network loss on directly connected links

schmurfy

schmurfy

Hello,

My setup is as follow:
- soekris NET4501 box
- FreeBSD 8.0
- sis network driver
- netgraph compiled in kernel but disabled
- pf compiled in kernel but disabled

My problem is rather strange, I connected the box with another computer on two physical interfaces (also two different interfaces on the other side) and when doing a "ping -f xxx" once everything works fine but if I do it on both interfaces at the same time (from the box to the other computer) ping -f start printing dots (request sent but no reply) and when I stop both ping I get 5% to 35% loss, the worst thing is that when I do a tcpdump -plni on one of the exiting interfaces I do see the icmp answers :(

Anyone experienced this ?
I tried same thing on linux with the same hardware and it works as expected (0% loss)
(don't tell me to use linux I just use it to test if it is an hardware problem, I prefer FreeBSD on all points)


This time there is no netgraph, no firewall, I do not see what could impact this. Could it be a bug with the sis driver ?
 
One thing I have noticed with the sis driver is that it can't seem to handle full duplex traffic well at all. Using a Soekris here with the quad port sis card installed.
 
Can I know how you did notice that? Because I also noticed some strange behavior aside of the one I describe but I cannot pinpoint them as specifically as this one. In my current state I think I will just try another freebsd release and see if it works, maybe FreeBSD 6 will works...
 
I was benchmarking another device on my network with iperf and noticed performance fall through the floor when I did full duplex tests. I narrowed it down to the sis based devices in my soekris.

I was running 8.0 at the time too. Curious to know if it works better with 6...
 
After some tests here are the results (all with a minimal kernel config):

- changing HZ with freebsd FreeBSD 8 (150, 400, 1000): same problem
- testing with freebsd FreeBSD 6: same problem
- using a tweaked version of monowall to allow shell access (http://www.xs4all.nl/~fredmol/m0n0/#downloads1_22, embedded image): same problem

So I have no configuration working currently, given the number of different versions I tried and even monowall I doubt it is something specific to me... Is there such a thing like icmp output limit?

If I ping flood the box from the other computer (after a "sysctl net.inet.icmp.icmplim=0" on the box) on the two interfaces there is no loss reported.

But if the ping flood comes from the box itself the two ping -f processes report something between 5% and 35% loss but still stay at 0% if there is only one ping flood running.

I don't know if this is a real problem of if only icmp packets are affected somehow but if it also happens with other network packets it could explains some problems we encounter.
 
You must log in or register to reply here.
Back
Top