Horovod on GPU¶
To use Horovod on GPU, read the options below and see which one applies to you best.
In most situations, using NCCL 2 will significantly improve performance over the CPU version. NCCL 2 provides the allreduce operation optimized for NVIDIA GPUs and a variety of networking devices, such as RoCE or InfiniBand.
If you have installed NCCL 2 using the
nccl-<version>.txzpackage, you should add the library path to
LD_LIBRARY_PATHenvironment variable or register it in
$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/nccl-<version>/lib
2. (Optional) If you’re using an NVIDIA Tesla GPU and NIC with GPUDirect RDMA support, you can further speed up NCCL 2 by installing an nv_peer_memory driver.
GPUDirect allows GPUs to transfer memory among each other without CPU involvement, which significantly reduces latency and load on CPU. NCCL 2 is able to use GPUDirect automatically for allreduce operation if it detects it.
Note: Open MPI 3.1.3 has an issue that may cause hangs. The recommended fix is to downgrade to Open MPI 3.1.2 or upgrade to Open MPI 4.0.0.
If you installed TensorFlow from PyPI, make sure that
g++-5or above is installed.
If you installed PyTorch from PyPI, make sure that
g++-5or above is installed.
If you installed either package from Conda, make sure that the
gxx_linux-64Conda package is installed.
If you installed NCCL 2 using the
nccl-<version>.txzpackage, you should specify the path to NCCL 2 using the
$ HOROVOD_NCCL_HOME=/usr/local/nccl-<version> HOROVOD_GPU_OPERATIONS=NCCL pip install --no-cache-dir horovod
If you installed NCCL 2 using the Ubuntu package, you can run:
$ HOROVOD_GPU_OPERATIONS=NCCL pip install --no-cache-dir horovod
If you installed NCCL 2 using the CentOS / RHEL package, you can run:
$ HOROVOD_NCCL_INCLUDE=/usr/include HOROVOD_NCCL_LIB=/usr/lib64 HOROVOD_GPU_OPERATIONS=NCCL pip install --no-cache-dir horovod
Note: Some models with a high computation to communication ratio benefit from doing allreduce on CPU, even if a
GPU version is available. To force allreduce to happen on CPU, pass
opt = hvd.DistributedOptimizer(opt, device_dense='/cpu:0')
Advanced: Have a proprietary MPI implementation with GPU support optimized for your network?¶
This section is only relevant if you have a proprietary MPI implementation with GPU support, i.e. not Open MPI or MPICH. Most users should follow one of the sections above.
If your MPI vendor’s implementation of allreduce operation on GPU is faster than NCCL 2, you can configure Horovod to use it instead:
$ HOROVOD_GPU_ALLREDUCE=MPI pip install --no-cache-dir horovod
Additionally, if your MPI vendor’s implementation supports allgather, broadcast, and reducescatter operations on GPU, you can configure Horovod to use them as well:
$ HOROVOD_GPU_OPERATIONS=MPI pip install --no-cache-dir horovod
Note: Allgather allocates an output tensor which is proportionate to the number of processes participating in the
training. If you find yourself running out of GPU memory, you can force allgather to happen on CPU by passing
opt = hvd.DistributedOptimizer(opt, device_sparse='/cpu:0')