30% Collisions - How to Resolve?

[Peter_17]

Hi all, I have a couple B132L boxes on a switched network which are running on forced 10MB Half-Dupe ports. I am getting anywhere from 25-35% collisions on the interfaces, but no tx/rx errors. I have switched cables & ports, but am still experiencing high collisions. Most other machines on the network have no indication of collisions at all.

Here's the lanadmin data. What I be interpreted from this?

Network Management ID = 4
Description = lan0 Hewlett-Packard LAN Interface Hw Rev 0
Type (value) = ethernet-csmacd(6)
MTU Size = 1500
Speed = 10000000
Station Address = xxxxxxxxxxx
Administration Status (value) = up(1)
Operation Status (value) = up(1)
Last Change = 1916
Inbound Octets = 119228131
Inbound Unicast Packets = 81255
Inbound Non-Unicast Packets = 59326
Inbound Discards = 0
Inbound Errors = 0
Inbound Unknown Protocols = 1865
Outbound Octets = 3191795
Outbound Unicast Packets = 44493
Outbound Non-Unicast Packets = 91
Outbound Discards = 0
Outbound Errors = 0
Outbound Queue Length = 0
Specific = 655367

Ethernet-like Statistics Group

Index = 4
Alignment Errors = 0
FCS Errors = 0
Single Collision Frames = 8236
Multiple Collision Frames = 4792
Deferred Transmissions = 14125
Late Collisions = 0
Excessive Collisions = 0
Internal MAC Transmit Errors = 0
Carrier Sense Errors = 0
Frames Too Long = 0
Internal MAC Receive Errors = 0


...and here's netstat -i

Name Mtu Network Address Ipkts Ierrs Opkts Oerrs Coll
ni0* 0 none none 0 0 0 0 0
ni1* 0 none none 0 0 0 0 0
lo0 4608 loopback localhost 65261 0 65261 0 0
lan0 1500 XX.XX.XXX wcalh0kn 147255 0 45148 0 13028



Any and all help appreciated.

Thanks,

Pete

[paul courry]

Are you running coax or CAT5? If coax then make sure all the terminators are in place and TIGHT.

BTW, what do your network people say about this and have they put a sniffer on this yet?

[Peter_17]

Everything is cat-6

[Sridhar Bhaskarla]

Did you check up the ports on the switch?. Most of the times bad switch ports / director cards cause these problems. You certainly need to check with your Network Operations.

-Sri

[Darrell Allen]

Hi,

Is it possible the switch is set to full duplex?

Darrell

[Stefan Farrelly]

Normally excessive collisions is down to;

1. too many TCP devices broadcasting on the same subnet. Just because other devices arent detecting it doesnt mean its not happening. Normally a reduction in the number of devices by adding an additional hub and distributing half your TCP devices to the new hub fixes the problem.

2. Try another lan cable. This is usually first thing to try as its so easy to do.

3. Your card could possibly be faulty, but in my experience very unlikely. After trying 1. or 2. above dont forget to reboot your HP server to reset all the stats and server/lan card.

[Bernie Vande Griend]

IF, you've checked for sure that your interface card is running at 10 HD and the switch port is also set to 10HD. (Make sure it is not using autonegotiation). and that the port you are using is working OK. Then your interface card could be faulty, although usually you would see other errors as well.
I'm assuming that everything was working well at one point.

[Uday_S_Ankolekar]

Hi,

1.Change the Lan cable,
2. Change the speed to Full duplex,
3. Ask network team to change the port to full duplex.

Good luck

-USA..

[Peter_17]

I'm sure that the B132L is capable only of 10/Half, the B132L+ however can to 100MB.... I'm gonna swap out that box with a known good one to see what happens....

[linuxfan]

Hi Peter,

This is for general information.

The B132L's core NIC is 10Mbit/s half-duplex only.
The B132L+'s core NIC is 10/100 autosensing, autoneg.
If the driver claiming the interface is "lan2" it is 10 Mbit/s, half-duplex only.
If the driver claiming the interface is btlan* it is a 10/100 card.

Secondly,
How long has your machine been up?
netstat -ain reports the cumulative collisions since the machine was last rebooted, did you try "netstat -ain interval" eg: "netstat -ain 5" to see the collisions?

