Re: Performance regression on kernels 3.10 and newer

[Alexander Duyck <alexander.h.duyck@intel.com>]

On 08/14/2014 12:59 PM, Rick Jones wrote:
> On 08/14/2014 11:46 AM, Eric Dumazet wrote:
> 
>> I believe you answered your own question : prequeue mode does not work
>> very well when one host has hundred of active TCP flows to one other.
>>
>> In real life, applications do not use prequeue, because nobody wants one
>> thread per flow.

My concern here is that netperf is a standard tool to use for testing
network performance, and the kernel default is to run with
tcp_low_latency disabled.  As such the prequeue is a part of the
standard path is it not?  If the prequeue isn't really useful anymore
should we consider pulling it out of the kernel, or disabling it by
making tcp_low_latency the default?

>> Each socket has its own dst now route cache was removed, but if your
>> netperf migrates cpu (and NUMA node), we do not detect the dst should be
>> re-created onto a different NUMA node.
> 
> Presumably, the -T $i,$j option in Alex's netperf command lines will
> have bound netperf and netserver to a specific CPU where they will have
> remained.
> 
> rick jones

Yes, my test was affinitized per CPU.  I was originally trying to test
some local vs remote NUMA performance numbers.  Also as I mentioned I
was using the ixgbe driver with 82599 and I had ATR enabled so the
receive flow was affinitized to the queue as well.  We shouldn't have
had any cross node chatter as a result of that.

Thanks,

Alex