[Click] some feedback on Click stability with e1000

[Eddie Kohler]

Hey,

Anyone else see problems like the below, from Yong Liao at UMass?

Eddie




I think I did not describe our click problem very clearly yesterday. You 
told to me that click crashes at certain time could because of pushing a 
pkt after freeing it. I hesitated to mention to you that the crashing 
problem happens when we use the original click code. In our virtual 
network project, we use the kernel click forwarding speed as a base line 
in evaluating the virtual router forwarding speed. So we downloaded the 
click code from ucla website and followed the instruction to patch and 
compile the  kernel, compile click, and test kernel click with a simple 
configuration that moves packets between two interfaces. The udpgen 
configuration in the click package is used to generate traffic.

We found that when the packet generation rate is high, say 500k~1000k 
pps for 64-byte pkts, the kernel click moving packets between two 
interfaces sometimes crashes by itself, no matter whether polling is 
used or not. Besides, the crashing is more likely to happen in the 
afternoon and on smp machines (we tried a dual-core intel cpu machine 
and a single-core but hyper-threading intel machine).

My guess is that maybe the reason is twofold. First is that the e1000 
driver in click may not work very well with relatively new intel e1000 
cards. Second is that kernel click may have some smp related issues, 
such as the kernel must disable or enable some configuration features 
but the click document does not explicitly point them out.