Horovod with MPI

MPI can be used as an alternative to Gloo for coordinating work between processes in Horovod. When using NCCL, performance will be similar between the two, but if you are doing CPU training, there are noticeable performance benefits to using MPI.

First install Open MPI or another MPI implementation. Learn how to install Open MPI on this page.

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.

mpirun

horovodrun introduces a convenient, Open MPI-based wrapper for running Horovod scripts.

In some cases it is desirable to have fine-grained control over options passed to Open MPI. This page describes running Horovod training directly using Open MPI.

  1. Run on a machine with 4 GPUs:

    horovodrun -np 4 python train.py
    

    Equivalent Open MPI command:

    mpirun -np 4 \
        -bind-to none -map-by slot \
        -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
        -mca pml ob1 -mca btl ^openib \
        python train.py
    
  2. Run on 4 machines with 4 GPUs each:

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

    Equivalent Open MPI command:

    mpirun -np 16 \
        -H server1:4,server2:4,server3:4,server4:4 \
        -bind-to none -map-by slot \
        -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
        -mca pml ob1 -mca btl ^openib \
        python train.py
    

Starting with the Open MPI 3, it’s important to add the -bind-to none and -map-by slot arguments. -bind-to none specifies Open MPI to not bind a training process to a single CPU core (which would hurt performance). -map-by slot allows you to have a mixture of different NUMA configurations because the default behavior is to bind to the socket.

The -mca pml ob1 and -mca btl ^openib flags force the use of TCP for MPI communication. This avoids many multiprocessing issues that Open MPI has with RDMA which typically results in segmentation faults. Using TCP for MPI does not have noticeable performance impact since most of the heavy communication is done by NCCL, which will use RDMA via RoCE or InfiniBand if they’re available (see Horovod on GPU). Notable exceptions from this rule are models that heavily use hvd.broadcast() and hvd.allgather() operations. To make those operations use RDMA, read the Open MPI with RDMA section below.

With the -x option you can specify (-x NCCL_DEBUG=INFO) or copy (-x LD_LIBRARY_PATH) an environment variable to all the workers.

Custom SSH ports

Specify custom SSH ports with -mca plm_rsh_args "-p <port>" as follows:

mpirun -np 16 \
    -H server1:4,server2:4,server3:4,server4:4 \
    -bind-to none -map-by slot \
    -mca plm_rsh_args "-p 12345"
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -mca pml ob1 -mca btl ^openib \
    python train.py

This is frequently useful in the case of running Horovod in Docker environment.

Open MPI with RDMA

As noted above, using TCP for MPI communication does not have any significant effects on performance in the majority of cases. Models that make heavy use of hvd.broadcast() and hvd.allgather() operations are exceptions to that rule.

Default Open MPI openib BTL that provides RDMA functionality does not work well with MPI multithreading. In order to use RDMA with openib, multithreading must be disabled via the -x HOROVOD_MPI_THREADS_DISABLE=1 option. See the example below:

mpirun -np 16 \
    -H server1:4,server2:4,server3:4,server4:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x HOROVOD_MPI_THREADS_DISABLE=1 -x PATH \
    -mca pml ob1 \
    python train.py

Other MPI RDMA implementations may or may not benefit from disabling multithreading, so please consult vendor documentation.

Horovod Parameter Knobs

Many of the configurable parameters available as command line arguments to horovodrun can be used with mpirun through the use of environment variables.

Tensor Fusion:

$ mpirun -x HOROVOD_FUSION_THRESHOLD=33554432 -x HOROVOD_CYCLE_TIME=3.5 ... python train.py

Timeline:

$ mpirun -x HOROVOD_TIMELINE=/path/to/timeline.json -x HOROVOD_TIMELINE_MARK_CYCLES=1 ... python train.py

Autotuning:

$ mpirun -x HOROVOD_AUTOTUNE=1 -x HOROVOD_AUTOTUNE_LOG=/tmp/autotune_log.csv ... python train.py

Note that when using horovodrun, any command line arguments will override values set in the environment.

Hangs due to non-routed network interfaces

Having network interfaces that are not routed can cause Open MPI to hang. An example of such interface is docker0.

If you see non-routed interfaces (like docker0) in the output of ifconfig, you should tell Open MPI and NCCL to not use them via the -mca btl_tcp_if_exclude <interface>[,<interface>] and NCCL_SOCKET_IFNAME=^<interface>[,<interface>] parameters.

ifconfig

Produces output like this:

docker0   Link encap:Ethernet  HWaddr 02:42:2d:17:ea:66
          inet addr:172.17.0.1  Bcast:0.0.0.0  Mask:255.255.0.0
          UP BROADCAST MULTICAST  MTU:1500  Metric:1
          RX packets:0 errors:0 dropped:0 overruns:0 frame:0
          TX packets:0 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:0
          RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)
eth0      Link encap:Ethernet  HWaddr 24:8a:07:b3:7d:8b
          inet addr:10.0.0.1  Bcast:10.0.0.255  Mask:255.255.255.0
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:900002410 errors:0 dropped:405 overruns:0 frame:0
          TX packets:1521598641 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000
          RX bytes:376184431726 (350.3 GiB)  TX bytes:954933846124 (889.3 GiB)
eth1      Link encap:Ethernet  HWaddr 24:8a:07:b3:7d:8a
          inet addr:192.168.0.1  Bcast:192.168.0.255  Mask:255.255.255.0
          UP BROADCAST RUNNING MULTICAST  MTU:1500  Metric:1
          RX packets:2410141 errors:0 dropped:0 overruns:0 frame:0
          TX packets:2312177 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000
          RX bytes:698398061 (666.0 MiB)  TX bytes:458504418 (437.2 MiB)
lo        Link encap:Local Loopback
          inet addr:127.0.0.1  Mask:255.0.0.0
          inet6 addr: ::1/128 Scope:Host
          UP LOOPBACK RUNNING  MTU:65536  Metric:1
          RX packets:497075633 errors:0 dropped:0 overruns:0 frame:0
          TX packets:497075633 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1
          RX bytes:72680421398 (67.6 GiB)  TX bytes:72680421398 (67.6 GiB)

Example mpirun command with lo and docker0 interfaces excluded:

mpirun -np 16 \
    -H server1:4,server2:4,server3:4,server4:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -x NCCL_SOCKET_IFNAME=^lo,docker0 \
    -mca pml ob1 -mca btl ^openib \
    -mca btl_tcp_if_exclude lo,docker0 \
    python train.py