full-duplex congestion symptoms

[DECNetPhaseV]

this is a real newbie question, but I can't seem to find an answer anywhere.

if I have three ports on e.g. a Gb switch, with an aggregate fabric rate of say 6 Gb, and two systems send packets to a third at a rate of 700 Mb each, ovbiously the receiver can't be sent all the data. (Assuming steady-state, full duplex for all systems.)

Specifically for a CISCO switch, what do the client and server see? Does the switch just silently throw away packets, or does it haev some way of reporting this problem?

(In the old days we would have seen packet retransmission failures, but that doesn't seem to apply in a switched environment.)

thanks

[tbaranski]

The switch will (silently) drop packets, and retransmission must be handled at a higher layer (e.g., TCP).

[Kevin Dorrell]

The switch will drop packets, as your other correspondent said.

It will count those drops, so you can see the counts from the CLI of the switch. For CatOS, look at "show counters". For IOS, look at "show int". If you want more details, I can post them next time I have access to my switches. Are you IOS or CatoS?

Do you have a network management system of any sort? If so, you could use it to monitor the discards and raise an alarm when they go above a certain level. Again, it depends on what NMS you are using.

There could also be messages on the syslog, but the detailed behaviour depends on which swicth, which IOS/CatOS, and the version.

Kevin Dorrell

Luxembourg

[pflunkert]

You must wrote: "The switch should count drops" The catalyst 2950, 3550,3560 and 3750 don't count any dropped packets. This is software independent!! I made several tests: With a smartbit a sent traffic through the switch. More traffic the switch can sent. He should count the dropped packets but he don't do it. I opened a case and Cisco creates bugs (CSCee70104 for example). Unfortunally you can't bank on the counters. On the catalyst 3750 you have the possibility to use another command to the the dropped packets: show platform port-asic stats drop asic 1

The complete story is very long, so i post only the short form ;-))

Regards

Peter

[Kevin Dorrell]

Thanks for that useful information Peter. I had a look at our 2950s in the light of what you said. On the "show int", I see a counter "Total output drops", but I don't see any value other than 0 on any of the switches. So, yes. I presume the drop would be on the output queue, that is, it wouldn't block/drop at the input queue.

On the input side, I am looking at a "show int" that says "Input queue: 0/75/1095/0 (size/max/drops/fluches)", so it looks as though I have seen some losses on input. Strange thing is that my Spectrum NMS looking at the same interface does not report any discards. I don't know which OID the Spectrum is looking at.

Unfortunately I dont have any 35s to try your command on.

I had a look at the "show counters" on CatOS 4000, but I didn't see anything that looked like drops or discards. The NMS reports discards, but I couldn't find any CatOS port that had ever discarded anything.

Kevin Dorrell

Luxembourg

[pflunkert]

Hi Kevin,

Nice that the information helps you. When you need some more infos, write a posting and i will sent you a mail.

Regards

Peter

[DECNetPhaseV]

Hi Kevin, et al:

thanks--I don't actually own the border switches, they belong to the central network folks, hence no access to the counters. Which means I have to start analyzing the TCP re-transmission rates on the host side. (Thankfully there isn't a lot of UDP used.)

Jeff