Optimize TensorFlow performance using the Profiler (original) (raw)

This guide demonstrates how to use the tools available with the TensorFlow Profiler to track the performance of your TensorFlow models. You will learn how to understand how your model performs on the host (CPU), the device (GPU), or on a combination of both the host and device(s).

Profiling helps understand the hardware resource consumption (time and memory) of the various TensorFlow operations (ops) in your model and resolve performance bottlenecks and, ultimately, make the model execute faster.

This guide will walk you through how to install the Profiler, the various tools available, the different modes of how the Profiler collects performance data, and some recommended best practices to optimize model performance.

If you want to profile your model performance on Cloud TPUs, refer to theCloud TPU guide.

Install the Profiler and GPU prerequisites

Install the Profiler plugin for TensorBoard with pip. Note that the Profiler requires the latest versions of TensorFlow and TensorBoard (>=2.2).

pip install -U tensorboard_plugin_profile

To profile on the GPU, you must:

  1. Meet the NVIDIA® GPU drivers and CUDA® Toolkit requirements listed onTensorFlow GPU support software requirements.
  2. Make sure theNVIDIA® CUDA® Profiling Tools Interface(CUPTI) exists on the path:
/sbin/ldconfig -N -v <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo stretchy="false">(</mo><mi>s</mi><mi>e</mi><msup><mi>d</mi><mo mathvariant="normal" lspace="0em" rspace="0em">′</mo></msup><mi>s</mi><mi mathvariant="normal">/</mi><mo>:</mo><mi mathvariant="normal">/</mi><mi mathvariant="normal">/</mi><msup><mi>g</mi><mo mathvariant="normal" lspace="0em" rspace="0em">′</mo></msup><mo>&lt;</mo><mo>&lt;</mo><mo>&lt;</mo></mrow><annotation encoding="application/x-tex">(sed &#x27;s/:/ /g&#x27; &lt;&lt;&lt; </annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1.0019em;vertical-align:-0.25em;"></span><span class="mopen">(</span><span class="mord mathnormal">se</span><span class="mord"><span class="mord mathnormal">d</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.7519em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">′</span></span></span></span></span></span></span></span></span><span class="mord mathnormal">s</span><span class="mord">/</span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">:</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="base"><span class="strut" style="height:1.0019em;vertical-align:-0.25em;"></span><span class="mord">//</span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">g</span><span class="msupsub"><span class="vlist-t"><span class="vlist-r"><span class="vlist" style="height:0.7519em;"><span style="top:-3.063em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mtight">′</span></span></span></span></span></span></span></span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">&lt;&lt;&lt;</span></span></span></span>LD_LIBRARY_PATH) | \  
grep libcupti

If you don't have CUPTI on the path, prepend its installation directory to the$LD_LIBRARY_PATH environment variable by running:

export LD_LIBRARY_PATH=/usr/local/cuda/extras/CUPTI/lib64:$LD_LIBRARY_PATH

Then, run the ldconfig command above again to verify that the CUPTI library is found.

Resolve privilege issues

When you run profiling with CUDA® Toolkit in a Docker environment or on Linux, you may encounter issues related to insufficient CUPTI privileges (CUPTI_ERROR_INSUFFICIENT_PRIVILEGES). Go to theNVIDIA Developer Docsto learn more about how you can resolve these issues on Linux.

To resolve CUPTI privilege issues in a Docker environment, run

docker run option '--privileged=true'

Access the Profiler from the Profile tab in TensorBoard, which appears only after you have captured some model data.

The Profiler has a selection of tools to help with performance analysis:

Overview page

The overview page provides a top level view of how your model performed during a profile run. The page shows you an aggregated overview page for your host and all devices, and some recommendations to improve your model training performance. You can also select individual hosts in the Host dropdown.

The overview page displays data as follows:

image

Input pipeline analyzer

When a TensorFlow program reads data from a file it begins at the top of the TensorFlow graph in a pipelined manner. The read process is divided into multiple data processing stages connected in series, where the output of one stage is the input to the next one. This system of reading data is called the_input pipeline_.

A typical pipeline for reading records from files has the following stages:

  1. File reading.
  2. File preprocessing (optional).
  3. File transfer from the host to the device.

An inefficient input pipeline can severely slow down your application. An application is considered input bound when it spends a significant portion of time in the input pipeline. Use the insights obtained from the input pipeline analyzer to understand where the input pipeline is inefficient.

The input pipeline analyzer tells you immediately whether your program is input bound and walks you through device- and host-side analysis to debug performance bottlenecks at any stage in the input pipeline.

