rahul003 commented on a change in pull request #8762: Gradient compression faq URL: https://github.com/apache/incubator-mxnet/pull/8762#discussion_r152460220
########## File path: docs/faq/gradient_compression.md ########## @@ -0,0 +1,98 @@ +# Gradient Compression + +Gradient Compression reduces communication bandwidth to make distributed training with GPUs more scalable and efficient without significant loss in convergence rate or accuracy. + + +## Benefits + +**Increased Speed** + +For tasks like acoustic modeling in speech recognition (like in Alexa), the gradient compression capability is observed to speedup training by about 2 times, depending on the size of the model and the network bandwidth of the instance. Bigger models see larger speedup with gradient compression. + +**Minimal Accuracy Loss** + +Gradient compression uses the approach of delaying the synchronization of weight updates which are small. Although small weight updates might not be sent for that batch, this information is not discarded. Once the weight updates for this location accumulate to become a larger value, they will be propagated. Since there is no information loss, but only delayed updates, it does not lead to a significant loss in accuracy or convergence rate. In distributed training experiments[1], it is observed a loss of accuracy as low as 1% for this technique. + + +## When to Use Gradient Compression + +When training models whose architectures include large fully connected components, it can be helpful to use gradient compression. For larger models, the communication cost becomes a major factor. Such models stand to benefit greatly with gradient compression. + + +### GPU versus CPU + +The greatest benefits from gradient compression are realized when using GPUs for both single-node multi-GPU and multi-node (single or multi-GPU) distributed training. Training on CPU would provide a lower compute density per compute node as compared to the massive compute density per compute node on a GPU. Due to this, the required communication bandwidth for CPU-based nodes during training is not as high as for GPU-based nodes. Hence, the benefits of gradient compression are lower for CPU-based nodes as compared to GPU-based nodes. + + +### Network Latency + +Benefits of gradient compression can be found when using distributed training with network connected nodes. Depending on the network latency between nodes and the model's size, these can contribute to slow performance such that gradient compression may provide speed improvements. + +You may not want to use gradient compression if you have low latency network communication. + + +### Model Size + +Distributed training involves synchronization of weights after each batch. Larger models have much higher communication costs during training, hence such models stand to benefit much more from gradient compression. +When running distributed training with gradient compression, the quantize and dequantize operations happen on CPU parallelized with OpenMP. For smaller models, when training on GPUs, it helps to set `OMP_NUM_THREADS=1` on each node, so that the overhead of launching OMP threads doesn't cause the compression and decompression to be slow. + +### Model Architecture + +The communication bandwidth requirements during training vary across various neural network architectures and hence the benefits of gradient compression vary accordingly. + +In networks which have significant fully connected components, since such layers have low compute cost on GPUs, communication becomes a bottleneck limiting the speed of distributed training. Gradient compression can help reduce the communication cost, and thus speed up training in such cases. We have observed speedup of about 2x on large fully connected neural networks. Models like AlexNet and VGG have large fully connected components as part of the network, hence stand to benefit from gradient compression. Long Short-Term Memory architectures require more communication bandwidth, so they also exhibit speed improvements with gradient compression. + +Architectures like Convolutional Neural Networks on the other hand have a higher compute cost, in which case some communication can be parallelized with compute. Since communication is not the bottleneck in such networks, gradient compression doesn't help much. + + +### Single Node Gradient Compression + +When the training is configured to use device to device communication on a single node with multiple GPUs, gradient compression can be used to reduce the cost communication. This can provide about 20% speedup for large models using older generation architectures. However, speed benefits may be negligible on a machine with a newer generation architecture where GPUs can communicate at low latency. + + +## Deep Neural Networks and Sparse Data + +It is well-known that typically the weights of a fully connected DNN (Deep Neural Networks) are sparsely distributed with most weights close to zero, and so it is not surprising that sub-gradients are also sparse [1]. Since sub-gradients are computed from a small part of the training data, they are even sparser than the weights. Hence, only a small fraction of the weights is required to be updated after each mini-batch. In other words, elements of the gradient that are near zero can safely be delayed longer than the typical mini-batch size. The sub-gradients are compressed significantly by considering only gradient elements whose absolute values exceed a threshold. The resulting sparse gradients are then encoded using 2-bit quantization thereby reducing the communication bandwidth. The delayed gradient values are aggregated into a gradient residual which is communicated when it reaches the threshold. + + +## Technical Implementation + +For data-parallel training, the model is replicated across compute nodes with the weight-updates synchronized across all the model replicas. The massive local computational density of the GPU nodes increases the required communication bandwidth for weight updates across model replicas in data-parallel distributed training. Instead of the uniform update-rate of weights imposed by the mini-batch size, the gradient compression capability controls the rate of weight-update per individual weight. Gradient compression uses the approach of delaying synchronization of weights whose updates (aka gradients) are small, and compressing the weight-updates which are synchronized. This reduction in communication bandwidth enables distributed training to be more efficient and scalable to more GPU nodes without significant loss in convergence rate or accuracy. Review comment: Please remove the current Technical Implementation and the Deep neural networks sections and create a new section for the below. I think its valuable to include the details in technical implementation because this is important information which a user would otherwise get only by reading and understanding the code. ## Approach The idea behind gradient compression comes from two observations. Firstly, the gradients of weights computed for a small mini-batch of training data, when training large neural networks, are typically sparse. Only a small fraction of the weights have significant updates after each mini-batch. The synchronization of updates that are near zero can safely be delayed longer than the typical mini-batch size. This essentially means that the rate of weight-update can vary depending on the value of an individual weight. Secondly, gradients can be compressed significantly by considering only those gradient elements whose absolute values exceed a threshold and then quantizing them to use lower bits per gradient value. By compressing the gradients, we can reduce communication bandwidth. The delayed gradient values, in the form of quantization error and values that don't meet the threshold, are aggregated into a gradient residual which is communicated when it reaches the threshold. ## Technical Implementation ### 2 bit Quantization Currently the supported type of quantization uses 2bits for each gradient value. Any positive value greater than or equal to the threshold sets two bits as `11`, any negative value whose absolute value is greater or equal to the threshold sets two bits as`10`, and others are set to `00`. This enables us to store 16 quantized gradients into one float. The error in quantization, which is `original_value - quantized_value` is stored in a gradient residual. ### Types of kvstore Supported types of kvstore are `device` and all distributed kvstores (like `dist_sync`, `dist_async`, `dist_sync_device`). When kvstore is `device`, the communication between GPUs is compressed. Please note that this increases the memory usage of GPUs because of the additional residual stored. When using a distributed kvstore, worker to server communication is compressed. Server to worker communication and device to device communication are not compressed to avoid multiple levels of compression. In this case, compression and decompression happen on CPU, gradient residuals will be stored on CPU. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
