Deep Gradient Compression [arXiv]

@inproceedings{lin2018dgc,
  title={{Deep Gradient Compression: Reducing the communication bandwidth for distributed training}},
  author={Lin, Yujun and Han, Song and Mao, Huizi and Wang, Yu and Dally, William J},
  booktitle={The International Conference on Learning Representations},
  year={2018}
}

Overview

We release the PyTorch code of the Deep Gradient Compression.

Figure 1. Deep Gradient Compression (DGC) can reduce the communication bandwidth (transmit less gradients by pruning away small gradients), improve the scalability, and speed up distributed training.


Figure 2. : DGC maintains accuracy: Learning curves of ResNet (the gradient sparsity is 99.9%).

Figure 3. DGC improves the scalability: speedup measured on NVIDIA TITAN RTX 2080Ti GPU cluster with 25 Gbps Ethernet.

Content

• Prerequisites
• Code
• Training
• Known Issues and TODOs

Prerequisites

The code is built with following libraries (see requirements.txt):

• Python >= 3.7
• PyTorch >= 1.5
• Horovod >= 0.19.4
• numpy
• tensorboardX >= 1.2
• tqdm
• openmpi >= 4.0

Code

The core code to implement DGC is in dgc/compression.py and dgc/memory.py.

• Gradient Accumulation and Momentum Correction

    mmt = self.momentums[name]
    vec = self.velocities[name]
    if self.nesterov:
        mmt.add_(grad).mul_(self.momentum)
        vec.add_(mmt).add_(grad)
    else:
        mmt.mul_(self.momentum).add_(grad)
        vec.add_(mmt)
    return vec

• Sparsification

    importance = tensor.abs()
    # sampling
    sample_start = random.randint(0, sample_stride - 1)
    samples = importance[sample_start::sample_stride]
    # thresholding
    threshold = torch.min(torch.topk(samples, top_k_samples, 0, largest=True, sorted=False)[0])
    mask = torch.ge(importance, threshold)
    indices = mask.nonzero().view(-1)

Training

We use Horovod to run distributed training:

• run on a machine with N GPUs,

horovodrun -np N python train.py --configs [config files]

e.g., resnet-20 on cifar-10 dataset with 8 GPUs:

# fp16 values, int32 indices
# warmup coeff: [0.25, 0.063, 0.015, 0.004, 0.001] -> 0.001
horovodrun -np 8 python train.py --configs configs/cifar/resnet20.py \
    configs/dgc/wm5.py configs/dgc/fp16.py configs/dgc/int32.py

• run on K machines with N GPUs each,

mpirun -np [K*N] -H server0:N,server1:N,...,serverK:N \
    -bind-to none -map-by slot -x NCCL_DEBUG=INFO \
    -x LD_LIBRARY_PATH -x PATH -mca pml ob1 \
    -mca btl ^openib -mca btl_tcp_if_exclude docker0,lo \
    python train.py --configs [config files]

e.g., resnet-50 on ImageNet dataset with 4 machines with 8 GPUs each,

# fp32 values, int64 indices, no warmup
mpirun -np 32 -H server0:8,server1:8,server2:8,server3:8 \
    -bind-to none -map-by slot -x NCCL_DEBUG=INFO \
    -x LD_LIBRARY_PATH -x PATH -mca pml ob1 \
    -mca btl ^openib -mca btl_tcp_if_exclude docker0,lo \
    python train.py --configs configs/imagenet/resnet50.py \
    configs/dgc/wm0.py

For more information on horovodrun, please read horovod documentations.

You can modify/add new config files under configs to change training settings. You can also modify some trivial configs in the command:

python train.py --configs [config files] --[config name] [config value] --suffix [suffix of experiment directory]

e.g.,

horovodrun -np 8 python train.py --configs configs/cifar/resnet20.py \
    configs/dgc/wm5.py --configs.train.num_epochs 500 --suffix .e500

Here are some reproduce results using 0.1% compression ratio (i.e., configs.train.compression.compress_ratio = 0.001):

#GPUs	Batch Size	#Sparsified Nodes	ResNet-50	VGG-16-BN	LR Scheduler
-	-	-	76.2	73.4	-
8	256	8	76.6	74.1	MultiStep
16	512	16	76.5	73.8	MultiStep
32	1024	32	76.3	73.3	MultiStep
32	1024	32	76.7	74.4	Cosine
64	2048	64	76.8	74.2	Cosine
64	2048	8	76.6	73.8	Cosine
128	4096	16	76.4	73.1	Cosine
256	8192	32	75.9	71.7	Cosine

Known Issues and TODOs

• Backend: We currently only support OpenMPI backend. We encountered some errors when calling allgather using NCCL2 backend: allgathered data are random data once in a while; if we set CUDA_LAUNCH_BLOCKING=1 for debugging, everything works well.
• #Sparsified Nodes: We currently treat each GPU as an independent node. However, communication is rarely a bottleneck within one machine. A better strategy should be performing allreduce dense gradients intra-machine and allgather sparse gradients inter-machines.
  • For accuracy/convergence verification, we can simulate this by setting configs.train.num_batches_per_step to desired #GPUs per machine (see accuracy table for batch size = 4096/8192).
• Sparsification Granularity: We naively perform fine-grained (i.e., element-wise) top-k to select gradients, and thus the communication will suffer from increased allgather data volume as #nodes increases.
  • Sun et.al. modified the process with coarse-grained sparsification: gradients are partioned into chunks, allreduce the gradient chunks selected based on allreduced L1-norm of each chunk, which gets rid of the allgather and solves the problem.
• Data Encoding: We did not perform any data quantization/encoding before transmission. Data encoding can further reduce data volume.
• Overhead: Performing sparsification (esp. adapting thresholding) in C/C++ may further reduce the DGC overhead.

License

This repository is released under the Apache license. See LICENSE for additional details.

Acknowledgement

• Our implementation is modified from grace which is an unified framework for all sorts of compressed distributed training algorithms.