Check the guidance on input pipeline performance for recommended best practices to optimize your data input pipelines.

Input pipeline dashboard

To open the input pipeline analyzer, select Profile, then selectinput_pipeline_analyzer from the Tools dropdown.

image

The dashboard contains three sections:

  1. Summary: Summarizes the overall input pipeline with information on whether your application is input bound and, if so, by how much.
  2. Device-side analysis: Displays detailed, device-side analysis results, including the device step-time and the range of device time spent waiting for input data across cores at each step.
  3. Host-side analysis: Shows a detailed analysis on the host side, including a breakdown of input processing time on the host.

Input pipeline summary

The Summary reports if your program is input bound by presenting the percentage of device time spent on waiting for input from the host. If you are using a standard input pipeline that has been instrumented, the tool reports where most of the input processing time is spent.

Device-side analysis

The device-side analysis provides insights on time spent on the device versus on the host and how much device time was spent waiting for input data from the host.

  1. Step time plotted against step number: Displays a graph of device step time (in milliseconds) over all the steps sampled. Each step is broken into the multiple categories (with different colors) of where time is spent. The red area corresponds to the portion of the step time the devices were sitting idle waiting for input data from the host. The green area shows how much of the time the device was actually working.
  2. Step time statistics: Reports the average, standard deviation, and range ([minimum, maximum]) of the device step time.

Host-side analysis

The host-side analysis reports a breakdown of the input processing time (the time spent on tf.data API ops) on the host into several categories:

Expand Input Op Statistics to inspect the statistics for individual input ops and their categories broken down by execution time.

image

A source data table will appear with each entry containing the following information:

  1. Input Op: Shows the TensorFlow op name of the input op.
  2. Count: Shows the total number of instances of op execution during the profiling period.
  3. Total Time (in ms): Shows the cumulative sum of time spent on each of those instances.
  4. Total Time %: Shows the total time spent on an op as a fraction of the total time spent in input processing.
  5. Total Self Time (in ms): Shows the cumulative sum of the self time spent on each of those instances. The self time here measures the time spent inside the function body, excluding the time spent in the function it calls.
  6. Total Self Time %. Shows the total self time as a fraction of the total time spent on input processing.
  7. Category. Shows the processing category of the input op.

TensorFlow stats

The TensorFlow Stats tool displays the performance of every TensorFlow op (op) that is executed on the host or device during a profiling session.

image

The tool displays performance information in two panes:

You can choose to include or exclude Idle time in the pie charts and table.

Trace viewer

The trace viewer displays a timeline that shows:

The trace viewer allows you to identify performance problems in your model, then take steps to resolve them. For example, at a high level, you can identify whether input or model training is taking the majority of the time. Drilling down, you can identify which ops take the longest to execute. Note that the trace viewer is limited to 1 million events per device.

Trace viewer interface

When you open the trace viewer, it appears displaying your most recent run:

image

This screen contains the following main elements:

  1. Timeline pane: Shows ops that the device and the host executed over time.
  2. Details pane: Shows additional information for ops selected in the Timeline pane.

The Timeline pane contains the following elements:

  1. Top bar: Contains various auxiliary controls.
  2. Time axis: Shows time relative to the beginning of the trace.
  3. Section and track labels: Each section contains multiple tracks and has a triangle on the left that you can click to expand and collapse the section. There is one section for every processing element in the system.
  4. Tool selector: Contains various tools for interacting with the trace viewer such as Zoom, Pan, Select, and Timing. Use the Timing tool to mark a time interval.
  5. Events: These show the time during which an op was executed or the duration of meta-events, such as training steps.
Sections and tracks

The trace viewer contains the following sections:

Events

Events within the timeline are displayed in different colors; the colors themselves have no specific meaning.

The trace viewer can also display traces of Python function calls in your TensorFlow program. If you use the tf.profiler.experimental.start API, you can enable Python tracing by using the ProfilerOptions namedtuple when starting profiling. Alternatively, if you use the sampling mode for profiling, you can select the level of tracing by using the dropdown options in the Capture Profile dialog.

image

GPU kernel stats

This tool shows performance statistics and the originating op for every GPU accelerated kernel.

image

The tool displays information in two panes:

Memory profile tool

The Memory Profile tool monitors the memory usage of your device during the profiling interval. You can use this tool to:

The memory profile tool displays data in three sections:

  1. Memory Profile Summary
  2. Memory Timeline Graph
  3. Memory Breakdown Table

Memory profile summary

This section displays a high-level summary of the memory profile of your TensorFlow program as shown below:

