apeforest commented on a change in pull request #17235: [DOC] Add a few tips
for running horovod
URL: https://github.com/apache/incubator-mxnet/pull/17235#discussion_r363619993
##########
File path: example/distributed_training-horovod/README.md
##########
@@ -199,3 +199,11 @@ $ mpirun -np 8 \
-mca pml ob1 -mca btl ^openib \
python train.py
```
+
+## Tuning Horovod Performance
+
+1. To analyse horovod performance, [horovod
timeline](https://github.com/horovod/horovod/blob/master/docs/timeline.rst) is
a handy tool to trace and visualize the time spent on horovod operations.
+
+2. A few tuning knobs affect horovod runtime performance (explained
[here](https://github.com/horovod/horovod/blob/master/docs/tensor-fusion.rst)).
Apart from `HOROVOD_FUSION_THRESHOLD`, sometimes we find increasing
`HOROVOD_CYCLE_TIME` (up to 100 ms), changing
[`NCCL_ALGO`](https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/env.html#nccl-algo),
and `NCCL_MIN_NCHANNELS` improves performance.
Review comment:
Can you add a reference to NCCL_MIN_NCHANNELS
https://docs.nvidia.com/deeplearning/sdk/nccl-developer-guide/docs/env.html#nccl-min-nchannels
here too?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
