Add keepalive support to the TCP OOB component #477
Conversation
|
Sent from
|
Refer to this link for build results (access rights to CI server needed): |
orte/mca/oob/tcp/oob_tcp_component.c
Outdated
There was a problem hiding this comment.
why have mca_oob_tcp_component.tcp_proto and this call to getprotobyname() instead of just using the IPPROTO_TCP constant everywhere? Is there some portability issue being dealt with here, or some sort of IPoIB compatibility?
…se the right define
orte/mca/oob/tcp/oob_tcp_component.c
Outdated
There was a problem hiding this comment.
nit: the => looks odd next to the <=, might be better to replace with --> or means
|
Great minds think alike on the proto lookup - I changed it literally as you sent your comment :-) Will fix the other comment |
orte/mca/oob/tcp/oob_tcp_component.c
Outdated
There was a problem hiding this comment.
This is a pretty frequent keepalive, compared to what I think most people would expect. I'd back off to something at least in the 30-60 second range.
|
|
Refer to this link for build results (access rights to CI server needed): |
|
|
Refer to this link for build results (access rights to CI server needed): |
… keepalive options
|
@goodell I've captured all your comments - see what you think |
|
|
Refer to this link for build results (access rights to CI server needed): |
There was a problem hiding this comment.
I think this should be > instead of <.
|
👍 |
Add keepalive support to the TCP OOB component
Refs open-mpi#627. Fix support for multi-threads with CUDA 7.0
3 participants
Add keepalive support for the TCP inter-node connections so we detect when the cluster loses a node. This adds three new MCA params:
oob_tcp_keepalive_time: Idle time in seconds before starting to send keepalives (num <= 0 => disable keepalive)
oob_tcp_keepalive_intvl: Time between keepalives, in seconds
oob_tcp_keepalive_probes: Number of keepalives that can be missed before declaring error
@goodell could you take a gander? I've got a minor compiler warning silence change in there as well. Just didn't feel like separating it out.