[Click] FromDevice (OS-X) buffering up packets...

[Nicholas Weaver]

This case however its not from the NIC FIFO, as a TCPdump running on the same interface at the same time doesn't experience the many second queuing delay.

On Jul 20, 2010, at 10:57 AM, Giorgio Calarco wrote:

> hi all,
> 
> having bursts of packets and increased jitter does not only depends on
> the Click itself, but also from the queueing at the NIC FIFO. The PCI
> bus is indeed fair (in the sense that it serves each peripheral
> with a token passing policy), but once the peripheral acquires
> access onto the bus, it's not time-limited in any way. If the bus
> occupancy probability is high, for any reason, the NICs tend to accumulate
> packets in their internal FIFO and once they can start the DMA
> transfer, they move all the packets back-to-back, generating
> a burst. I 'm not sure if this is your case (which hardware are you using ?
> could the bus possibly be close to saturation ?),
> anyway, you can find more details in one of the second (?) chapther
> of my past PhD
> thesis (it's published on the Click site - Publications).
> Hope it can help in some way.
> 
> Ciao, Giorgio
> 
> On Tue, Jul 20, 2010 at 6:55 PM,  <rchertov at cs.ucsb.edu> wrote:
>> On 5:25 pm 07/19/10 <rchertov at cs.ucsb.edu> wrote:
>>> I am seeing similar behavior on Linux as well.  In my case, I carry some
>> 
>> Just wanted to state that the behavior on Linux is similar but not as bad
>> as Nick is reporting for OSX.  I get about 20ms worth of extra jitter by
>> using the To/FromDevice compared to sending the data via the RawSocket.
>> 
>>> data between two PCs which run user-level click.  If I use RawSocket
>>> element, I get much better performance compared to using To/FromDevice
>>> and manually composing the frames.
>>> 
>>> 
>>> On 10:42 am 07/19/10 Nicholas Weaver <nweaver at icsi.berkeley.edu> wrote:
>>>> The test configuration:
>>>> 
>>>> FromDevice(en0, PROMISC true) -> CheckIPHeader(14) ->
>>>>            IPPrint('reply') -> Discard;
>>>> 
>>>> 
>>>> A TCPdump running at the same time gets all the packets as they are
>>>> received, but the click configuration gets them in bursts (but the
>>>> timestamps are right)
>>>> 
>>>> On Jul 19, 2010, at 10:38 AM, Nicholas Weaver wrote:
>>>> 
>>>>> I remember having a similar problem before and I don't remember what
>>>>> the solution is...
>>>>> On OS-X (latest version, latest devtools), FromDevice is buffering
>>>>> up a large number of packets rather than receiving them in real
>>>>> time. A TCPdump running at the same time does not have this problem.
>>>>> 
>>>>> 
>>>>> Suggestions?
>>>>> 
>>>>> 
>>>>> 
>>>> 
>>>> 
>>>> _______________________________________________
>>>> 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
>> 
>> _______________________________________________
>> 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
>