MXNet supports training with multiple CPUs and GPUs, which may be located on different physical machines.
In default, MXNet uses data parallelism to partition the workload over multiple devices. Assume there are n devices, then each one will get the complete model and train it on 1/n of the data. The results such as the gradient and updated model are communicated cross these devices.
Model parallelism is also supported. In this parallelism, each device maintains a part of the model. It is useful when the model is too large to fit into a single device. There is a tutorial showing how to do model parallelism for a multi-layer LSTM model. This tutorial will focus on data parallelism.
In default, MXNet will partition a data batch evenly into each GPU. Assume batch size b and k GPUs, then in one iteration each GPU will perform forward and backward on b/k examples. The gradients are then summed over all GPUs before updating the model.
To use GPUs, we need to compiled MXNet with GPU support. For example, set
USE_CUDA=1inconfig.mkbeforemake. (see MXNet installation guide for more options).
If a machine has one or more than one GPU cards installed, then each card is labeled by a number starting from 0. To use a particular GPU, one can often either specify the context ctx in codes or pass --gpus in the command line. For example, to use GPU 0 and 2 in python one can often create a model with
import mxnet as mx model = mx.model.FeedForward(ctx=[mx.gpu(0), mx.gpu(2)], ...)
while if the program accepts a --gpus flag such as example/image-classification, then we can try
python train_mnist.py --gpus 0,2 ...
If the GPUs have different computation power, we can partition the workload according to their powers. For example, if GPU 0 is 3 times faster than GPU 2, then we provide an additional workload option work_load_list=[3, 1], see model.fit for more details.
Training with multiple GPUs should have the same results as a single GPU if all other hyper-parameters are the same. But in practice, the results vary mainly due to the randomness of I/O (random order or other augmentations), weight initialization with different seeds, and CUDNN.
We can control where the gradient is aggregated and model updating if performed by creating different KVStore, which is the module for data communication. One can either use mx.kvstore.create(type) to get an instance or use the program flag --kv-store type.
There are two commonly used types,
local: all gradients are copied to CPU memory and weights are updated there.device: both gradients' aggregation and weight updating are run on GPUs. It also attempts to use GPU peer-to-peer communication, which potentially accelerates the communication. But this option may result in higher GPU memory usage.When there is a large number of GPUs, e.g. >=4, we suggest using device for better performance.
We can simply change the KVStore type to run with multiple machines.
dist_sync behaviors similarly to local. But one major difference is that batch-size now means the batch size used on each machine. So if there are n machines and we use batch size b, then dist_sync behaviors equally to local with batch size n*b.dist_device_sync is identical to dist_sync with the difference similar to device vs local.dist_async performs asynchronous updating. The weight is updated once received gradient from any machine. The update is atomic, namely, no two updates happen on the same weight at the same time. However, the order is not guaranteed.To use distributed training, we need to compile with
USE_DIST_KVSTORE=1(see MXNet installation guide for more options).
Launching a distributed job is a bit different from running on a single machine. MXNet provides tools/launch.py to start a job by using ssh, mpi, sge, or yarn.
Assume we are at the directory mxnet/example/image-classification. and want to train mnist with lenet by using train_mnist.py. On a single machine, we can run by
python train_mnist.py --network lenet
Now if there are two ssh-able machines, and we want to train it on these two machines. First, we save the IPs (or hostname) of these two machines in file hosts, e.g.
$ cat hosts 172.30.0.172 172.30.0.171
Next, if the mxnet folder is accessible by both machines, e.g. on a network filesystem, then we can run by
../../tools/launch.py -n 2 --launcher ssh -H hosts python train_mnist.py --network lenet --kv-store dist_sync
Note that, besides the single machine arguments, here we
launch.py to submit the jobssh if all machines are ssh-able, mpi if mpirun is available, sge for Sun Grid Engine, and yarn for Apache Yarn.-n number of worker nodes to run-H the host file which is required by ssh and mpi--kv-store use either dist_sync or dist_asyncNow consider if the mxnet folder is not accessible. We can first copy the MXNet library to this folder by
cp -r ../../python/mxnet . cp -r ../../lib/libmxnet.so mxnet
then ask launch.py to synchronize the current directory to all machines' /tmp/mxnet directory with --sync-dst-dir
../../tools/launch.py -n 2 -H hosts --sync-dst-dir /tmp/mxnet \ python train_mnist.py --network lenet --kv-store dist_sync
MXNet often chooses the first available network interface. But for machines have multiple interfaces, we can specify which network interface to use for data communication by the environment variable DMLC_INTERFACE. For example, to use the interface eth0, we can
export DMLC_INTERFACE=eth0; ../../tools/launch.py ...
SetPS_VERBOSE=1 to see the debug logging, e.g
export PS_VERBOSE=1; ../../tools/launch.py ...
../../tools/launch.py -h