Skip to contentSkip to conversation

Longstanding networking / SMP issue? (duplextest)

DISCUSSION jamal 1 participant 2 messages Feb 20, 2003 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

jamal

author
#2 lore raw

I got you. I think its a clever idea. You have better chances if you make your packets larger.
Netdev is the network developers mailing list at [email protected] Linux-net is network users mailing list;

cheers,
jamal

Quoted 56 lines · Simon Kirby
Quoted 3 lines · jamal

Hi, could you please either posting or ccing netdev on network related issues? There are a lot of people who could help you but are not subscribed to lk.

Is netdev separate from linux-net? If so, where is it? :)

Quoted 8 lines

I dont have an answer for you. I think testing a different card like Dave says would be a good start. I am curious about the theory of operation of your program; by duplex mismatch i take it you mean one end has one speed setting but the other has a conflicting one?
If thats so i am not sure i understand why sending a burst of packets and waiting for responses catches this problem. Are you saying you could successfully send but fail to receive? I dont see the connection and i am curious.

Here's the story...

Once upon a time, Donald wrote (or helped write, I don't know) the eepro100 driver. Because Intel had so many variants of the eepro100, it was difficult to implement link autonegotiation because different cards had the link state detection wired to different pins. Even today, the eepro100 driver doesn't seem to autonegotiate with any of the switches I've tried it with. (10/100 autonegotiates fine, but the duplex does not.)

To get around this problem, we had to force our switch ports to 100/full and force the eepro100 driver to 100/full also. This works well, but because we have hundreds of boxes running diferent OSs and we tend to move things aroud fairly often, it's easy to forget to set up the port properly. When this happens, everything appears to work, but depending on timing, some packets are dropped due to one side believing there is a collision while other side sees no problem. (This assumes you understand how Ethernet works -- if not, let me know!)

The timing required to hit the packet loss is actually quite important. Web site visitors can often download at fairly high speeds over the Internet (600 kB/sec, for example) from a machine with a duplex mismatch, but internally (on the same LAN), transfer rates can be as low as 20 kB/sec, although usually about 60 kB/sec, depending on the number of switches being traversed, luck, etc.

The duplextest program attempts to set up a timing situation that exploits the half-collision scenario by sending multiple echo packets at once, so that the first will be on the way back while the latter are still being sent. "ping -f" doesn't work for this because it always uses a timer and only sends one packet.

I hope this explains things...

Simon-

[        Simon Kirby        ][        Network Operations        ]
[     [email protected]     ][     NetNation Communications     ]

[ Opinions expressed are not necessarily those of my employer. ]

author
#3 lore raw

Pls ignore this. Wasnt meant to go to netdev - this is what happens when i get too lazy (i invoked it from pine so i could cunpaste it and then forgot to delete it ;->).

cheers,
jamal

Quoted 70 lines · jamal

I got you. I think its a clever idea. You have better chances if you make your packets larger.
Netdev is the network developers mailing list at [email protected] Linux-net is network users mailing list;

cheers,
jamal

Quoted 56 lines · Simon Kirby
Quoted 3 lines · jamal

Hi, could you please either posting or ccing netdev on network related issues? There are a lot of people who could help you but are not subscribed to lk.

Is netdev separate from linux-net? If so, where is it? :)

Quoted 8 lines

I dont have an answer for you. I think testing a different card like Dave says would be a good start. I am curious about the theory of operation of your program; by duplex mismatch i take it you mean one end has one speed setting but the other has a conflicting one?
If thats so i am not sure i understand why sending a burst of packets and waiting for responses catches this problem. Are you saying you could successfully send but fail to receive? I dont see the connection and i am curious.

Here's the story...

Once upon a time, Donald wrote (or helped write, I don't know) the eepro100 driver. Because Intel had so many variants of the eepro100, it was difficult to implement link autonegotiation because different cards had the link state detection wired to different pins. Even today, the eepro100 driver doesn't seem to autonegotiate with any of the switches I've tried it with. (10/100 autonegotiates fine, but the duplex does not.)

To get around this problem, we had to force our switch ports to 100/full and force the eepro100 driver to 100/full also. This works well, but because we have hundreds of boxes running diferent OSs and we tend to move things aroud fairly often, it's easy to forget to set up the port properly. When this happens, everything appears to work, but depending on timing, some packets are dropped due to one side believing there is a collision while other side sees no problem. (This assumes you understand how Ethernet works -- if not, let me know!)

The timing required to hit the packet loss is actually quite important. Web site visitors can often download at fairly high speeds over the Internet (600 kB/sec, for example) from a machine with a duplex mismatch, but internally (on the same LAN), transfer rates can be as low as 20 kB/sec, although usually about 60 kB/sec, depending on the number of switches being traversed, luck, etc.

The duplextest program attempts to set up a timing situation that exploits the half-collision scenario by sending multiple echo packets at once, so that the first will be on the way back while the latter are still being sent. "ping -f" doesn't work for this because it always uses a timer and only sends one packet.

I hope this explains things...

Simon-

[        Simon Kirby        ][        Network Operations        ]
[     [email protected]     ][     NetNation Communications     ]

[ Opinions expressed are not necessarily those of my employer. ]