Mikael,
I see. In my setup the T60 is able to upload data to PE860 with much
higher bandwidth than to download data from PE860 because it's the
PE860 NIC which is in half-duplex mode and makes the retransmits. If I
put the switch port to 100/half and PE860 NIC to 100/full, I get
opposite results in terms of speed: download in T60 is much higher
than upload from T60 to PE860.

BTW, it looks like duplex mismatch tends to sometimes break Iperf.
Probably due to duplex mismatch TCP sessions drop which affect the
Iperf.


Gert,
what do you mean by "telco gear"? And I understand that late collision
is a collision which happens after the first 512b of frame are on the
wire, but why shouldn't late collisions occur in case the network is
fine? Is it because in time it takes Tx in one end to send 512b of
data to a wire, the Rx of the other end should have already received
the first bits of the frame and doesn't try to transmit anything until
the frame is completely received?


regards,
martin


2011/7/24 Gert Doering <[email protected]>:
> Hi,
>
> On Sun, Jul 24, 2011 at 04:52:33AM +0300, Martin T wrote:
>> WS-C2950C-24 port Fa0/1 is configured to "speed 100" and "duplex full":
>>
>> Fa0/1     -> PE860           connected    1            full    100 
>> 10/100BaseTX
>>
>> ..and as PE860 eth0 interface is configured to autonegotiate it's
>> speed, there is a duplex mismatch between the switch and PE860
>> machine.
>
> Of course.  So just don't do that.  Never configure anything for
> "autonegotiation off" unless you know for sure that one of the devices
> will not do autonegotiation properly (which is nowadays basically
> "the single-port fastethernet ports on cisco 7200s", "telco gear", and
> nothing else).
>
>> I configured IP address 10.10.10.1 to eth0 of T60 and 10.10.10.2 to
>> eth0 of PE860. There is no packet loss between the T60 and PE860 if
>> there is no traffic other than ICMP "echo requests" and "echo replies"
>> between the hosts:
>
> This is not surprising.  Packet loss will hit if both sides want to send
> traffic at the same time - the "full duplex" side will just send it, while
> the "half duplex" side will detect it as "late collision" and drop both
> packets.
>
> Ping is strictly request/response, with only one packet ever in-flight
> at the same time, so no problems here.
>
> Try two parallel ping streams with "ping -i 0.01 -s 1000" and you'll see
> loss.
>
> [..]
>> So I guess duplex mismatch does affect UDP traffic as well? >
>
> Sure.  It breaks *Ethernet* traffic, no matter what's in there.
>
> gert
> --
> USENET is *not* the non-clickable part of WWW!
>                                                           //www.muc.de/~gert/
> Gert Doering - Munich, Germany                             [email protected]
> fax: +49-89-35655025                        [email protected]
>

_______________________________________________
cisco-nsp mailing list  [email protected]
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/

Reply via email to