...
The existing parameter server approach to distributed MXNet faces limitations in performance and feature completeness (tensor fusion, gradient compression, single-bit gradient compression and ability to use MPI and NCCL.).
Horovod is an open-source distributed training framework , that has shown 2x speedup compared to distributed TensorFlow using innovative techniques [1, 2].
We propose to add Horovod support to MXNet. This will help our users achieve goal of linear scalability up to 256 GPUs and beyond. Naturally, we will support multi-machine CPU training too.
Value Proposition
This project is seeking to provide an alternative distributed training solution for MXNet customers. It offers customers the following value proposition:
Usability - Users do not have to experiment with number of workers and number of servers to get best performance out-of-the-box.
Performance - Horovod + Tensorflow has shown 2x performance of Distributed Tensorflow [1], so we expect it to compare well to parameter servershow similar gains.
Cost savings - Parameter servers are not needed when they use Horovod.
Simplified architecture - Leverage battle-tested libraries such as MPI and NCCL, as well as network optimizations such as RDMA.
Profiler - Horovod has an excellent profiler for finding bottlenecks.
Online learning - Due to its MPI paradigm, Horovod can save checkpoints which enables online learning and fine-tuning of your model. With parameter server, it takes some additional work to save Optimizer state located on servers, but with Horovod , this feature comes for free. Note: this feature is not currently supported.
Community - Horovod is a way for MXNet to leverage the Deep Learning community for advancements in distributed training, and for increasing MXNet's visibility.
Proposed Approach
User Interface
...
$ mpirun -np 8 --hostfile ~/hosts --bind-to none --map-by slot -x NCCL_DEBUG=INFO -x NCCL_MIN_NRINGS=4 -x LD_LIBRARY_PATH -x PATH -x MXNET_USE_OPERATOR_TUNING=0 -mca pml ob1 -mca btl ^openib python3 /home/ubuntu/master/3rdparty/horovod/examples/mxnet_imagenet_resnet50.py --benchmark 1 --batch-size=256 --network resnet-v1 --num-layers=50 --num-epochs 1 --kv-store horovod None --dtype float16 --gpus 0
Note: This call is valid when using OpenMPI. To improve user experience, this training script may be wrapped into a Bash script in the future.
...
hvd.allreduce
hvd.broadcast_parameters
hvd.local_rank
hvd.rank
hvd.local_size
hvd.size
hvd.DistributedOptimizerhvd.BroadcastVariablesHook
The following are the two key calls from the Horovod API warrant additional explanation:
...
(a) hvd.allreduce (b) hvd.broadcast_parameters
Figure 1. How two key Horovod operations are implemented using Horovod API
Going through how Allreduce works, the DistributedOptimizer is used to wrap an MXNet Optimizer class:
- In every iteration the DistributedOptimizer wrapper will insert an Allreduce of the gradients before the weight update is done.
- This is done by calling hvd.allreduce.
- This calls down into the C API horovod_mxnet_allreduce_async
- This calls MXNet's PushAsync, which creates a callback for Horovod to call upon completion of the Allreduce.
- After the Allreduce is complete, Optimizer's weight update is done.
Horovod Interaction with MXNet
...
This behaviour is the same as when doing pip install horovod for TensorFlow and PyTorch support. When those libraries are not present, Horovod installation will fail.
Figure 2. How Horovod interacts with MXNet engine
...
Code Changes to MXNet Repository
We need to:
Introduce the new C API method Modify Module.fit to take a list of hvd.BroadcastVariableHooksMXWaitforHorovodAllreduce and MXWaitforHorovodBroadcast
Make a folder called "include" that will be located in the same folder as the MXNet pip package. This will allow Horovod to call into MXNet for both the NDArray operations and the MXWaitForHorovod call that tells the MXNet engine which NDArrays must be locked for the duration of the Allreduce and Broadcast.
...
Proof of concept that MXNet can be made to work with Horovod
Code reuse (~90% of the code from the prototype will go into the final design)
Fast prototyping possible due to sidestepping challenges such as:
Compilation of Horovod pip package separate from MXNet, which requires some arcane mechanisms such as CFFI
Building DistributedOptimizer class that wraps Optimizer class in MXNetAdding hook to Module.fit to support calling hvd.broadcast at time of kv.init in MXNet
Our prototype uses the KVStore API to call Horovod backend. We expose a new KVStore class that can be selected by the user.
