3x100Mbps bonding: works!

Laurent Itti itti at cco.caltech.edu
Wed Oct 4 22:59:07 PDT 2000

Hi all:

ok, what was wrong is that the MAC addresses did not truly get all set to
the same value for the three branches in the bonded channel.
A peculiarity of the Realtek driver I guess.

however, there is a very cool program "rtl8139-diag" written by Donald
Becker (available on his driver page) that allows for low-level access to
the realtek eeprom; it seems that the program was all setup to change the
HWADDR except that there was no implementation for doing the actual change
(i.e., the option existed, the argument would get parsed, and then nothing
would be done with it!) That was just a matter of adding a few lines of
code, and it works flawlessly.


> the overhead for sharing an irq is just that some of the time,
> the wrong instance of a driver will get called.  it checks its 
> hardware, decides the irq is not for it, and returns.  the code
> overhead is trivial, but you could make the case for worrying
> a little about the time overhead, since PCI transactions are 
> many, many CPU cycles.  ie, O(1us).  still, even with ~5K int/second,
> that's not a major problem.

yes, that's what I intuitively thought. But is there any way to instruct
the bios to better distribute the hardware over the many free IRQs?
I think that for CPI there is nothing we can do after the bios has done
its job, right?

right now in the dual-CPU console, I have 4 ether and an IDE RAID and it
just pisses me off to see nothing on irqs 3,5,7 but 2*IDE+1*ether on 11!

anybody has experience maybe with playing with the 4 PCI IRQ line settings
in a bios?

[[[machine gets sluggish when doing 3 simultaneous transfers]]]
> you could trivially answer that question by looking at /proc/interrupts
> (or even "vmstat 1").  I suspect the problem is actually memory,
> that you need to extend the backlog queues, perhaps also configure
> the kernel to keep more memory free.

good point!  now the machine is not sluggish anymore...
maybe it was because I was using 2.4.0 before and now downgraded to
2.2.17 (2.4.0 was too unstable with my stripped RAID disk).

I'll post a summary on our website once we are all done. So far this
Beowulf looks fabulous!

  -- laurent

More information about the Beowulf mailing list