On 29/04/2025 22:19, Otto Moerbeek wrote:
If you looks more closely you'll see that resuseport and UDP do not work well together for the multithreaded case. Alls packets end up in one thread.
Yes, so in you comment bellow I believe you meant "does NOT handle ... multi-threaded case very well", correct?
According to a few tests I've done with pthread code as well as fork() and SO_REUSEPORT,
all connections end up in the same thread/process.
Fixing this would require something like SO_REUSEPORT_LB, like FreeBSD and DragonFly did, or intervention on the daemon side to do the load balancing?
Alternatively running multiple 1-thread daemons and do some kind of load balancing on the system, say with relayd.
thanks,
G
On 29/04/2025 18:11, Otto Moerbeek wrote:
IIRC OpenBSD does handle sending to a non-blocking socket in the multi-threaded case very well. If multiple threads try to do that simulteneously, some of them will get a EAGAIN. To test this hypothesis, try running with a single thread and see if the errors disappear.
No comments:
Post a Comment