eric-haibin-lin commented on a change in pull request #17555: [MXNET-#16795]
Byteps-KVStore: Intergrate Byteps into mxnet as new type of kvstore backend
URL: https://github.com/apache/incubator-mxnet/pull/17555#discussion_r376827971
##########
File path: tools/launch.py
##########
@@ -63,14 +63,27 @@ def main():
in default it is equal to NUM_WORKERS')
parser.add_argument('-H', '--hostfile', type=str,
help = 'the hostfile of slave machines which will run \
- the job. Required for ssh and mpi launcher')
+ the job. Required for ssh and mpi launcher.\
+ When -SH is set, the file provided by -H will \
+ be used to recognize worker machines only. Otherwise, \
+ -H is used for both server and worker machines.')
+ parser.add_argument('-SH', '--server-hostfile', type=str,
+ help = 'the hostfile of server machines which will run
\
+ the job. Required for byteps multi-machine launching.')
parser.add_argument('--sync-dst-dir', type=str,
help = 'if specificed, it will sync the current \
directory into slave machines\'s SYNC_DST_DIR if ssh \
launcher is used')
parser.add_argument('--launcher', type=str, default='ssh',
choices = ['local', 'ssh', 'mpi', 'sge', 'yarn'],
help = 'the launcher to use')
+ bps_group = parser.add_argument_group('byteps-backend')
+ bps_group.add_argument('--byteps', action='store_true',
+ help = 'Whether use byteps launcher to launch')
+ bps_group.add_argument('--scheduler-ip', type=str,
Review comment:
If you add flags for scheduler ip/port, they shall support other launcher
type too (e.g. native distributed kvstore). Do you plan to support that?
Alternatively it's also ok to remove them and use localhost and a random ip for
the scheduler.
----------------------------------------------------------------
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