The memory profile summary has six fields:

  1. Memory ID: Dropdown which lists all available device memory systems. Select the memory system you want to view from the dropdown.
  2. #Allocation: The number of memory allocations made during the profiling interval.
  3. #Deallocation: The number of memory deallocations in the profiling interval
  4. Memory Capacity: The total capacity (in GiBs) of the memory system that you select.
  5. Peak Heap Usage: The peak memory usage (in GiBs) since the model started running.
  6. Peak Memory Usage: The peak memory usage (in GiBs) in the profiling interval. This field contains the following sub-fields:
    1. Timestamp: The timestamp of when the peak memory usage occurred on the Timeline Graph.
    2. Stack Reservation: Amount of memory reserved on the stack (in GiBs).
    3. Heap Allocation: Amount of memory allocated on the heap (in GiBs).
    4. Free Memory: Amount of free memory (in GiBs). The Memory Capacity is the sum total of the Stack Reservation, Heap Allocation, and Free Memory.
    5. Fragmentation: The percentage of fragmentation (lower is better). It is calculated as a percentage of (1 - Size of the largest chunk of free memory / Total free memory).

Memory timeline graph

This section displays a plot of the memory usage (in GiBs) and the percentage of fragmentation versus time (in ms).

image

The X-axis represents the timeline (in ms) of the profiling interval. The Y-axis on the left represents the memory usage (in GiBs) and the Y-axis on the right represents the percentage of fragmentation. At each point in time on the X-axis, the total memory is broken down into three categories: stack (in red), heap (in orange), and free (in green). Hover over a specific timestamp to view the details about the memory allocation/deallocation events at that point like below:

image

The pop-up window displays the following information:

Memory breakdown table

This table shows the active memory allocations at the point of peak memory usage in the profiling interval.

image

There is one row for each TensorFlow Op and each row has the following columns:

Pod viewer

The Pod Viewer tool shows the breakdown of a training step across all workers.

image

tf.data bottleneck analysis

The tf.data bottleneck analysis tool automatically detects bottlenecks intf.data input pipelines in your program and provides recommendations on how to fix them. It works with any program using tf.data regardless of the platform (CPU/GPU/TPU). Its analysis and recommendations are based on thisguide.

It detects a bottleneck by following these steps:

  1. Find the most input bound host.
  2. Find the slowest execution of a tf.data input pipeline.
  3. Reconstruct the input pipeline graph from the profiler trace.
  4. Find the critical path in the input pipeline graph.
  5. Identify the slowest transformation on the critical path as a bottleneck.

The UI is divided into three sections: Performance Analysis Summary,Summary of All Input Pipelines and Input Pipeline Graph.

Performance analysis summary

image

This section provides the summary of the analysis. It reports on slow tf.datainput pipelines detected in the profile. This section also shows the most input bound host and its slowest input pipeline with the max latency. Most importantly, it identifies which part of the input pipeline is the bottleneck and how to fix it. The bottleneck information is provided with the iterator type and its long name.

How to read tf.data iterator's long name

A long name is formatted as Iterator::<Dataset_1>::...::<Dataset_n>. In the long name, <Dataset_n> matches the iterator type and the other datasets in the long name represent downstream transformations.

For example, consider the following input pipeline dataset:

dataset = tf.data.Dataset.range(10).map(lambda x: x).repeat(2).batch(5)

The long names for the iterators from the above dataset will be:

Iterator Type Long Name
Range Iterator::Batch::Repeat::Map::Range
Map Iterator::Batch::Repeat::Map
Repeat Iterator::Batch::Repeat
Batch Iterator::Batch

Summary of all input pipelines

image

This section provides the summary of all input pipelines across all hosts. Typically there is one input pipeline. When using the distribution strategy, there is one host input pipeline running the program's tf.data code and multiple device input pipelines retrieving data from the host input pipeline and transferring it to the devices.

For each input pipeline, it shows the statistics of its execution time. A call is counted as slow if it takes longer than 50 μs.

Input pipeline graph

image

This section shows the input pipeline graph with the execution time information. You can use "Host" and "Input Pipeline" to choose which host and input pipeline to see. Executions of the input pipeline are sorted by the execution time in descending order which you can choose using the Rank dropdown.

image

The nodes on the critical path have bold outlines. The bottleneck node, which is the node with the longest self time on the critical path, has a red outline. The other non-critical nodes have gray dashed outlines.

In each node,Start Time indicates the start time of the execution. The same node may be executed multiple times, for example, if there is a Batch op in the input pipeline. If it is executed multiple times, it is the start time of the first execution.

