Bjarke Istrup Pedersen wrote:
> Indeed, I know it can.
>
> What I'm trying to figure out is how I can get something useful out of
> it when it crashes, so I can track the problem down. (I'm kind of
> hoping it is a software issue)
> Any ideas? (I'm running Linux 2.6.34.1 on it)

Can you remove the "bridging" and make it routable ?  It is possible
(unlikely) that the bridging code can somehow starve the kernel of
resources (be it memory, spare CPU power, out of interrupt handler).


What is your interface tx queue length ?  Maybe turn it down to 25 
packets "/sbin/ifconfig eth0 txqueuelen 25" (or less).  The idea of this 
is to make it start to drop packets BEFORE the system/kernel gets into 
trouble, dropped packets should show up in the interface statistical 
counters (so if it remains zero under high load keep turning it down). 
The connection being bridged will automatically react and recover and 
lower the throughput.


Are you able to knock something up to monitor /proc/interrupts also 
/proc/meminfo maybe even /proc/schedstat (I'm not sure myself how to 
interpret that one) also interface statistical counters.  Sync values to 
disk to disk to provide some kind of graph.


The other aspect is overheating (and other instablity effects), maybe 
just removing the case lid would be enough to provide additional 
cooling.  Doing a memtest (to check for bad bit patterns).  Although 
well engineered it doesn't work forever.


Linux does support traffic shaping via such things are Token Bucket 
Filters (TBF) and although it may sound kind of strange to enable 
something that consumes an additional amount of CPU time when the 
problem maybe CPU starvation.  However the goal here is to maintain a 
stable system by providing rate limiting.    TBF works by deliberately 
delaying the packets as they pass through the device by tiny amounts 
this makes the sender and receiver think the maximum throughput across 
the network is lower than it really is.


Also take a look for some Intel Pro/100 Dual-Port (82559 based PCI-X 
cards).  They are literally being thrown away on ebay as all enterprises 
uses 1000bT now and all consumers devices have then on motherboard, they 
are PCI-X (have 64bit/66MHz extra finger which will fit/overhang net5501 
with just milli-metres to spare) and sometimes come in multipacks boxes 
of 5 for $25 total (yes that is $5 each).  Google the Part Code: 
PILA8472C3PAK5. That is what I have been able to pickup.

They don't have a huge internal buffer but are "Bus-Master" capable.


Darryl
_______________________________________________
Soekris-tech mailing list
[email protected]
http://lists.soekris.com/mailman/listinfo/soekris-tech

Reply via email to