...
Easier integration with other MXNet bindings, because those bindings already support KVStore
User does not have to install another dependency in order to do distributed training, because MXNet build tool includes Horovod source code as a 3rd party dependency.
However, there is a trade-off, because then the Horovod source code would need to be maintained to ensure there are no regressions
Language bindings for languages other than Python are available without additional work
Performance Benchmarks
Final API Benchmarks
Model: resnet-v1, 50 layers
Dataset: synthetic
Dtype: float32
Instance types: Horovod+X (16 32 p3.16xlarge), parameter server (16 32 p3.16xlarge, 32 r4.16xlarge).
...
Figure 4. Preliminary benchmark on synthetic data comparing parameter server co-located
...
(servers
...
on same
...
node as workers
...
) and Horovod+MXNet
Prototype Benchmarks
Model: resnet-v1, 50 layers
Dataset: synthetic
Dtype: float32
Instance types: Horovod+X (16 p3.16xlarge), parameter server (16 p3.16xlarge, 32 r4.16xlarge).
...
Figure 45. Preliminary benchmark on synthetic data comparing parameter server co-located (servers on same node as workers), parameter server 2 servers:1 worker) , Intel MPI+MXNet, Horovod+Tensorflow, and Horovod+MXNet.
Addition of New APIs
We are introducing a new MXWaitForHorovodAllreduce and MXWaitForHorovodBroadcast function to the MXNet C API. This function will takes the form of:
- void MXWaitforHorovodAllreduce( NDArray* input, NDArray* output, bool average, char* name, void (*func)(NDArray*, NDArray*, bool, char*, void (*cb)()))
- void MXWaitforHorovodBroadcast( NDArray* input, NDArray* output, bool average, char* name, void (*func)(NDArray*, NDArray*, bool, char*, void (*cb)()))
The parameters are:
...
CPU support and GPU fp16 support
Since CPU support and GPU fp16 support are listed as experimental at the moment, we do not have performance numbers for them.
Test Plan
Functionality Tests
We will be introducing unit tests for most public Horovod functions at the Python level:
hvd.allreduce
hvd.broadcast_parameters
hvd.local_rank
hvd.rank
hvd.local_size
hvd.size
These will be contained under the path "horovod/test/test_mxnet.py", next to "test_tensorflow.py" and "test_torch.py". To run the test:
$ mpirun -np 8 --hostfile ~/hosts --bind-to none --map-by slot -x NCCL_DEBUG=INFO -x NCCL_MIN_NRINGS=4 -x LD_LIBRARY_PATH -x PATH -x MXNET_USE_OPERATOR_TUNING=0 -mca pml ob1 -mca btl ^openib test_mxnet.py
Performance Tests
...
Automated performance tests will be outside the scope of this project. For example, the Horovod repository itself does not have any performance tests. They provide a pointer to the repo from https://github.com/tensorflow/benchmarks, and say that the user will be able to replicate the performance numbers shown in the paper [1].
Technical Challenges
MXNet not intended to be used in 1 process/1 GPU mode
...
Since we are linking with the MXNet shared library, we need to include the correct headers in the PyPi package. In order to avoid ABI compatibility issues, we may need to add additional APIs (e.g. mx.config.get_compile_flags or mx.config.get_link_flags) that return the compilation and linker flags respectively. Then, the Horovod installation can proceed using the exact same flags.
Gluon support
Gluon support can be added by:
- making sure to pass the DistributedOptimizer object into Trainer instead of into Module
- using hvd.broadcast_parameters on the exposed initialized parameters
Milestones
Aug. 10, 2018: Prototype API available for testing: https://github.com/ctcyang/horovod/tree/mxnet_fp16_divide_before_sum/examples/mxnet
- GPU fp32 support has been tested
- GPU fp16 support is still experimental
- CPU support is experimental
Oct. 5, 2018: Beta release of Final API
...