This guide is for TensorFlow users utilizing GPUs to improve model performance. Using the TensorFlow Profiler as the main tool to gain insight into performance, this guide will help you debug when one or more of your GPUs are underutilized. A quickstart guide to the TensorFlow Profiler can be found in the TensorFlow Profiler tutorial, and additional ways to obtain a profile are documented in the Optimize TensorFlow performance using the Profiler guide.
Keep in mind that offloading computations to GPU might not always be beneficial, particularly for small models. There are overheads due to data transfer between host (CPU) and device (GPU), as well as overheads due to the latency involved when the host launches GPU kernels. Good performance is achieved when the host successfully keeps the GPU occupied by offloading enough work.
Performance Optimization Workflow
This guide outlines how to debug performance issues starting with a single GPU, then moving to a single host with multiple GPUs. It is recommended to debug performance issues in this order. For example, if you are using a TensorFlow distribution strategy to train a model on a single host with multiple GPUs and notice suboptimal GPU utilization, you should first optimize and debug performance for 1 GPU, before debugging the multi-GPU system. The recommended order is as follows:
- Optimize and debug performance on 1 GPU
- Check if input pipeline is a bottleneck
- Debug performance of 1 GPU
- Enable fp16 and optionally enable XLA
- Optimize and debug performance on multi-GPU single host
As a baseline for getting performant code on GPU, this guide assumes you are
tf.function. The Keras compile/fit API will utilize
tf.function automatically under the hood. When writing a custom training loop,
refer to this guide on how to
The next sections discuss suggested approaches for each of the scenarios above to help identify and fix performance bottlenecks.
Optimize Performance on 1 GPU
In an ideal case, your program should have high GPU utilization, minimal CPU (host) to GPU (device) communication, and no overhead from the input pipeline. The first step to analyzing performance is to get a profile for a model running with one GPU.
The Overview Page of the TensorFlow Profiler provides an idea of how far away your program is from the ideal scenario.
The key numbers to look for on the overview page are:
- How much of the step time is from actual device execution
- The percentage of ops placed on device vs host
- How many kernels use fp16
Achieving optimal performance means maximizing these numbers in all three cases. To get an in-depth understanding of your program, you will need to be familiar with the TensorFlow Profiler trace viewer. The sections below show some common trace viewer patterns that you should look for when diagnosing performance bottlenecks.
Below is an image of a model trace view running on 1 GPU. From the Tensorflow Name Scope and Tensorflow Ops sections, you can identify different parts of the model, like the forward pass, the loss function, backward pass/gradient calculation, and the optimizer weight update. You can also see the operations running on the GPU next to each Stream, which refer to CUDA streams. Each stream is used for specific tasks. In this trace, Stream#118 is used to launch compute kernels and device to device copies. Stream#119 is used for host to device copy and Stream#120 for device to host copy.
This trace shows common characteristics of a performant model. For example, the GPU compute timeline (Stream#118) looks busy with very few gaps. There are minimal copies from host to device (Stream #119) and from device to host (Stream #120), as well as minimal gaps between steps. When you run the TensorFlow Profiler for your program, you might not see these ideal characteristics in your trace view. The rest of this guide covers common scenarios and how to fix them.
Debug Input Pipeline
The first step in GPU performance debugging is to determine if your program is input-bound. The easiest way to figure this out is to use the TensorFlow Profiler’s Input-Pipeline Analyzer, which gives an overview of time spent in the input pipeline.
The following are potential actions you can take if your input-pipeline contributes significantly to step time:
- Refer to the
tf.dataspecific guide to learn how to debug your input pipeline.
- Another quick way to check if the input pipeline is the bottleneck is to use randomly generated input data that does not need any pre-processing. Here is an example of using this technique for a ResNet model. If the input pipeline is optimal you should see similar performance with real data and with generated random/synthetic data. The only overhead in the synthetic data case will be due to input data copy which again can be prefetched and optimized.
