Currently I am studying the usage of Apache Spark 3.0 with Rapids GPU Acceleration. In the official spark-rapids docs I came across this page which states:
There are cases where you may want to get access to the raw data on the GPU, preferably without copying it. One use case for this is exporting the data to an ML framework after doing feature extraction.
To me this sounds as if one could make data that is already available on the GPU from some upstream Spark ETL process directly available to a framework such as Tensorflow or PyTorch. If this is the case how can I access the data from within any of these frameworks? If I am misunderstanding something here, what is the quote exactly referring to?
The link you references really only allows you to get access to the data still sitting on the GPU, but using that data in another framework, like Tensorflow or PyTorch is not that simple.
TL;DR; Unless you have a library explicitly setup to work with the RAPIDS accelerator you probably want to run your ETL with RAPIDS, then save it, and launch a new job to train your models using that data.
There are still a number of issues that you would need to solve. We have worked on these in the case of XGBoost, but it has not been something that we have tried to tackle for Tensorflow or PyTorch yet.
The big issues are
Getting the data to the correct process. Even if the data is on the GPU, because of security, it is tied to a given user process. PyTorch and Tensorflow generally run as python processes and not in the same JVM that Spark is running in. This means that the data has to be sent to the other process. There are several ways to do this, but it is non-trivial to try and do it as a zero-copy operation.
The format of the data is not what Tensorflow or PyTorch want. The data for RAPIDs is in an arrow compatible format. Tensorflow and PyTorch have APIs for importing data in standard formats from the CPU, but it might take a bit of work to get the data into a format that the frameworks want and to find an API to let you pull it in directly from the GPU.
Sharing GPU resources. Spark only recently added in support for scheduling GPUs. Prior to that people would just launch a single spark task per executor and a single python process so that the python process would own the entire GPU when doing training or inference. With the RAPIDS accelerator the GPU is not free any more and you need a way to share the resources. RMM provides some of this if both libraries are updated to use it and they are in the same process, but in the case of Pytorch and and Tensoflow they are typically in python processes so figuring out how to share the GPU is hard.
What strategies and forms of parallelization are feasible and available for training and serving a neural network?:
inside a machine across cores (e.g. GPU / TPU / CPU)
across machines on a network or a rack
I'm also looking for evidence for how they may also be used in e.g. TensorFlow, PyTorch or MXNet.
Training
To my knowledge, when training large neural networks on large datasets, one could at least have:
Different cores or machines operate on different parts of the graph ("graph splitting"). E.g. backpropagation through the graph itself can be parallelized e.g. by having different layers hosted on different machines since (I think?) the autodiff graph is always a DAG.
Different cores or machines operate on different samples of data ("data splitting"). In SGD, the computation of gradients across batches or samples can also be parallelized (e.g. the gradients can be combined after computing them independently on different batches). I believe this is also called gradient accumulation (?).
When is each strategy better for what type of problem or neural network? Which modes are supported by modern libraries? and can one combine all four (2x2) strategies?
On top of that, I have read about:
Asynchronous training
Synchronous training
but I don't know what exactly that refers to, e.g. is it the computation of gradients on different data batches or the computation of gradients on different subgraphs? Or perhaps it refers to something else altogether?
Serving
If the network is huge, prediction / inference may also be slow, and the model may not fit on a single machine in memory at serving time. Are there any known multi-core and multi-node prediction solutions that work that can handle such models?
Training
In general, there are two strategies of parallelizing model training: data parallelism and model parallelism.
1. Data parallelism
This strategy splits training data into N partitions, each of which will be trained on different “devices” (different CPU cores, GPUs, or even machines). In contrast to training without data parallelism which produces one gradient per minibatch, we now have N gradients for each minibatch step. The next question is how we should combine these N gradients.
One way to do it is by averaging all the N gradients and then updating the model parameters once based on the average. This technique is called synchronous distributed SGD. By doing the average, we have a more accurate gradient, but with a cost of waiting all the devices to finish computing its own local gradient.
Another way is by not combining the gradients — each gradient will instead be used to update the model parameters independently. So, there will be N parameter updates for each minibatch step, in contrast to only one for the previous technique. This technique is called asynchronous distributed SGD. Because it doesn't have to wait other devices to finish, the async approach will take less time to complete a minibatch step than the sync approach will do. However, the async approach will produce a more noisy gradient, so it might need to complete more minibatch steps to catch up with the performance (in terms of loss) of the sync approach.