-Regards
Ramesh

[Peter_17]

What I've been doing is running netstat -i 5 and then xfering a big file via ftp into the machines /tmp directory. Here's the result of that... you can see when the xfer starts.... :[myname@b132lbox]% netstat -ian 5
(ni0) -> input output (Total)-> input output
packets errs packets errs colls packets errs packets errs colls
0 0 0 0 0 613672 0 570947 0 14545
0 0 0 0 0 77 0 59 0 0
0 0 0 0 0 1905 0 962 0 360
0 0 0 0 0 3094 0 1574 0 541
0 0 0 0 0 3023 0 1535 0 568


Also, the box which I'm ftp'ing from indicates 0 collisions....

Any thoughts?

[Darrell Allen]

I've had similar issues when there was a mismatch on the duplex settings between the server and the switch: collisions sky-rocket when sending a large amount of data. Otherwise, you don't see the collisions and things seem to be working. Please excuse me if you've already verified this but my net guys like to set auto-neg on their switches or force full-duplex.

Darrell

[rick jones]

Peter -

The stats you posted look perfectly clean. It is quite normal to have a non-trivial collision rate when transfering files between two hosts over a half-duplex link. "Collisions" in and of themselves are _NOT_ errors. The only "collision" that is an error is a "Late Collision" (poor naming choices).

The impact of normal collisions is actually quite small. Now, what _might_ happen in HD mode is something called the "capture effect" where the transmitting system is able to monopolize the wire. The receiver, trying to send ACK's, has his NIC backoff timer grow quite large, the sender runs-out of window (no ACKs) and stops. you wait for the reciever's backoff timer to expire, a flood of ACK's, then you start again. You would be able to see this is a packet trace.

The workarounds are to either shrink the TCP window (something <= 8K I think) or increase the TCP window (>= 56KB perhaps?)

The "fix" is to upgrate your network to full-duplex :)

If a NIC is forced into HD operation, and the other side is doing FD, one would see FCS errors, "late collisions" and the like.

[Printaporn_1]

any error message from netfmt
#netfmt -t 50 -f /var/adm/nettl.LOG00

[Darrell Allen]

Hi Peter,

Would you be so kind as to post your solution in this thread? There's a number of interested people and it will be beneficial for future readers.

Additionally, would you please assign points? Especially to any who may have helped you find the solution but also some "thank you for trying" points would encourage your peers to continue offering assistance, not only to you but to the community in general.

Thanks,
Darrell

[Peter_17]

Firstly, thanks to all who responded so far... Here'a a bit more of my scenerio... I've logged into numerous 10HD machines (mostly B132L's) and performed the same tests with the big xfer while monitoring `netstat -i 2` output. I'm seeing the same statistics (approx 30% collisions) occur on each 10HD box I'm doing this on, regardless if the port it's located on is forced to 10HD or set to augoneg. So with this said, and also with what Rick said, I'm assuming that 30% collisions is a fairly 'normal' and 'predictable' thing when xfering very large files across 10HD capable machines on a predominently 100FD network. Does anyone know where I can find some further information on this topic? Something which explains this with more detail.

As far as points go Darrell, I typically assign them when everything is resolved...

[Wodisch]

Hello Peter,

use ("nettl" and "netfmt") or "ethereal" (from the proting archives) to get an idea what is going on in your network segment. Especially the "top talkers" and "top listeners"!
Then look for broadcasts (ethernet and IP), and the MTU (some stations might have it set too big).
Then you should get a closer idea, who is to blame.

Just my $0.02,
Wodisch

[Darrell Allen]

Hi Peter,

No offense was meant about points. I struggle about when to assign them myself. 8-10 points too early discourages other comments. Non assigned seems to do likewise.

Anyway, I just don't buy the high number of collisions for the number of packets you report:
lan0 1500 XX.XX.XXX wcalh0kn 147255 0 45148 0 13028

I've had 10HD boxes on primarily 100FD lans and they did not experience that high a percent of collisions. This was with large file transfers and network backups both.

Something just smells wrong with your network. All I can think of is to verify the network itself (cable distances, cable specs, terminators, etc).

Wish I could be of more help.

Darrell

[rick jones]

what the rest of the network happens to be doesn't really matter, what matters is what the link into your machine happens to be. since it is half-duplex, it is running CSMA/CD - Carrier Sense (deferred transmission stat) Multiple Access (hubs versus pt-to-pt)/ Collision Detection - when two nodes on the network happen to try to talk at the same time. CSMA/CD is how the nodes of an Ethernet network control access to the network.

CSMA/CD is not used on a full-duplex link because the nodes at either end of the cable are not trying to share access to specific wires. there are wires dedicated for machine A transmitting to B, and for B transmitting to A.

in the past, the presence of CSMA/CD was used as a "litmus" test of something being ethernet or not, regardless of the header format. one could argue then that full-duplex is not really "Ethernet" ...