Open qnikst opened 9 years ago
On the other hand taking NC away from local communication removes a number of possible errors with NC lock or bottoms in values that are sent.
I don't follow that argument. Maybe an example is in order.
Mine? If NC is not used in local communication then unsafe send of message that will evaluate to error will not break NC.
@qnikst That's not a good argument I think - more behaviour special cases is a bad thing, not a good thing.
Agree.
Yes, I agree too. The purpose of skipping the NC was to reduce latency, and avoid bottlenecks in the NC thread. I think we can probably leave this as an open question for now, as nobody is yet screaming at us to speed up intra-node comms.
[Imported from JIRA, posted by @hyperthunk Tim Watson at 26/Apr/13 7:16 PM as DP-76]
Some initial investigation suggests that this doesn't have much effect on message passing efficiency though. I'm not totally surprised, and I suspect any micro-benchmarks we come up with will vary considerably between applications. Hmph.