Debug Performance of 1 GPU
There are several factors that can contribute to low GPU utilization. Below are some scenarios commonly observed when looking at the trace viewer and potential solutions.
Analyze Gaps Between Steps
A common observation when your program is not running optimally is gaps between training steps. In the image below, there is a large gap between steps 8 and 9, meaning that the GPU is idle during that time.
If your trace viewer shows large gaps between steps, this could be an indication
that your program is input bound. In that case you should refer to the previous
section on debugging your input pipeline if you haven’t already done so.
However, even with an optimized input pipeline, you can still see gaps between
the end of one step and the start of another due to CPU thread contention.
tf.data makes use of background threads to parallelize pipeline processing.
These threads may interfere with GPU host-side activity that happens at the
beginning of each step, such as copying data or scheduling GPU operations.
If you see large gaps on the host side, which schedules these ops on the GPU,
you can set the environment variable
ensures that GPU kernels are launched from their own dedicated threads, and
don't get queued behind
Gaps between steps can also be caused by metric calculations, Keras callbacks,
or ops outside of
tf.function that run on the host. These ops don’t have as
good performance as the ops inside a TensorFlow graph. Additionally, some of
these ops run on the CPU and copy tensors back and forth from the GPU.
If after optimizing your input pipeline you still notice gaps between steps in
the trace viewer, you should look at the model code between steps, and see if
disabling callbacks/metrics improves performance. Some details of these ops are
also on the trace viewer (both device and host side).The recommendation in this
scenario is to amortize the overhead of these ops by executing them after a
fixed number of steps instead of every step. When using the
compile method in
tf.keras API, setting the
experimental_steps_per_execution flag does
this automatically. For custom training loops, use
Achieve Higher Device Utilization
Small GPU Kernels and Host Kernel Launch Delays
The host enqueues kernels to be run on the GPU, but there is a latency (around 20-40 μs) involved before kernels are actually executed on the GPU. In an ideal case, the host enqueues enough kernels on the GPU such that the GPU spends most of its time executing, rather than waiting on the host to enqueue more kernels.
The TensorFlow Profiler's Overview Page shows how much time the GPU was idle due to waiting on the host to launch kernels. In the image below, the GPU is idle for about 10% of the step time waiting on kernels to be launched.
The trace viewer for this same program shows small gaps between kernels where the host is busy launching kernels on the GPU.
By launching a lot of small ops on the GPU (like a scalar add for example), the host might not keep up with the GPU. The Tensorflow Stats page for the same TensorFlow Profile shows 126,224 Mul operations taking 2.77 seconds. Thus, each kernel is about 21.9 μs, which is very small (around the same time as launch latency) and can potentially result in host kernel launch delays.
If your trace viewer shows many small gaps between ops on the GPU like the image above, you can:
- Concatenate small tensors and use vectorized operations or use a larger batch size to make each kernel launched do more work, which will keep the GPU busy for longer.
- Make sure you’re using
tf.functionto create TF graphs and not running operations in a pure eager mode (Using
tf.keras.Model.compileautomatically does this).
- Fuse kernels using XLA. For more details, see the section below on how to enable XLA to get higher performance. This is an experimental feature, but leads to high device utilization.
Tensorflow Op Placement
The TensorFlow Profiler Overview Page tells you the percentage of ops placed on the host vs device (you can also verify the placement of specific ops by looking at the trace viewer). Like in the image below, you want the percentage of ops on the host to be very small compared to the device.
Ideally most of the compute intensive ops should be placed on the GPU. To find
out which devices the operations and tensors in your model are assigned to, set
tf.debugging.set_log_device_placement(True) as the first statement of your
program. Note that in some cases, even if you specify an op to be placed on a
particular device, its implementation might override this condition
tf.unique). Even for single GPU training, specifying a distribution
strategy, such as
tf.distribute.OneDeviceStrategy, can result in more
deterministic placement of ops on your device.
