emil14
commented
1 month ago As we know from #644 in new design we have one queue for all senders. This design isn't perfect because we have a bottle-neck because of 2 reasons
- slow receiver will slow down sending from sender and receiving by fast receivers
- all senders share same queue so slow sending slows down other faster sendings ("sending" here is both actual sending and receiving by all receivers)
Obvious idea to scale this is to add more queues. One queue per fan-in
Each fan-in has N senders to 1 receiver. We could use that N as a buffer size.
We have opportunity to set buffer size for channels (ports) in IR but we don't do that.
About second part - we might try to somehow automatically calculate (e.g. based on amount of senders/receivers) or expose to user. Or even combine!
Related to #663 and #644