There are many papers proposing some improvements and optimizations on either approach, but the main idea is generally the same as described above.
In the literature there's been some disagreement on which technique is better in practice. At the end most people now settle on the synchronous approach.
Data Parallelism in PyTorch
To do synchronous SGD, we can wrap our model with torch.nn.parallel.DistributedDataParallel:
from torch.nn.parallel import DistributedDataParallel as DDP
# `model` is the model we previously initialized
model = ...
# `rank` is a device number starting from 0
model = model.to(rank)
ddp_model = DDP(model, device_ids=[rank])
Then we can train it similarly. For more details, you can refer to the official tutorial.
For doing asynchronous SGD in PyTorch, we need to implement it more manually since there is no wrapper similar to DistributedDataParallel for it.
Data Parallelism in TensorFlow/Keras
For synchronous SGD, we can use tf.distribute.MirroredStrategy to wrap the model initalization:
import tensorflow as tf
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
model = Model(...)
model.compile(...)
Then we can train it as usual. For more details, you can refer to the official guides on Keras website and TensorFlow website.
For asynchronous SGD, we can use tf.distribute.experimental.ParameterServerStrategy similarly.
2. Model Parallelism
This strategy splits the model into N parts, each of which will be computed on different devices. A common way to split the model is based on layers: different sets of layers are placed on different devices. But we can also split it more intricately depending on the model architecture.
Model Parallelism in TensorFlow and PyTorch
To implement model parallelism in either TensorFlow or PyTorch, the idea is the same: to move some model parameters into a different device.
In PyTorch we can use torch.nn.Module.to method to move a module into a different device. For example, suppose we want to create two linear layers each of which is placed on a different GPU:
import torch.nn as nn
linear1 = nn.Linear(16, 8).to('cuda:0')
linear2 = nn.Linear(8, 4).to('cuda:1')
In TensorFlow we can use tf.device to place an operation into a specific device. To implement the PyTorch example above in TensorFlow:
import tensorflow as tf
from tensorflow.keras import layers
with tf.device('/GPU:0'):
linear1 = layers.Dense(8, input_dim=16)
with tf.device('/GPU:1'):
linear2 = layers.Dense(4, input_dim=8)
For more details you can refer to the official PyTorch tutorial; or if you use TensorFlow you can even use a more high-level library like mesh.
3. Hybrid: Data and Model Parallelism
Recall that data parallelism only splits the training data, whereas model parallelism only splits the model structures. If we have a model so large that even after using either parallelism strategy it still doesn't fit in the memory, we can always do both.
In practice most people prefer data parallelism to model parallelism since the former is more decoupled (in fact, independent) from the model architecture than the latter. That is, by using data parallelism they can change the model architecture as they like, without worrying which part of the model should be parallelized.
Model Inference / Serving
Parallelizing model serving is easier than parallelizing model training since the model parameters are already fixed and each request can be processed independently. Similar to scaling a regular Python web service, we can scale model serving by spawning more processes (to workaround Python's GIL) in a single machine, or even spawning more machine instances.
When we use a GPU to serve the model, though, we need to do more work to scale it. Because of how concurrency is handled differently by a GPU compared to a CPU, in order to maximize the performance, we need to do inference request batching. The idea is when a request comes, instead of immediately processing it, we wait some timeout duration for other requests to come. When the timeout is up, even if the number of requests is only one, we batch them all to be processed on the GPU.
In order to minimize the average request latency, we need to find the optimal timeout duration. To find it we need to observe that there is a trade-off between minimizing the timeout duration and maximizing the number of batch size. If the timeout is too low, the batch size will be small, so the GPU will be underutilized. But if the timeout is too high, the requests that come early will wait too long before they get processed. So, the optimal timeout duration depends on the model complexity (hence, the inference duration) and the average requests per second to receive.
Implementing a scheduler to do request batching is not a trivial task, so instead of doing it manually, we'd better use TensorFlow Serving or PyTorch Serve which already supports it.
To learn more about parallel and distributed learning, you can read this review paper.
As the question is quite broad, I'll try to shed a little different light and touch on different topics than what was shown in
#Daniel's in-depth answer.
Training
Data parallelization vs model parallelization
As mentioned by #Daniel data parallelism is used way more often and is easier to do correctly. Major caveat of model parallelism is the need to wait for part of neural network and synchronization between them.
