Strange things on GBit / 1000->100 / net.inet.tcp.inflight.*

[Andre Oppermann]

Raphael H. Becker wrote:
> On Fri, Sep 17, 2004 at 11:26:20PM +0200, Andre Oppermann wrote:
> 
>>Raphael H. Becker wrote:
>>
>>>># sysctl net.inet.tcp
>>>># sysctl net.inet.tcp.hostcache.list
>>>># netstat -s -p tcp
>>>># netstat -s -p ip
>>>
>>>http://rhb.uugrn.org/FreeBSD/bugs/5.x/1000-to-100/check_before.txt
>>>http://rhb.uugrn.org/FreeBSD/bugs/5.x/1000-to-100/check_after.txt
>>
>>Could you please provide the same information (except the second sysctl one)
>>from the target machine as well.  You don't have to show the difference
> 
> http://rhb.uugrn.org/FreeBSD/bugs/5.x/1000-to-100/check_target.txt

Ok, this give some very stange information:

  On the taget machine (4.10) we see a huge amount of OO packets arriving:

		92920 out-of-order packets (134546620 bytes)

  but we don't see nearly that many on retransmits on the source machine (5.3):

		281 data packets (405792 bytes) retransmitted

However the data set on the target machine seem to be skewed.  There is a
lot of other TCP traffic on there as well.  So this probably doesn't really
relate directly to the test you did.

>>between before and after but try to make sure that not much other traffic
>>was going than the test.
> 
> Just my ssh-Session. But I can't guarantee for each packet passing the
> router (= the target-machine). There's no "production"-traffic on that
> subnet, just my testboxes.

Hmmm... Without getting the TCP statictics back to zero it's hard to correlate
any data.

Would you mind trying a 5.3 to 5.3 transfer but the target 5.3 box forced to
only 100Mbit full-duplex?  Same statistics again.

-- 
Andre