|
|
With this pair of programs I did some performance tests of the messaging
system. Guess what -- I found two bugs :-)
See the subsequent patches.
With 1500 msg_source processes I can generate message overload: A
Intel(R) Xeon(R) CPU L5640 @ 2.27GHz
can receive roughly 100k messages per second. When using
messaging_read_send/recv user/system time is roughly even, a bit more
work done in user space. When using messaging_register, due to less
malloc activity, user space chews a lot less.
By the way: 1.500 helper threads in a blocking sendto() against a single
datagram socket reading as fast as it can (with epoll_wait in between)
only drove the loadavg to 12 on a 24-core machine. So I guess unix domain
datagram sockets are pretty well protected against overload. No thundering
herd or so. Interestingly "top" showed msg_sink at less than 90% CPU,
although it was clearly the bottleneck. But that must be a "top" artifact.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
|
