Parallelization strategies for deep learning

Question

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:

1. 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.
2. 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?

Nide

I think this will get a better response on StackOverflow. – Shannashannah

Thanks @TimMak Happy to move the question to StackOverflow. Any advice on how to do that? – Nide

I am not an expert on this exact topic but I have used Horovod with AWS SageMaker in the past which is the recommended approach when doing distributed computing with TS/Keras. Here is the link horovod.readthedocs.io/en/stable/summary_include.html – Coulter

I don't know much on this topic either, but some pointers on asynchronous training. #27435784 realpython.com/async-io-python – Shannashannah

Cryptomining. Look at cryptoming software, such as xmr-stak github.com/fireice-uk/xmr-stak and you will find perfect examples of parallelization both at the CPU as well as the GPU level. When I was into this, I did some work on a 72-core Xeon Phi 7295 and that's what helped me. – Morphia

@TimMak Python's asyncio has nothing to do with asynchronous training in the context of machine learning. – Hakodate

Answer 1

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.

Bibliofilm

Thanks Szymon! What about within-GPU (or within-TPU) computation? What makes them a good fit for what type of parallelism? Are there any forms of parallelism that are better witthin a machine (e.g. GPU/ TPU) vs across machines? (nodes in a network). Or is everything you said here ("sync vs async", "data vs model parallelism", etc) as good of a fit in both cases? (if so, why bother build custom ML-specific hardware such as TPUs?) – Nide

@Nide is it what you are after more or less? I'm not into hardware too much, so you may read up on CPU vs GPU vs TPU on your own, here are some basics and how it looks more or less at the moment. – Bibliofilm

@Nide The techniques apply both to parallelism within and across machines. However, as Szymon mentioned that it's better to do vertical scaling first (within machine) than horizontal scaling (across machines) because across-machine communication is costly. For example, it's better to have one instance with 4 GPUs than two instances with 2 GPUs each. As for why building TPUs, they are better than GPUs in terms of performance/wattage and perf/price. – Hakodate

Thanks @Hakodate - QQ on gradient accumulation (GA). How exactly is it different from mini batch gradient descent? "[GA] 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)." – Nide

@Nide minibatch gradient descent isn't a different concept. You can do minibatch gradient descent with gradient accumulation. Traditionally (without gradient accumulation), what you do is: run forward pass on minibatch, calculate loss, backpropagate to get gradients, apply updates via optimizer (e.g. plain gradient descent). With gradient accumulation: run forward pass on minibatch, calculate loss, backprop (repeat those three steps N times), average gradients from multiple backpropagations (divide by N) and apply optimization afterwards. Is it clear now? Is TPU/GPU clear after answer update? – Bibliofilm

Awesome. Thanks @SzymonMaszke! and the use of a parameter server helps mostly with bandwidth / data transfers when doing all-reduce right? Also, is it true that none of these frameworks currently store / persist in a DB the parameters (they are perhaps logged) as the algorithm runs? I left this comment to try to relate parameter servers to what you wrote here on all-reduce, but I may have it wrong, so I'd be curious whether or not you agree with those calculations, and how you think it relates to what you wrote here. – Nide

@Nide parameter server - not only that. As the comment you linked to shows, each worker has to send it's packet to server (say 1GB), server merges the update somehow and updates all the workers. So only 2 1GB transfer per worker in this case. In all-reduce without parameter server it would take 8 transfers per worker (each send data to the others, assuming 9 workers in total) and each should receive data from other workers (8 again). This would be 16 1GB transfers per worker, hence way more wasteful. Yes, it helps with bandwith if each worker needs the same data locally. – Bibliofilm

@Nide DB - I don't see a point in storing parameters in database (assuming neural network), but someone correct me if I'm wrong. Those are created as highly optimized directed graphs, often irregular and don't need much description. pytorch uses pickle and loads the model to RAM, tensorflow uses protobuf (which can be used for fast transfers over the network), keras uses HDF5 which again is a file format. – Bibliofilm

Answer 2

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.

Hakodate