Say you have a simple feedforward 5 layer neural network spread across 5 different GPUs, each layer for one device. In this case, during each forward pass each device has to wait for computations from the previous layers. In this simplistic case, copying data between devices and synchronization would take a lot longer and won't bring benefits.
On the other hand, there are models better suited for model parallelization like Inception networks, see picture below:
Here you can see 4 independent paths from previous layer which could go in parallel and only 2 synchronization points (Filter concatenation and Previous Layer).
Questions
E.g. backpropagation through the graph itself can be parallelized e.g.
by having different layers hosted on different machines since (I
think?) the autodiff graph is always a DAG.
It's not that easy. Gradients are calculated based on the loss value (usually) and you need to know gradients of deeper layers to calculate gradients for the more shallow ones. As above, if you have independent paths it's easier and may help, but it's way easier on a single device.
I believe this is also called gradient accumulation (?)
No, it's actually reduction across multiple devices. You can see some of that in PyTorch tutorial. Gradient accumulation is when you run your forward pass (either on single or multiple devices) N times and backpropagate (the gradient is kept in the graph and the values are added during each pass) and optimizer only makes a single step to change neural network's weights (and clears the gradient). In this case, loss is usually divided by the number of steps without optimizer. This is used for more reliable gradient estimation, usually when you are unable to use large batches.
Reduction across devices looks like this:
This is all-reduce in data parallelization, each device calculates the values which are send to all other devices and backpropagated there.
When is each strategy better for what type of problem or neural
network?
Described above, data parallel is almost always fine if you have enough of data and the samples are big (up to 8k samples or more can be done at once without very big struggle).
Which modes are supported by modern libraries?
tensorflow and pytorch both support either, most modern and maintained libraries have those functionalities implemented one way or another
can one combine all four (2x2) strategies
Yes, you can parallelize both model and data across and within machines.
synchronous vs asynchronous
asynchronous
Described by #Daniel in brief, but it's worth mentioning updates are not totally separate. That would make little sense, as we would essentially train N different models based on their batches.
Instead, there is a global parameter space, where each replica is supposed to share calculated updates asynchronously (so forward pass, backward, calculate update with optimizer and share this update to global params).
This approach has one problem though: there is no guarantee that when one worker calculated forward pass another worker updated the parameters, so the update is calculated with respect to old set of params and this is called stale gradients. Due to this, convergence might be hurt.
Other approach is to calculate N steps and updates for each worker and synchronize them afterwards, though it's not used as often.
This part was based on great blogpost and you should definitely read it if interested (there is more about staleness and some solutions).
synchronous
Mostly described previously, there are different approaches but PyTorch gathers output from network and backpropagates on them (torch.nn.parallel.DistributedDataParallel)[https://pytorch.org/docs/stable/nn.html#torch.nn.parallel.DistributedDataParallel]. BTW. You should solely this (no torch.nn.DataParallel) as it overcomes Python's GIL problem.
Takeaways
Data parallelization is always almost used when going for speed up as you "only" have to replicate neural network on each device (either over the network or within single machine), run part of batch on each during the forward pass, concatenate them into a single batch (synchronization) on one device and backpropagate on said.
There are multiple ways to do data parallelization, already introduced by #Daniel
Model parallelization is done when the model is too large to fit on single machine (OpenAI's GPT-3 would be an extreme case) or when the architecture is suited for this task, but both are rarely the case AFAIK.
The more and the longer parallel paths the model has (synchronization points), the better it might be suited for model parallelization
It's important to start workers at similar times with similar loads in order not to way for synchronization processes in synchronous approach or not to get stale gradients in asynchronous (though in the latter case it's not enough).
Serving
Small models
As you are after large models I won't delve into options for smaller ones, just a brief mention.
If you want to serve multiple users over the network you need some way to scale your architecture (usually cloud like GCP or AWS). You could do that using Kubernetes and it's PODs or pre-allocate some servers to handle requests, but that approach would be inefficient (small number of users and running servers would generate pointless costs, while large numbers of users may halt the infrastructure and take too long to process resuests).
Other way is to use autoscaling based on serverless approach. Resources will be provided based on each request so it has large scaling abilities + you don't pay when the traffic is low. You can see Azure Functions as they are on the path to improve it for ML/DL tasks, or torchlambda for PyTorch (disclaimer, I'm the author) for smaller models.
