Subject: Re: poor network performance on de0's on 21164's.
To: Andrew Gallatin <gallatin@cs.duke.edu>
From: Michael T. Stolarchuk <mts@off.to>
List: port-alpha
Date: 07/09/1999 11:46:25
In message <14214.4453.355809.780092@grasshopper.cs.duke.edu>, Andrew Gallatin writes:
>
>
>Older tulips (we have a batch of no-name 21140A cards in some older
>alphas & PCs here) also work reliably in full-duplex mode.
>
Since we had two old 21140's lying around, a netgear, and a cogent,
I decided it would make sense to try these two cards between
the two test alpha's...
To review: both alpha's have 10basetx-fdx set... both in the SRM
and in /etc/ifconfig.de0. Visual verification of the lights on
the switch says they're in fdx.
alpha1# ifconfig de0
de0: flags=8863<UP,BROADCAST,NOTRAILERS,RUNNING,SIMPLEX,MULTICAST> mtu 1500
media: Ethernet 100baseTX full-duplex
status: active
inet 192.168.129.1 netmask 0xfffff800 broadcast 192.168.135.255
alpha1#
and
alpha2# ifconfig -a
de0: flags=8863<UP,BROADCAST,NOTRAILERS,RUNNING,SIMPLEX,MULTICAST> mtu 1500
address: 00:00:92:96:92:f1
media: Ethernet 100baseTX full-duplex
status: active
inet 192.168.130.1 netmask 0xfffff800 broadcast 192.168.135.255
and when i ftp, the performance is still awful:
ftp> get netbsd.hold
local: netbsd.hold remote: netbsd.hold
227 Entering Passive Mode (192,168,130,1,255,252)
150 Opening BINARY mode data connection for 'netbsd.hold' (3890305 bytes).
100% |*************************************| 3799 KB 71.73 KB/s 00:00 ETA
226 Transfer complete.
3890305 bytes received in 00:52 (71.73 KB/s)
To see if collisions is the issue, I check netstat output:
alpha1# netstat -i
Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll
de0 1500 <Link> 00:40:05:36:22:33 2978 105 2130 0 0
de0 1500 192.168.128/2 alpha1 2978 105 2130 0 0
lo0 33200 <Link> 26 0 26 0 0
alpha2# netstat -i
Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll
de0 1500 <Link> 00:00:92:96:92:f1 2126 0 3098 6 0
de0 1500 192.168.128/2 alpha2 2126 0 3098 6 0
No collisions, but the receiving machine is getting lots of input errors.
I also recall trying fxp's on these two machines, not getting any
collisions, but still getting very poor performance. As i recall
when i tried running openBsd on one of the alpha's it told me
(when i used de0's) that the problem was framing errors.
When I was using tcpdump on both the sender, and the receiver, I could
see the sender was sending packets back to back. When the sender would
ship more than, say, four packets back to back, the receiving machine
didn't receive all the packets.
It is as if the receiving alpha can't get four packets in a row back to back.
I tried modifying the if_de to allow for more interpacket spacing to
get the back-back data received... but that also didn't work.
mts.