device to host and host to device at the same time for device N
sum_bidir Float Sum of the average bandwidth from device to host and host to device for all devices.
sum_h2d Float Sum of the average bandwidth from host to device for all devices.
sum_d2h Float Sum of the average bandwidth from device to host for all devices.

h2d_d2h_latency_pinned/h2d_d2h_latency_unpinned

Stat Name Type Description
N_h2d Float Average latency from host to device for device N
N_d2h Float Average latency from device to host for device N
N_bidir Float Average latency from device to host and host to device at the same time for device N

p2p_bw_p2p_enabled/p2p_bw_p2p_disabled

Stat Name Type Description
N_M_onedir Float Average bandwidth from device N to device M, copying one direction at a time.
N_M_bidir Float Average bandwidth from device N to device M, copying both directions at the same time.

p2p_bw_concurrent_p2p_enabled/p2p_bw_concurrent_p2p_disabled

Stat Name Type Description
l2r_N_M Float Average bandwidth from device N to device M
r2l_N_M Float Average bandwidth from device M to device N
bidir_N_M Float Average bandwidth from device M to device N, copying concurrently
r2l_sum Float Sum of average bandwidth for all right (M) to left (N) copies
r2l_sum Float Sum of average bidirectional bandwidth for all right (M) to left (N) and left to right copies copies

1d_exch_bw_p2p_enabled/1d_exch_bw_p2p_disabled

Stat Name Type Description
l2r_N Float Average bandwidth from device N to device N+1
r2l_N Float Average bandwidth from device N to device N-1
l2r_sum Float Sum of all l2r average bandwidth stats
r2l_sum Float Sum of all l2r average bandwidth stats

p2p_latency_p2p_enabled/p2p_latency_p2p_disabled

Stat Name Type Description
N_M Float Average latency from device N to device M

Failure

The plugin will fail if:
  • The latency exceeds the configured threshold for relevant tests.
  • The bandwidth cannot exceed the configured threshold for relevant tests.
  • If the number of PCIe retransmits exceeds a user-provided threshold.

SM Performance Plugin

The performance plugin’s purpose is to bring the GPU to a target performance level in gigaflops by doing large matrix multiplications using cublas. Unlike the GPU performance plugin, the SM performance plugin does not copy the source arrays to the GPU before every matrix multiplication. This allows the SM performance plugin's performance to not be capped by device to host bandwidth. The plugin calculates how many matrix operations per second are necessary to achieve the configured performance target and fails if it cannot achieve that target.

This plugin should be used to watch for thermal, power and related anomalies while the target GPU(s) are under realistic load conditions. By setting the appropriate parameters a user can ensure that all GPUs in a node or cluster reach desired performance levels. Further analysis of the generated stats can also show variations in the required power, clocks or temperatures to reach these targets, and thus highlight GPUs or nodes that are operating less efficiently.

Preconditions

  • This plugin must be run as root or be granted the ability to set application clocks on supporting hardware (K10 or higher). To allow non-root to change application clocks on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change application clocks with nvidia-smi -i deviceIndex -acp 0
  • If the GPUs the plugin is being run on are K80 or higher, the plugin must be run as root or be granted the ability to disable auto boosted clocks. Non-root is allowed to set auto boost by default on K80 or higher. To enable non-root setting auto boost on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change auto boosted clocks with nvidia-smi -i deviceIndex --auto-boost-permission=0

Configuration Parameters

Parameter Name Type Default Value Range Description
test_duration Float 90.0 30.0 - 3600.0 How long the performance test should run for in seconds. It is recommended to set this to at least 30 seconds for performance to stabilize.
temperature_max Float Null 30.0 - 120.0 The maximum temperature in C the card is allowed to reach during the test. Note that this check is disabled by default. Use nvidia-smi -q to see the normal temperature limits of your device.
target_performance Float Null SKU dependent The maximum relative performance each card will attempt to achieve.

