Run Horovod

Run Horovod

https://horovod.readthedocs.io/en/stable/running_include.html
https://horovod.readthedocs.io/en/stable/running.html

horovodrun训练，通过-np 指定线程数

1. 单机 4卡

$ horovodrun -np 4 -H localhost:4 python train.py

2. 4太机器，每台机器4卡

$ horovodrun -np 16 -H server1:4,server2:4,server3:4,server4:4 python train.py

3. 通过配置文件指定节点，节点的线程数

$ cat myhostfile

aa slots=2
bb slots=2
cc slots=2

• 其中 aa，bb，cc 是主机的名字

• slots 是每个主机上可以执行的线程数

$ horovodrun -np 6 -hostfile myhostfile python train.py

3. To run using Open MPI

https://horovod.readthedocs.io/en/stable/mpirun.html

4. Horovod in Docker

https://horovod.readthedocs.io/en/latest/docker_include.html

5.To run in Kubernetes, see Kubeflow, MPI Operator, Helm Chart, FfDL, and Polyaxon.

https://github.com/kubeflow/examples/tree/master/demos/yelp_demo/ks_app/vendor/kubeflow/mpi-job

6.To run in Spark, see Spark.

7.To run in Singularity, see Singularity.

8.To run in a LSF HPC cluster (e.g. Summit), see LSF.