One reason for having the majority of ops placed on the GPU is to prevent excessive memory copies between host and device (memory copies for model input/output data between host and device are expected). An example of excessive copying can be seen in the trace view below on GPU streams #167, #168, and #169.
These copies can sometimes hurt performance if they block GPU kernels from executing. Memory copy operations in the trace viewer have more information about the ops that are the source of these copied tensors, but it might not always be easy to associate a memCopy with an op. In these cases, it is helpful to look at the ops nearby to see if the memory copy happens at the same location in every step.
More Efficient Kernels on GPUs
Once your program's GPU utilization is acceptable, the next step is to look into increasing the efficiency of the GPU kernels by utilizing Tensor Cores or fusing ops.
Utilizing Tensor Cores
Modern GPUs have specialized tensor cores that can significantly improve the performance of kernels that are eligible. The GPU kernel stats page indicates which GPU kernels are Tensor Core eligible, and which kernels are using the Tensor Core. Enabling fp16 (see Enabling Mixed Precision section below) is one way to make your program’s General Matrix Multiply (GEMM) kernels (matmul ops) utilize the Tensor Core.
For other detailed recommendations on how to make kernels efficient for GPUs, refer to the NVIDIA Deep Learning Performance guide, which covers a variety of techniques you can experiment with such as using NCHW vs NHWC format to represent inputs, or making the input dimensions to be a multiple of 8.
tf.xla.xperimental_compile feature, TensorFlow can fuse smaller ops
to form bigger kernels leading to significant performance gains. More details
are discussed in the Enable XLA section below.
Enable fp16 and XLA
After following the above steps, enabling mixed-precision and XLA are two optional steps you can take to improve performance further. The suggested approach is to enable them one by one and verify that the performance benefits are as expected.
Enabling Mixed Precision
The TensorFlow Mixed Precision guide shows how to enable fp16 precision on GPUs. When it comes to realizing the performance benefits of fp16 there are a few pointers to keep in mind.
Using Optimal fp16 Kernels
With fp16 enabled, your program’s matrix multiplications (GEMM) kernels, should use the corresponding fp16 version that utilizes the Tensor Cores. However, in some cases, this does not happen and you do not see the expected speedup from enabling fp16, as your program falls back to the inefficient implementation instead.
The GPU kernel stats page shows which ops are Tensor Core eligible and which kernels are actually using the efficient Tensor Core. The NVIDIA guide on deep learning performance contains additional suggestions on how to leverage Tensor Cores. Additionally, the benefits of fp16 will also show in kernels that were previously memory bound, as now the operation will take half the time.
Dynamic vs Static Loss Scaling
Loss scaling is necessary when using fp16 to prevent underflow due to low precision. There are two types of loss scaling, dynamic and static, both of which are explained in greater detail in the Mixed Precision guide. When trying to optimize performance, it is important to remember that dynamic loss scaling can introduce additional conditional ops that run on the host, and lead to gaps that will be visible between steps in the trace viewer. On the other hand, static loss scaling does not have such overheads and can be a better option in terms of performance with the catch that you need to specify the correct static-loss scale value.
As the final step to getting the best performance with a single GPU, you can experiment with enabling XLA, which will fuse ops and lead to better device utilization and a lower memory footprint. For details on how to enable XLA in your program, refer to the XLA: Optimizing Compiler for Machine Learning guide.
Optimize Performance on Multi-GPU Single Host
tf.distribute.MirroredStrategy API can be used to scale model training
from 1 GPU to multiple GPUs on a single host. To learn more about how to do
distributed training with Tensorflow, please refer to the
Distributed training with Keras
guide. Although the transition from one GPU to multiple GPUs should ideally be
scalable out of the box, you can sometimes encounter performance issues.
When going from training with a single GPU to multiple GPUs on the same host, ideally you should see performance scaling with only the additional overhead of gradient communication and increased host thread utilization. Because of this overhead, you will not see an exact 2x speedup if you move from 1 to 2 GPUs for example. The trace view below shows an example of the extra communication overhead when training on multiple GPUs. There is some overhead to concatenate the gradients, communicate them across replicas, and split them before doing the weight update.
