Detection performance?

Bogdan Costescu bogdan.costescu at iwr.uni-heidelberg.de
Wed Aug 20 09:39:16 EDT 2003


On Wed, 20 Aug 2003,  wrote:

> Using Mii-diag, what could be the average delay 
> between the link-status change ( phisically ) 
> and the detection of this event.

Depends on the card capabilities and the driver. Most drivers poll for 
change, some use an interrupt.

> I'm using a 3Com 905 Tornado PC card; is there 
> a different delay for each PC card in changing the 
> status register? 

I don't understand the question...

> I'd like to have a delay minor than 1 ms
> between the time in which i phisically 
> disconnect the cable and the time in which
> I have the detection (in example with a printf on video, ...)

The 3c59x driver polls every 60 seconds for media status when using 
autonegotiation (default). People from the HA and bonding projects have 
modified this to allow very fine polling of the media registers, however 
this has a big disadvantage: the CPU spends a lot of time waiting for 
completion of in/out operations - the finer the poll, the more CPU lost.
The time taken to talk to th MII does not depend on the CPU speed, but on 
th PCI speed, so the faster the CPU, the more instruction cycles are lost 
to I/O.

> In your experience, is it reasonable? 

No, because the network card should transfer data, not be a watchdog.
There is one other solution, but there is no code for it yet. At least 
the Tornado cards allow generating an interrupt whenever the media 
changes. This would alleviate the need to continually poll the media 
registers and would give an indication very soon after the event happened. 
This was on my to-do list for a long time, but it was never done and 
probably won't be done soon.

> Normally do I have to wait  for a greater delay?

If by "normally" you mean "the 359x driver distributed with the kernel" or 
"the 3c59x driver from Scyld", then yes.

> Thanks in advance for your kind answer and observations.

This isn't really beowulf related. Please use vortex at scyld.com for 
discussing the 3c59x driver.

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De

_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list