Large models
As mentioned previously, you could use Kubernetes with your custom code or ready to use tools.
In the first case, you can spread the model just the same as for training, but only do forward pass. In this way even giant models can be put up on the network (once again, GPT-3 with 175B parameters), but requires a lot of work.
In the second, #Daniel provided two possibilities. Others worth mentioning could be (read respective docs as those have a lot of functionalities):
KubeFlow - multiple frameworks, based on Kubernetes (so auto-scaling, multi-node), training, serving and what not, connects with other things like MLFlow below
AWS SageMaker - training and serving with Python API, supported by Amazon
MLFlow - multiple frameworks, for experiment handling and serving
BentoML - multiple frameworks, training and serving
For PyTorch, you could read more here, while tensorflow has a lot of serving functionality out of the box via Tensorflow EXtended (TFX).
Questions from OP's comment
Are there any forms of parallelism that are better within a machine vs
across machines
The best for of parallelism would probably be within one giant computer as to minimize transfer between devices.
Additionally, there are different backends (at least in PyTorch) one can choose from (mpi, gloo, nccl) and not all of them support direct sending, receiving, reducing etc. data between devices (some may support CPU to CPU, others GPU to GPU). If there is no direct link between devices, those have to be first copied to another device and copied again to target device (e.g. GPU on other machine -> CPU on host -> GPU on host). See pytorch info.
The more data and the bigger network, the more profitable it should be to parallelize computations. If whole dataset can be fit on a single device there is no need for parallelization. Additionally, one should take into account things like internet transfer speed, network reliability etc. Those costs may outweigh benefits.
In general, go for data parallelization if you have lots of of data (say ImageNet with 1.000.000 images) or big samples (say images 2000x2000). If possible, within a single machine as to minimize between-machines transfer. Distribute model only if there is no way around it (e.g. it doesn't fit on GPU). Don't otherwise (there is little to no point to parallelize when training MNIST as the whole dataset will easily fit in RAM and the read will be fastest from it).
why bother build custom ML-specific hardware such as TPUs?
CPUs are not the best suited for highly parallel computations (e.g. matrices multiplication) + CPU may be occupied with many other tasks (like data loading), hence it makes sense to use GPU.
As GPU was created with graphics in mind (so algebraic transformation), it can take some of CPU duties and can be specialized (many more cores when compared to CPU but simpler ones, see V100 for example).
Now, TPUs are tailored specificially for tensor computations (so deep learning mainly) and originated in Google, still WIP when compared to GPUs. Those are suited for certain types of models (mainly convolutional neural networks) and can bring speedups in this case. Additionally one should use the largest batches with this device (see here), best to be divisible by 128. You can compare that to NVidia's Tensor Cores technology (GPU) where you are fine with batches (or layer sizes) divisible by 16 or 8 (float16 precision and int8 respectively) for good utilization (although the more the better and depends on number of cores, exact graphic card and many other stuff, see some guidelines here).
On the other hand, TPUs support still isn't the best, although two major frameworks support it (tensorflow officially, while PyTorch with torch_xla package).
In general, GPU is a good default choice in deep learning right now, TPUs for convolution heavy architectures, though might give some headache tbh. Also (once again thanks #Daniel), TPUs are more power effective, hence should be cheaper when comparing single floating point operation cost.
I am running TensorFlow on a machine which has two GPUs, each with 3 GB memory. My batch size is only 2GB, and so can fit on one GPU. Is there any point in training with both GPUs (using CUDA_VISIBLE_DEVICES)? If I did, how would TensorFlow distribute the training?
With regards to memory: I assume that you mean that one data batch is 2GB. However, Tensorflow also requires memory to store variables as well as hidden layer results etc. (to compute gradients). For this reason it also depends on your specific model whether or not the memory will be enough. Your best bet would be to just try with one GPU and see if the program crashes due to memory errors.
With regards to distribution: Tensorflow doesn't do this automatically at all. Each op is placed on some device. By default, if you have any number of GPUs available, all GPU-compatible ops will be placed on the first GPU and the rest on the CPU. This is despite Tensorflow reserving all memory on all GPUs by default.