The following checklist will help you to achieve better performance when optimizing performance in the multi-GPU scenario:
- Try to maximize the batch size, which will lead to higher device utilization and amortize the costs of communication across multiple GPUs. Using the memory profiler helps get a sense of how close your program is to peak memory utilization. Note that while a higher batch size can affect convergence, this is usually outweighed by the performance benefits.
- When moving from a single GPU to multiple GPUs, the same host now has to process much more input data. So after (1) it is recommended to re-check the input pipeline performance and make sure it is not a bottleneck.
- Check the GPU timeline in your program's trace view to see if there are uncecessary AllReduce calls, as this results in a synchronization across all devices. In the trace view shown above, the AllReduce is done via the NCCL kernel, and there is only one NCCL call on each GPU for the gradients on each step.
- Check for unnecessary D2H, H2D and D2D copy operations and see if they can be minimized.
- Check the step time to make sure each replica is doing the same work. It can happen that one GPU (typically GPU0) is oversubscribed because the host mistakenly ends up putting more work on it.
- Lastly, check the training step across all GPUs in your trace view for any ops that are executing sequntially. This usually happens when your program includes control dependencies from one GPU to another. Debugging performance in this situation has been solved on a case-by-case basis in the past. If you observe this behavior in your program, file a Github issue with images of your trace view.
Optimize Gradient AllReduce
When training with a synchronous strategy, each device receives a portion of the
input data. After computing the forward and backwards passes through the model,
the gradients calculated on each device need to be aggregated and reduced. This
gradient AllReduce happens after the gradient calculation on each device, and
before the optimizer updates the model weights. Each GPU first concatenates the
gradients across the model layers, communicates them across GPUs using
tf.distribute.NcclAllReduce is the default),
and then returns the gradients after reduction per layer. The optimizer will use
these reduced gradients to update the weights of your model. Ideally, this
process should happen at the same time on all GPUs to prevent any overheads. The
time to AllReduce should be approximately the same as:
(number of parameters * 4bytes)/ (communication bandwidth)
This calculation is useful as a quick check to understand whether the
performance you see when running a distributed training job is as expected, or
if you need to do further performance debugging. You can get the number of
parameters in your model from
Note that each model parameter is 4 bytes since Tensorflow uses fp32 to communicate gradients. Even when you have enabled fp16, NCCL AllReduce utilizes fp32 parameters. In the future, Tensorflow will support AllReduce operations using fp16, as well as pipelining the gradient AllReduce so it overlaps with the gradient computation.
To see benefits of scaling, the step-time needs to be much higher compared to these overheads. One way to achieve this is to use a higher-batch size since batch size affects step time, but does not affect the communication overheads.
GPU Host Thread Contention
When running multiple GPUs, the CPU’s job is to keep all of the devices busy by efficiently launching GPU kernels across the devices. However, when there are a lot of independent operations that the CPU can schedule on one GPU, the CPU can decide to use a lot of its host threads to keep one GPU busy, and then launch kernels on another GPU in a non-deterministic order. This can cause a skew or negative scaling, which can hurt performance.
The trace viewer below shows the overhead when the CPU staggers GPU kernel launchs inefficiently, as GPU1 is idle and then starts running ops after GPU2 has started.
The trace view for the host shows that the host is launching kernels on GPU2 before launching them on GPU1 (note that the below tf_Compute* ops are not indicative of CPU threads).
If you see this staggering of GPU kernels in your program’s trace view, the recommended action is to:
- Set the TensorFlow environment variable
gpu_private. This environment variable will tell the host to keep threads for a GPU private.
- By default,
TF_GPU_THREAD_MODE=gpu_privatesets the number of threads to 2, which is sufficient in most cases. However, that number can be changed by setting the TensorFlow environment variable
TF_GPU_THREAD_COUNTto the desired number of threads.