Distributed learning of CNNs on heterogeneous CPU/GPU architectures

TL;DR

This paper introduces a novel distributed training method for CNNs that efficiently leverages heterogeneous CPU/GPU architectures, significantly reducing training time without sacrificing accuracy.

Contribution

It presents a new model parallelism approach focusing on distributing only convolutional layers across heterogeneous devices, optimizing training speed.

Findings

Achieves up to 3.28x speedup with four CPUs.

Achieves up to 2.45x speedup with three GPUs.

Effective for CNNs on datasets like CIFAR-10.

Abstract

Convolutional Neural Networks (CNNs) have shown to be powerful classification tools in tasks that range from check reading to medical diagnosis, reaching close to human perception, and in some cases surpassing it. However, the problems to solve are becoming larger and more complex, which translates to larger CNNs, leading to longer training times that not even the adoption of Graphics Processing Units (GPUs) could keep up to. This problem is partially solved by using more processing units and distributed training methods that are offered by several frameworks dedicated to neural network training. However, these techniques do not take full advantage of the possible parallelization offered by CNNs and the cooperative use of heterogeneous devices with different processing capabilities, clock speeds, memory size, among others. This paper presents a new method for the parallel training of CNNs that can be considered as a particular instantiation of model parallelism, where only the convolutional layer is distributed. In fact, the convolutions processed during training (forward and backward propagation included) represent from $60$-$90$\% of global processing time. The paper analyzes the influence of network size, bandwidth, batch size, number of devices, including their processing capabilities, and other parameters. Results show that this technique is capable of diminishing the training time without affecting the classification performance for both CPUs and GPUs. For the CIFAR-10 dataset, using a CNN with two convolutional layers, and $500$ and $1500$ kernels, respectively, best speedups achieve $3.28\times$ using four CPUs and $2.45\times$ with three GPUs. Modern imaging datasets, larger and more complex than CIFAR-10 will certainly require more than $60$-$90$\% of processing time calculating convolutions, and speedups will tend to increase accordingly.