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.
Suppose we have the following architecture:
Multiple CNN layers
RNN layer
(Time-distributed) Dense classification layer
We want to train this architecture now. Our fancy GPU is very fast at solving the CNN layers. Although using a lower clockrate, it can perform many convolutions in parallel, thus the speed. Our fancy CPU however is faster for the (very long) resulting time series, because the time steps cannot be parallelized, and the processing profits from the higher CPU clockrate. So the (supposedly) smart idea for execution would look like this:
Multiple CNN layers (run on GPU)
RNN layer (run on CPU)
(Time-distributed) Dense classification layer (run on GPU/CPU)
This lead me to two important questions:
Is it possible, with any of the frameworks mentioned in the title, to distribute certain layers to certain hardware, and how?
If it is possible, would the overhead for the additional memory operations, e.g. tranferring between GPU-/CPU-RAM, render the whole idea useless?
Basically, in Pytorch you can control the device on which variables/parameters reside. AFAIK, it is your responsibility to make sure that for each operation all the arguments reside on the same device: i.e., you cannot conv(x, y) where x is on GPU and y is on CPU.
This is done via pytorch's .to() method that moves a module/variable .to('cpu') or .to('cuda:0')
As Shai mentioned you can control this yourself in pytorch so in theory you can have parts of your model on different devices. You have then to move data between devices in your forward pass.
I think the overhead as you mentioned would make the performance worst. The cuda RNN implementation benefits greatly from running on a gpu anyways :)
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.
When I was learning tensorflow, one basic concept of tensorflow was computational graphs, and the graphs was said to be static.
And I found in Pytorch, the graphs was said to be dynamic.
What's the difference of static Computational Graphs in tensorflow and dynamic Computational Graphs in Pytorch?
Both frameworks operate on tensors and view any model as a directed acyclic graph (DAG), but they differ drastically on how you can define them.
TensorFlow follows ‘data as code and code is data’ idiom. In TensorFlow you define graph statically before a model can run. All communication with outer world is performed via tf.Session object and tf.Placeholder which are tensors that will be substituted by external data at runtime.
In PyTorch things are way more imperative and dynamic: you can define, change and execute nodes as you go, no special session interfaces or placeholders. Overall, the framework is more tightly integrated with Python language and feels more native most of the times. When you write in TensorFlow sometimes you feel that your model is behind a brick wall with several tiny holes to communicate over. Anyways, this still sounds like a matter of taste more or less.
However, those approaches differ not only in a software engineering perspective: there are several dynamic neural network architectures that can benefit from the dynamic approach. Recall RNNs: with static graphs, the input sequence length will stay constant. This means that if you develop a sentiment analysis model for English sentences you must fix the sentence length to some maximum value and pad all smaller sequences with zeros. Not too convenient, huh. And you will get more problems in the domain of recursive RNNs and tree-RNNs. Currently Tensorflow has limited support for dynamic inputs via Tensorflow Fold. PyTorch has it by-default.
Reference:
https://medium.com/towards-data-science/pytorch-vs-tensorflow-spotting-the-difference-25c75777377b
https://www.reddit.com/r/MachineLearning/comments/5w3q74/d_so_pytorch_vs_tensorflow_whats_the_verdict_on/
Both TensorFlow and PyTorch allow specifying new computations at any point in time. However, TensorFlow has a "compilation" steps which incurs performance penalty every time you modify the graph. So TensorFlow optimal performance is achieved when you specify the computation once, and then flow new data through the same sequence of computations.
It's similar to interpreters vs. compilers -- the compilation step makes things faster, but also discourages people from modifying the program too often.
To make things concrete, when you modify the graph in TensorFlow (by appending new computations using regular API, or removing some computation using tf.contrib.graph_editor), this line is triggered in session.py. It will serialize the graph, and then the underlying runtime will rerun some optimizations which can take extra time, perhaps 200usec. In contrast, running an op in previously defined graph, or in numpy/PyTorch can be as low as 1 usec.
In tensorflow you first have to define the graph, then you execute it.
Once defined you graph is immutable: you can't add/remove nodes at runtime.
In pytorch, instead, you can change the structure of the graph at runtime: you can thus add/remove nodes at runtime, dynamically changing its structure.
My GPU which is Titan X should have be faster than the CPU which is Intel(R) Xeon(R) CPU E5-2643 v3 # 3.40GHz. But two of my models runs a little slower on the GPU. One model runs much faster on the GPU. Among those two models, one is implemented with tensorflow, the other is implemented with theano. The common character of the two models is they all belong to hierarchical Bi-LSTM model which means the last outputs of the bottom Bi-LSTM are fed into the other as inputs. So neither of the models are too simple.
So I would like to inquire what are the possible reasons that they run slower on GPU thant on CPU?
I could provide some info for the theano side:
Theano has been having multiple issues with scan, which is its workhorse for RNN loops.
Here's some of them:
Since theano does not know shape information at compile time, the resulting compiled routine can be suboptimal (like using gemv for vector-vector dot).
(as of Nov 2016) The current version of scan is implemented in cython, which have some overhead over a pure C++ version. If the RNN don't have much computation density on a single step, this can be significant.
It does not pipeline well. Using a scan to implement a map operation can often slower than using the underlying operation directly. Apparently the optimizer is premature and still can't recognize this kind of problem.
Solution:
Try upgrading to dev version. They have been making various improvements overtime.
Try unrolling the RNN (using a plain loop to build graph instead of scan), if you can afford the compilation time.
I made a PR to address gemv issue, only for old GPU backend. Give it a try (if not merged yet). Now it's part of dev master branch.