Re: questions on NAPI processing latency and dropped network packets

[Eric Dumazet <dada1@cosmosbay.com>]

Chris Friesen a écrit :
> Eric Dumazet wrote:
>> Chris Friesen a écrit :
>>
>>> I've done some further digging, and it appears that one of the 
>>> problems we may be facing is very high instantaneous traffic rates.
>>>
>>> Instrumentation showed up to 222K packets/sec for short periods (at 
>>> least 1.1 ms, possibly longer), although the long-term average is 
>>> down around 14-16K packets/sec.
>>
>>
>> Instrumentation done where exactly ?
> 
> I added some code to e1000_clean_rx_irq() to track rx_fifo drops, total 
> packets received, and an accurate timestamp.
> 
> If rx_fifo errors changed, it would dump the information.
> 
>>> Is there anything else we can do to minimize the latency of network 
>>> packet processing and avoid having to crank the rx ring size up so high?
> 
>> You have some tasks that disable softirqs too long. Sometimes, bumping 
>> RX ring size is OK (but you will still have delays), sometimes it is 
>> not an option, since 4096 is the limit on current hardware.
> 
> I added some instrumentation to take timestamps in __do_softirq() as 
> well.  Based on these timestamps, I can see the following code sequence:
> 
> 2374604616 usec, start processing softirqs in __do_softirq()
> 2374610337 usec, log values in e1000_clean_rx_irq()
> 2374611411 usec, log values in e1000_clean_rx_irq()
> 
> In between the successive calls to e1000_clean_rx_irq() the rx_fifo 
> counts went up.
> 
> Does anyone have any patchsets to track down what softirqs are taking a 
> long time, and/or who's disabling softirqs?
> 

Not for linux-2.6.10 unfortunatly.

Check net/ipv4/route.c, where many improvements can be done, especially if you 
have a large rt cache

grep . /proc/sys/net/ipv4/route/*