You should have a look at the GPU guide on the Tensorflow website. The most important thing is that you can use the with tf.device context manager to place ops on other GPUs. Using this, the idea would be to split your batch into X chunks (X = number of GPUs) and define your model on each device, each time taking the respective chunk as input and making sure to reuse variables.
If you are using tf.Estimator, there is some information in this question. It is very easy to do distributed execution here using just two simple wrappers, but I personally haven't been able to use it successfully (pretty slow and crashes randomly with a segfault).
I am trying to implement a neural network architecture (Self Organizing Maps) for execution on GPUs. I am exploring TensorFlow for this task.
In TensorFlow, I noticed that you just have to specify gpu as the device to execute something on the gpu like in this post. It seems that the way the operations are parallelized is decided by TF and the user does not have options to take optimization decisions. The "Optimizing for GPU" section on TensorFlow Performance Guide also does not talk about explicit control over parallelizing operations.
My question is, can I do CUDA-like optimization in TensorFlow? More elaborately, is it possible to define which operation will be parallelized (like defining CUDA kernels for parallel operations)?
Yes, but you probably don't want to.
At the most extreme you can define your own op (as described here: https://www.tensorflow.org/extend/adding_an_op).
You can implement it as a GPU Kernel and write whatever you want.
You probably don't want to. The default operations are likely well optimized. I doubt you would be able to squeeze anything out significant out of them.
You can decide the device placement for each individual operation (by using tf.device), but you will incur data transfer overhead every time you switch. This should cover the cases where there's some operation that it slow to execute on the GPU.
If you want to process part of the data on CPU and part on the GPU you can slice your data and do 2 operations (one on CPU and one on GPU).
By default, in TF, in graph mode (not in eager mode), everything, all the TF ops run in parallel. There is a thread pool for that, and its size is controlled via inter_op_parallelism_threads. (See also.)
That does not necessarily mean that e.g. multiple matmul will really run in parallel, if they are internally synchronized. That is the case for most CUDA ops, as there is only a single CUDA stream. See here.
I have a small web server that gets input in terms of sentences and needs to return a model prediction using Tensorflow Serving. It's working all fine and well using our single GPU, but now I'd like to enable batching such that Tensorflow Serving waits a bit to group incoming sentences before processing them together in one batch on the GPU.
I'm using the predesigned server framework with the predesigned batching framework using the initial release of Tensorflow Serving. I'm enabling batching using the --batching flag and have set batch_timeout_micros = 10000 and max_batch_size = 1000. The logging does confirm that batching is enabled and that the GPU is being used.
However, when sending requests to the serving server the batching has minimal effect. Sending 50 requests at the same time almost linearly scales in terms of time usage with sending 5 requests. Interestingly, the predict() function of the server is run once for each request (see here), which suggests to me that the batching is not being handled properly.
Am I missing something? How do I check what's wrong with the batching?
Note that this is different from How to do batching in Tensorflow Serving? as that question only examines how to send multiple requests from a single client, but not how to enable Tensorflow Serving's behind-the-scenes batching for multiple separate requests.
(I am not familiar with the server framework, but I'm quite familiar with HPC and with cuBLAS and cuDNN, the libraries TF uses to do its dot products and convolutions on GPU)
There are several issues that could cause disappointing performance scaling with the batch size.
I/O overhead, by which I mean network transfers, disk access (for large data), serialization, deserialization and similar cruft. These things tend to be linear in the size of the data.
To look into this overhead, I suggest you deploy 2 models: one that you actually need, and one that's trivial, but uses the same I/O, then subtract the time needed by one from another.
This time difference should be similar to the time running the complex model takes, when you use it directly, without the I/O overhead.
If the bottleneck is in the I/O, speeding up the GPU work is inconsequential.
Note that even if increasing the batch size makes the GPU faster, it might make the whole thing slower, because the GPU now has to wait for the I/O of the whole batch to finish to even start working.
cuDNN scaling: Things like matmul need large batch sizes to achieve their optimal throughput, but convolutions using cuDNN might not (At least it hasn't been my experience, but this might depend on the version and the GPU arch)
RAM, GPU RAM, or PCIe bandwidth-limited models: If your model's bottleneck is in any of these, it probably won't benefit from bigger batch sizes.
The way to check this is to run your model directly (perhaps with mock input), compare the timing to the aforementioned time difference and plot it as a function of the batch size.
By the way, as per the performance guide, one thing you could try is using the NCHW layout, if you are not already. There are other tips there.