Attention is currently required from: laforge.
1 comment:
Patchset:
just a random comment: As we do a lot of io_uring in libosmocore and programs using it in recent tim […]
Harald, thanks for feedback.
Current epoll + send/recv seems to be enough for BTS + 1-2-3 Ms, but if we are to go with more Ms it might indeed be better to optimize more. In my experience most of the overhead was in python-related wrapping, but once that is removed it indeed the IO syscalls that will become the bottleneck. We do not forward a lot of data, as GSM bursts are small, but we do forward a lot of packets, as each GSM burst is being sent via 1 UDP message and to keep the forarding latency minimal we cannot coalesce those messages even for different slots in the same GSM frame. So using io_uring might indeed make sense at some point.
When/if we are to do that I would suggest to cosider using libosmocore directly instead of going through additional third-party library. It should hopefully easy to do that from Cython.
Kirill
To view, visit change 40055. To unsubscribe, or for help writing mail filters, visit settings.