Total Duration is the wall time of the execution. If it is executed multiple times, it is the sum of the wall times of all executions.

Self Time is Total Time without the overlapped time with its immediate child nodes.

"# Calls" is the number of times the input pipeline is executed.

Collect performance data

The TensorFlow Profiler collects host activities and GPU traces of your TensorFlow model. You can configure the Profiler to collect performance data through either the programmatic mode or the sampling mode.

Profiling APIs

You can use the following APIs to perform profiling.

# Profile from batches 10 to 15  
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir,  
                                             profile_batch='10, 15')  
# Train the model and use the TensorBoard Keras callback to collect  
# performance profiling data  
model.fit(train_data,  
          steps_per_epoch=20,  
          epochs=5,  
          callbacks=[tb_callback])  
tf.profiler.experimental.start('logdir')  
# Train the model here  
tf.profiler.experimental.stop()  
with tf.profiler.experimental.Profile('logdir'):  
    # Train the model here  
    pass  
# Start a profiler server before your model runs.  
tf.profiler.experimental.server.start(6009)  
# (Model code goes here).  
#  Send a request to the profiler server to collect a trace of your model.  
tf.profiler.experimental.client.trace('grpc://localhost:6009',  
                                      'gs://your_tb_logdir', 2000)

An example for profiling multiple workers: 

# E.g., your worker IP addresses are 10.0.0.2, 10.0.0.3, 10.0.0.4, and you  
# would like to profile for a duration of 2 seconds.  
tf.profiler.experimental.client.trace(  
    'grpc://10.0.0.2:8466,grpc://10.0.0.3:8466,grpc://10.0.0.4:8466',  
    'gs://your_tb_logdir',  
    2000)

Use the Capture Profile dialog to specify:

Profiling custom training loops

To profile custom training loops in your TensorFlow code, instrument the training loop with the tf.profiler.experimental.Trace API to mark the step boundaries for the Profiler.

The name argument is used as a prefix for the step names, the step_numkeyword argument is appended in the step names, and the _r keyword argument makes this trace event get processed as a step event by the Profiler.

As an example,

for step in range(NUM_STEPS):
    with tf.profiler.experimental.Trace('train', step_num=step, _r=1):
        train_data = next(dataset)
        train_step(train_data)

This will enable the Profiler's step-based performance analysis and cause the step events to show up in the trace viewer.

Make sure that you include the dataset iterator within thetf.profiler.experimental.Trace context for accurate analysis of the input pipeline.

The code snippet below is an anti-pattern:

for step, train_data in enumerate(dataset):
    with tf.profiler.experimental.Trace('train', step_num=step, _r=1):
        train_step(train_data)

Profiling use cases

The profiler covers a number of use cases along four different axes. Some of the combinations are currently supported and others will be added in the future. Some of the use cases are:

The table below provides a quick overview of the TensorFlow-supported use cases mentioned above:

Profiling API Local Remote Multiple workers Hardware Platforms
TensorBoard Keras Callback Supported Not Supported Not Supported CPU, GPU
tf.profiler.experimentalstart/stop API Supported Not Supported Not Supported CPU, GPU
tf.profiler.experimentalclient.trace API Supported Supported Supported CPU, GPU, TPU
Context manager API Supported Not supported Not Supported CPU, GPU

Best practices for optimal model performance

Use the following recommendations as applicable for your TensorFlow models to achieve optimal performance.

In general, perform all transformations on the device and ensure that you use the latest compatible version of libraries like cuDNN and Intel MKL for your platform.

Optimize the input data pipeline

Use the data from the [#input_pipeline_analyzer] to optimize your data input pipeline. An efficient data input pipeline can drastically improve the speed of your model execution by reducing device idle time. Try to incorporate the best practices detailed in theBetter performance with the tf.data APIguide and below to make your data input pipeline more efficient.

Also read the tf.data performance analysisguide for more guidance on optimizing your input pipeline.

Optimize data augmentation

When working with image data, make yourdata augmentationmore efficient by casting to different data types after applying spatial transformations, such as flipping, cropping, rotating, etc.

Use NVIDIA® DALI

In some instances, such as when you have a system with a high GPU to CPU ratio, all of the above optimizations may not be enough to eliminate bottlenecks in the data loader caused due to limitations of CPU cycles.

If you are using NVIDIA® GPUs for computer vision and audio deep learning applications, consider using the Data Loading Library (DALI) to accelerate the data pipeline.

Check theNVIDIA® DALI: Operationsdocumentation for a list of supported DALI ops.