Stat Outputs

Stat Name Stat Scope Type Description
power_usage GPU Time series Float Per second power usage of each GPU in watts. Note that for multi-GPU boards, each GPU gets a fraction of the power budget of the board.
graphics_clock GPU Time series Float Per second clock rate of each GPU in MHZ
memory_clock GPU Time series Float Per second clock rate of the GPU’s memory in MHZ
nvml_events GPU Time series Int64 Any events that were read with nvmlEventSetWait during the test and the timestamp it was read it.
power_violation GPU Time series Float Percentage of time this GPU was violating power constraints.
gpu_temperature GPU Time series Float Per second temperature of the GPU in degrees C
perf_gflops GPU Time series Float The per second reading of average gflops since the test began.
flops_per_op GPU Float Flops (floating point operations) per operation queued to the GPU stream. One operation is one call to cublasSgemm or cublasDgemm
bytes_copied_per_op GPU Float How many bytes are copied to + from the GPU per operation
num_cuda_streams GPU Float How many cuda streams were used per gpu to queue operations to the GPUs
try_ops_per_sec GPU Float Calculated number of ops/second necessary to achieve target gigaflops

Failure

The plugin will fail if:
  • The GPU temperature exceeds a user-provided threshold.
  • If temperature violation counters increase
  • If the target performance level cannot be reached
  • If GPU double bit ECC errors occur

Performance Plugin

The performance plugin’s purpose is to bring the GPU to a target performance level in gigaflops by doing large matrix multiplications using cublas. The plugin calculates how many matrix operations per second are necessary to achieve the configured performance target and fails if it cannot achieve that target.

This plugin should be used to watch for thermal, power and related anomalies while the target GPU(s) are under realistic load conditions. By setting the appropriate parameters a user can ensure that all GPUs in a node or cluster reach desired performance levels. Further analysis of the generated stats can also show variations in the required power, clocks or temperatures to reach these targets, and thus highlight GPUs or nodes that are operating less efficiently.

Preconditions

  • This plugin must be run as root or be granted the ability to set application clocks on supporting hardware (K10 or higher). To allow non-root to change application clocks on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change application clocks with nvidia-smi -i deviceIndex -acp 0
  • If the GPUs the plugin is being run on are K80 or higher, the plugin must be run as root or be granted the ability to disable auto boosted clocks. Non-root is allowed to set auto boost by default on K80 or higher. To enable non-root setting auto boost on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change auto boosted clocks with nvidia-smi -i deviceIndex --auto-boost-permission=0

Configuration Parameters

Parameter Name Type Default Value Range Description
test_duration Float 120.0 30.0 - 3600.0 How long the performance test should run for in seconds. It is recommended to set this to at least 30 seconds for performance to stabilize.
temperature_max Float Null 30.0 - 120.0 The maximum temperature in C the card is allowed to reach during the test. Note that this check is disabled by default. Use nvidia-smi -q to see the normal temperature limits of your device.
target_performance Float Null SKU dependent The maximum relative performance each card will attempt to achieve.
max_pcie_replays Float 160.0 1.0 - 1000000.0 Maximum number of PCIe replays to allow per GPU for the duration of this plugin. This is based on an expected replay rate <8 per minute for PCIe Gen 3.0, assuming this plugin will run for 2 minutes (configurable) and allowing 10x as many replays before failure.

Stat Outputs

Stat Name Stat Scope Type Description
power_usage GPU Time series Float Per second power usage of each GPU in watts. Note that for multi-GPU boards, each GPU gets a fraction of the power budget of the board.
graphics_clock GPU Time series Float Per second clock rate of each GPU in MHZ
memory_clock GPU Time series Float Per second clock rate of the GPU’s memory in MHZ
nvml_events GPU Time series Int64 Any events that were read with nvmlEventSetWait during the test and the timestamp it was read it.
power_violation GPU Time series Float Percentage of time this GPU was violating power constraints.
gpu_temperature GPU Time series Float Per second temperature of the GPU in degrees C
perf_gflops GPU Time series Float The per second reading of average gflops since the test began.
flops_per_op GPU Float Flops (floating point operations) per operation queued to the GPU stream. One operation is one call to cublasSgemm or cublasDgemm
bytes_copied_per_op GPU Float How many bytes are copied to + from the GPU per operation
num_cuda_streams GPU Float How many cuda streams were used per gpu to queue operations to the GPUs
try_ops_per_sec GPU Float Calculated number of ops/second necessary to achieve target gigaflops
pcie_replay_count GPU Float The per second reading of PCIe replays that have occurred since the start of the GPU Performance plugin.

