Add cmake flag -DHPX_WITH_FAULT_TOLERANCE=ON (OFF by default) #3170
Conversation
…re connection reset signals of connections, so HPX does not terminate if a node crashes.
hpx/runtime.hpp
Outdated
| /// component type. This a purely internal function allowing to work | ||
| /// around certain library specific problems related to dynamic | ||
| /// loading of external libraries. | ||
| HPX_EXPORT bool tolerate_node_faults(); |
There was a problem hiding this comment.
This comment does not match the code.
There was a problem hiding this comment.
sorry, copy/paste :-( I'll add a proper comment
There was a problem hiding this comment.
I just updated the branch. Not sure if anything else is required of me here. Never did PR before.
|
What is the rationale of making this functionality optional? Wouldn't it be better to be always able to handle this kind of exception? |
hkaiser
changed the title
|
The functionality is optional, because I'm not sure if ignoring a "connection_reset" signal due to suddenly failing remote endpoints is the right way of handling this -- it is a workaround to avoid HPX raising an exception, and terminating, and allowing to explore recovery (which we are working on). |
In the end this should be the default, it shouldn't even be possible to disable this. I would have no objections to go ahead with merging this PR if that's the eventual goal. |
|
The goal should be to have a fault tolerance HPX in the end. I am all for having it optional, since there's always a trade off between fault tolerance and performance. This is also the first step in that direction so it makes sense to have it defaulted to being turned off for now. |
Ok. Fair enough. |
|
I fixed the comment and pushed the changes to the branch Let me know if you need anything else. |
to ignore connection reset signals of connections, so HPX does not terminate if a node crashes.