full duplex degrading distrbution layer switch performance. Why?

ccuse · ‎04-10-2002

Hi All,

Christopher CUSE writing to you from Antananrivo, MADAGASCAR.

I have a rather classic setup, but have found that using FULL DUPLEX on my distribution layer switch degrades performance practically to stand-still. Replicated problem at 2nd site with different hosts and cabling.

Here's an abbreviated network schema:

HOST1 .. HOSTn

|

ACCESS LAYER HUB

|

DISTRIBUTION LAYER SWITCH (1548M)

|

SERVERS (NT, AIX, LINUX) + ROUTER

CASE1: 1548M set for 100MB FULL DUPLEX on each SERVER port, 100MB HALF DUPLEX for HUB

HOST1 throughput to SERVERS around 20 megabits per second, bursts to 35 mbits/s

SERVER throuput to another SERVER 568 kilobytes per second. WOW.

CASE1 was replicated onsite with another Cisco 2800 switch with same result

CASE1 was replicated offsite with different hosts different cables with both 1548M and 2800

CASE2: 1548M set for 100MB HALF DUPLEX on each SERVER port, 100MB HALF DUPLEX for HUB

HOST1 throughput to SERVERS unchanged from CASE1

SERVER throughput to antother SERVER aroubnd 20 megabits per second, burts to 35 mbits/s

CABLING: CAT5 patch cable (no 9th wire) AMP connectors put together by me :-)

pin color

1 green w/white

2 green

3 orange w/white

4 blue

5 blue w/white

6 orange

7 brown w/white

8 brown

I really do not think that it is a cable quality problem.

PROBLEM STATEMENT:

Full duplex degrades performance between servers but not between hosts and servers. Theoretically, full duplex should improve over all throughput between servers and between hosts and servers.

Any ideas???

Cheers,

Christopher CUSE

CCNA

--I'm truying to conserve bandwidth now that most tuna is dolphin safe--

msjouw · ‎04-11-2002

How about making sure the servers are set for FULL Duplex also DO NOT let them auto negotiate. This sounds like a classic duplex mismatch problem, one side to half and the other to full. This might explain a little bit on that:

When Full duplex is turned on the CSMA/CD principle is turned off on that port for full duplex does not need to Sense the Carrier and does not need to do Collision Detection for there is no Multiple Access.

Half Duplex still checks to see if the link is available to be used to send data and when the other side, at the moment it starts sending, just decides to send also (remember it does not check for collisions) then the half duplex side says it has a collision and backs of.

This way this speed does not surprise me.

Regards, Maarten Sjouw.

ccuse · ‎04-11-2002

Hi Maarten,

Yup ... I went through quite a bit of effort to assure that SERVERS were specifically set to 100mbit FD during the the two cases. I'm going to repeat the experiments again.

Just plain old bizzare!

Is there a minimum cable length for 100mbits? It's about the only open item ...

Cheers & Thanks!

Christopher CUSE

PS __ I know i am on half-duplex now because each SERVER port is reporting collisions, whereas before they were not.

ccuse · ‎04-11-2002

Good Day Everyone,

Sincerest thanks for the help.

Last evening was spent trying to verify whether my adapters are really in full-duplex. I ass-u-me-d that when you tell the OS that you want full-duplex you would get it. Redhat Linux 7.2 on an IBM Netfinity with an Intel 82557 eepro100 card says that is in full-duplex and I tend to believe it ... plug the cable into a 10mbit or 100hd port and no linkbeat. The other Netfinity 5500 running NT Terminal Server with an embedded AMD card uses an IBM intermediate support driver, and I am starting to be convinced that it is not in FD although it says it is.

I think that I shall make some 2 wire patch cables (obligatory half-duplex) and see whether I get a linkbeat and can ping.

Will report on results.

Have nice weekend

Christopher CUSE

rfroom · ‎04-11-2002

If the servers are still set for AUTO SPEED and DUPLEX, you just created a duplex mismatch. Proper configuration requires the setting on both link partners to be indentical.

DELL ACORD · ‎04-11-2002

Chris,

Does the log show any messages concerning Duplex mismatch or errors for the ports that the servers are connected too! Also on the interface is there any counters for runts, giants, CRC, derfereds and input errors? Also, I had an issue after I had hardcoded both speed and duplex to match on the NIC and the switch. I still had problems. I was using 3COM NIC's and there was a know issue with the nIC driver. I updated the NIC driver and the issue was resolved. There is a web page on TAC about NIC cards and known issues. I can't remember the exact link.

Good luck

rfroom · ‎04-11-2002

This document will be useful for this type of troubleshooting:

www.cisco.com/warp/customer/473/46.html