Failure

The plugin will fail if:
  • The GPU temperature exceeds a user-provided threshold.
  • If temperature violation counters increase
  • If the target performance level cannot be reached
  • If GPU double bit ECC errors occur
  • If the number of PCIe retransmits exceeds a user-provided threshold.

Power Plugin

The purpose of the power plugin is to bring the GPUs to a preconfigured power level in watts by gradually increasing the compute load on the GPUs until the desired power level is achieved. This verifies that the GPUs can sustain a power level for a reasonable amount of time without problems like thermal violations arising.

Preconditions

  • This plugin must be run as root or be granted the ability to set application clocks on supporting hardware (K10 or higher). To allow non-root to change application clocks on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change application clocks with nvidia-smi -i deviceIndex -acp 0
  • If the GPUs the plugin is being run on are K80 or higher, the plugin must be run as root or be granted the ability to disable auto boosted clocks. Non-root is allowed to set auto boost by default on K80 or higher. To enable non-root setting auto boost on device deviceIndex, as root, first enable persistence mode with nvidia-smi -i deviceIndex -pm 1 and then allow non-root to change auto boosted clocks with nvidia-smi -i deviceIndex --auto-boost-permission=0

Configuration Parameters

Parameter Name Type Default Value Range Description
test_duration Float 120.0 30.0 - 3600.0 How long the performance test should run for in seconds. It is recommended to set this to at least 60 seconds for performance to stabilize.
temperature_max Float Null 30.0 - 120.0 The maximum temperature in C the card is allowed to reach during the test. Note that this check is disabled by default. Use nvidia-smi -q to see the normal temperature limits of your device.
target_power Float Differs per GPU Differs per GPU What power level in wattage we should try to maintain. If this is set to greater than the enforced power limit of the GPU, then we will try to power cap the device

Stat Outputs

Stat Name Stat Scope Type Description
power_usage GPU Time series Float Per second power usage of each GPU in watts. Note that for multi-GPU boards, each GPU gets a fraction of the power budget of the board.
graphics_clock GPU Time series Float Per second clock rate of each GPU in MHZ
memory_clock GPU Time series Float Per second clock rate of the GPU’s memory in MHZ
nvml_events GPU Time series Int64 Any events that were read with nvmlEventSetWait during the test and the timestamp it was read it.
power_violation GPU Time series Float Percentage of time this GPU was violating power constraints.
gpu_temperature GPU Time series Float Per second temperature of the GPU in degrees C

Failure

The plugin will fail if:
  • The GPU temperature exceeds a user-provided threshold.
  • If temperature violation counters increase
  • If the target performance level cannot be reached
  • If GPU double bit ECC errors occur

Test Output

The output of tests can be collected by setting the "logfile" global parameter which represents the prefix for the detailed outputs produced by each test. The default type of output is JSON but text and binary outputs are available as well. The latter two are meant more for parsing and direct reading by custom consumers respectively so this portion of the document will focus on the JSON output.

JSON Output

The JSON output format is keyed based off of the "stats" keys given in each test overview from Chapter 3. These standard JSON files can be processed in any number of ways but two example Python scripts have been provided to aid in visualization in the default installation directory.. The first is a JSON to comma-separated value script (json2csv.py) which can be used to import key values in to a graphing spreadsheet. Proper usage would be:

user@hostname
$ python json2csv.py -i stats_targeted_performance.json -o stats.csv -k gpu_temperature,power_usage
    

Also provided is an example Python script that uses the pygal library to generate readily viewable scalar vector graphics charts (json2svg.py), able to be opened in any browser. Proper usage would be:

user@hostname
$ python json2svg.py -i stats_targeted_performance.json -o stats.svg -k gpu_temperature,power_usage
    

Notices

Notice

ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.

Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.

Trademarks

NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.


NVIDIA Validation Suite User Guide (PDF) - vR352 (older) - Last updated May 7, 2015 - Send Feedback