Re: [dpdk-dev] Performance of rte_ring APIs

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, Aug 07, 2015 at 10:07:07AM -0700, Gopakumar Choorakkot Edakkunni wrote:
> Hi All,
> 
> I have an extremely simple test - I have just one single DPDK EAL
> thread which pulls packets from one 10G port and just puts the packet
> exactly as is (no changes) on another 10G port - I get 9.5million pps,
> so far so good. So its like this
> 
> dpdk_rx
> dpdk_tx
> 
> 9.5 Millionpps
> 
> Now I do the below and the performance comes down to like 4 Million
> pps, less than half !
> 
> dpdk_rx
> rte_ring_mc_dequeue_bulk(my_ring1, my_array, nb_rx)
> dpdk_tx
> rte_ring_mp_enqueue_bulk(my_ring1, my_array, nb_rx)
> 
> Note that I do nothing with the things I dequeue from the ring, I just
> enqueue it back. So is there any gotchas in using these rings ? I am
> sure I am missing something, it cant drop from 9.5Mpps to 4Mpps just
> because of a dequeue/enqueue ?
> 
You are using the multi-producer and multi-consumer ring functions, which tend to
be much more expensive than the single producer single consumer function. Depending
on the contention on those rings, performance can vary considerably in your application.

/Bruce