Use threading and parallel execution

Run ops on multiple CPU threads with the tf.config.threading API to execute them faster.

TensorFlow automatically sets the number of parallelism threads by default. The thread pool available for running TensorFlow ops depends on the number of CPU threads available.

Control the maximum parallel speedup for a single op by usingtf.config.threading.set_intra_op_parallelism_threads. Note that if you run multiple ops in parallel, they will all share the available thread pool.

If you have independent non-blocking ops (ops with no directed path between them on the graph), use tf.config.threading.set_inter_op_parallelism_threads to run them concurrently using the available thread pool.

Miscellaneous

When working with smaller models on NVIDIA® GPUs, you can settf.compat.v1.ConfigProto.force_gpu_compatible=True to force all CPU tensors to be allocated with CUDA pinned memory to give a significant boost to model performance. However, exercise caution while using this option for unknown/very large models as this might negatively impact the host (CPU) performance.

Improve device performance

Follow the best practices detailed here and in theGPU performance optimization guideto optimize on-device TensorFlow model performance.

If you are using NVIDIA GPUs, log the GPU and memory utilization to a CSV file by running:

nvidia-smi
--query-gpu=utilization.gpu,utilization.memory,memory.total,
memory.free,memory.used --format=csv

Configure data layout

When working with data that contains channel information (like images), optimize the data layout format to prefer channels last (NHWC over NCHW).

Channel-last data formats improveTensor Coreutilization and provide significant performance improvements especially in convolutional models when coupled with AMP. NCHW data layouts can still be operated on by Tensor Cores, but introduce additional overhead due to automatic transpose ops.

You can optimize the data layout to prefer NHWC layouts by settingdata_format="channels_last" for layers such as tf.keras.layers.Conv2D,tf.keras.layers.Conv3D, andtf.keras.layers.RandomRotation.

Use tf.keras.backend.set_image_data_format to set the default data layout format for the Keras backend API.

Max out the L2 cache

When working with NVIDIA® GPUs, execute the code snippet below before the training loop to max out the L2 fetch granularity to 128 bytes.

import ctypes

_libcudart = ctypes.CDLL('libcudart.so')
# Set device limit on the current device
# cudaLimitMaxL2FetchGranularity = 0x05
pValue = ctypes.cast((ctypes.c_int*1)(), ctypes.POINTER(ctypes.c_int))
_libcudart.cudaDeviceSetLimit(ctypes.c_int(0x05), ctypes.c_int(128))
_libcudart.cudaDeviceGetLimit(pValue, ctypes.c_int(0x05))
assert pValue.contents.value == 128

Configure GPU thread usage

The GPU thread mode decides how GPU threads are used.

Set the thread mode to gpu_private to make sure that preprocessing does not steal all the GPU threads. This will reduce the kernel launch delay during training. You can also set the number of threads per GPU. Set these values using environment variables.

import os

os.environ['TF_GPU_THREAD_MODE']='gpu_private'
os.environ['TF_GPU_THREAD_COUNT']='1'

Configure GPU memory options

In general, increase the batch size and scale the model to better utilize GPUs and get higher throughput. Note that increasing the batch size will change the model’s accuracy so the model needs to be scaled by tuning hyperparameters like the learning rate to meet the target accuracy.

Also, use tf.config.experimental.set_memory_growth to allow GPU memory to grow to prevent all the available memory from being fully allocated to ops that require only a fraction of the memory. This allows other processes which consume GPU memory to run on the same device.

To learn more, check out theLimiting GPU memory growthguidance in the GPU guide to learn more.

Miscellaneous

Additional resources

Known limitations

Profiling multiple GPUs on TensorFlow 2.2 and TensorFlow 2.3

TensorFlow 2.2 and 2.3 support multiple GPU profiling for single host systems only; multiple GPU profiling for multi-host systems is not supported. To profile multi-worker GPU configurations, each worker has to be profiled independently. From TensorFlow 2.4 multiple workers can be profiled using thetf.profiler.experimental.client.trace API.

CUDA® Toolkit 10.2 or later is required to profile multiple GPUs. As TensorFlow 2.2 and 2.3 support CUDA® Toolkit versions only up to 10.1, you need to create symbolic links to libcudart.so.10.1 and libcupti.so.10.1:

sudo ln -s /usr/local/cuda/lib64/libcudart.so.10.2 /usr/local/cuda/lib64/libcudart.so.10.1
sudo ln -s /usr/local/cuda/extras/CUPTI/lib64/libcupti.so.10.2 /usr/local/cuda/extras/CUPTI/lib64/libcupti.so.10.1