We propose HOlistic MEmory management (HOME), a new framework for performing tensor placements in large DNN training when GPU memory space is not enough. HOME combines tensor swapping with tensor recomputation to reduce GPU memory footprint. Different from existing work that only considers partial DNN model information, HOME takes the holistic DNN model information into account in tensor placement decisions. More specifically, HOME uses a custom-designed particle swarm optimization algorithm to achieve the globally optimized placement for each tensor of the DNN model with a greatly reduced searching space. This holistic awareness of the whole model information enables HOME to obtain high performance under the given GPU memory constraint. We implement HOME in PyTorch and conduct our experiments using six popular DNN models. Experimental results show that HOME can outperform vDNN and Capuchin by up to 5.7× and 1.3× in throughput. Furthermore, HOME can improve the maximum batch size by up to 2.8× than the original PyTorch and up to 1.3× than Capuchin.

Data swapping between CPUs and GPUs is widely used to address the GPU memory shortage issue when training deep neural networks (DNNs) requiring a larger amount of memory than that a GPU may have. Data swapping may become a bottleneck when its latency is longer than the latency of DNN computations. Tensor compression in GPUs can reduce the data swapping time. However, existing works on compressing tensors in the virtual memory of GPUs have three major issues: lack of portability because its implementation requires additional (de)compression units in memory controllers, sub-optimal compression performance for varying tensor compression ratios and sizes, and poor adaptation to dense tensors because they only focus on sparse tensors. We propose a self-tuning tensor compression framework, namedCSwap+, for improving the virtual memory management of GPUs. It uses GPUs for (de)compression directly and thus has high portability and is minimally dependent on GPU architecture features. Furthermore, it only applies compression on tensors that are deemed to be cost-effective considering their compression ratio, size, and the characteristics of compression algorithms at runtime. Finally, to adapt to DNN models with dense tensors, it also supports cost-effective lossy compression for dense tensors with nearly no model training accuracy degradation. We conduct the experiments through six representative memory-intensive DNN models. Compared to vDNN,CSwap+reduces tensor swapping latency by up to 50.9% and 46.1% with NVIDIA V100 GPU, for DNN models with sparse and dense tensors, respectively.

Graphic Processing Units (GPUs) have limited memory capacity. Training popular deep neural networks (DNNs) often requires a larger amount of memory than that a GPU may have. Consequently, training data needs to be swapped between CPUs and GPUs. Data swapping may become a bottleneck when its latency is longer than the latency of DNN computations. Tensor compression in GPUs can reduce the data swapping time. However, existing works on compressing tensors in the virtual memory of GPUs have two major issues: sub-optimal compression performance for varying tensor sparsity and sizes and lack of portability because its implementation requires additional (de)compression units in memory controllers. We propose a self-tuning tensor compression framework, named CSWAP, for improving the virtual memory management of GPUs. It has high portability and is minimally dependent on GPU architecture features. Furthermore, its runtime only applies compression on tensors that are deemed to be cost-effective considering their sparsity and size and the characteristics of compression algorithms. Finally, our framework is fully automated and can customize the compression policy for different neural network architectures and GPU architectures. Our experimental results using six representative memory-intensive DNN models show that CSWAP reduces tensor swapping latency by up to 50.9% and reduces the DNN training time by 20.7% on average with NVIDIA V100 GPUs compared to vDNN.