[Click] Problem with click fake-router.click and delay elements

[Eddie Kohler]

Sahan,

I wonder whether the bug you saw here may have been due to the bug fixed here:

http://www.read.cs.ucla.edu/gitweb?p=click;a=commit;h=01c8f4e084036338e83a6bff7a8e74dc49caa014

Does the bug still occur with current sources?

E


Sahan Gamage wrote:
> Hi Cliff,
> 
> thanks for the reply.
> Yes I am running the latest sources.
> I tried following 2 configurations :
> InfiniteSource(LENGTH 1000) -> EtherEncap(0x0800, 00:11:43:58:19:D1,
> 00:11:43:58:19:D1) -> Queue(2000) -> ToDevice(eth1);
> and
> InfiniteSource(LENGTH 1000) -> EtherEncap(0x0800, 00:11:43:58:19:D1,
> 00:11:43:58:19:D1) -> DelayUnqueue(0.005s) -> Queue(2000) ->
> ToDevice(eth1);
> 
> Other than click trying to occupy 100% of the CPU nothing special
> happened. Also I didn't see the "BUG! Tx Ring full when queue awake!"
> thing showing up.
> 
> I am still trying to figure out what causes the actual problem.
> 
> Thanks
> -sahan
> 
> On Thu, Jan 21, 2010 at 6:42 PM, Cliff Frey <cliff at meraki.com> wrote:
>> Without knowing anything for sure, I suspect that your ethernet driver has a
>> bug in the transmit side when transmitting traffic faster than the wire
>> rate.
>> You could try a configuration like
>> InfiniteSource(LENGTH 1000) -> EtherEncap(...) -> Queue -> ToDevice;
>> and I imagine that you would see the same behavior.
>> You could also try plugging eth0 into a slower ethernet connection (10 or
>> 100Mbps) and you might see the same behavior.
>> Are you running the most recent click sources (from git)?  There might have
>> been some ToDevice changes since then...
>> In any case, just my random ideas... I doubt that LinkUnqueue or Queue are
>> actually misbehaving.
>> Cliff
>>
>> On Thu, Jan 21, 2010 at 2:07 PM, Sahan Gamage <sahans at gmail.com> wrote:
>>> Hi All,
>>>
>>> I was working with click router to implement a delay inducing network.
>>> What I wanted was to configure a click router with standard routing
>>> components and once the routing is done, add some extra delay
>>> (specified). So I modified the fake-router.click to include real
>>> network devices (which is much similar to the router discussed in SOSP
>>> 99 paper). I encountered two problems.
>>> 1. When I ran the router without delay elements and use an iperf
>>> client and server on the two sides of the router, the maximum
>>> throughput I got was 330Mbps. The network drivers were tg3 drivers. My
>>> feeling is that the throughput is somewhat lower than that is
>>> expected.
>>>
>>> 2. Then I changed the configuration a little bit to introduce delays.
>>> What I did was change the output component from:
>>> out0 :: Queue(2000) -> ToDevice(eth0)
>>> to
>>> out0 :: Queue(2000) -> LinkUnqueue(0.005s, 1000Mbps) -> Queue(2000) ->
>>> ToDevice(eth0).
>>>
>>> I did this for both eth0 and eth1.
>>> But when I ran this configuration, the machine always locks up.
>>> Sometimes I can see there's a log entry in kernel message log saying
>>> "tg3: eth1: BUG! Tx Ring full when queue awake!".  Few  options I
>>> tried :
>>>  - I tried different delay elements such as LinkUnqueue and DelayUnqueue.
>>>  - I tried different Queue sizes (from 200 to 5000).
>>>  - I changed the Tx ring size of the devices to 511 (maximum allowed).
>>>
>>> But none of them resolved the problem.
>>>
>>> My feeling is the crash (or the lockup) is primarily due to the delay
>>> introducing elements (LinkUnqueue or DelayUnqueue). Does anybody know
>>> the reasons for reduced performance numbers and the lockups ?
>>>
>>> (I have attached my configuration file along with this)
>>>
>>> Thanks in advance
>>> -sahan
>>>
>>> _______________________________________________
>>> click mailing list
>>> click at amsterdam.lcs.mit.edu
>>> https://amsterdam.lcs.mit.edu/mailman/listinfo/click
>>>
>>
> 
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click