On Wed, Oct 13, 2010 at 03:33:27PM +0600, Anton Maksimenkov wrote: > 2010/10/13 Claudio Jeker <[email protected]>: > > Most probably re(4) was unable to allocate > > clusters and now the RX ring is empty and stuck, in the worst case you hit > > an interrupt storm. > > BTW, in worst case (I used ping -f, just can't find anything useful to > generate many small packets) "re" generate about 5000 interrupts. > While "vr" on same machine can do about 20 000 interrupts and the > machine still stay. >
As Henning mentioned, number of interrupts do not matter at all. My re(4) is happyly pumping 75kpps bidirectional with 6500 interrupts per second. The system is still mostly idle. ping -f is a bad test as it is a ping pong protocol and not mad to max out a line. > > As you can see my i386 had mbuf and mcluster failures because I run the > > kernel out of kvm (this would not have happend if I increased > > kern.maxclusters a bit more carefully). > > Anyway, you need to properly tune your system to handle 1000 and more > > TCP connections. > > As mikeb said it was done and so on. I still have problems to reproduce this. I tried the concurrent HTTP client I got from mikeb and my box has no problem running it with a 1000 concurrent TCP connections. The only problem I see is connect delays because SYN packets are dropped (most probably on the server). -- :wq Claudio
