Re: [dpdk-dev] [EXT] Re: [RFC PATCH v2] app/testpmd: tx pkt clones parameter in flowgen

[Igor Russkikh <irusskikh@marvell.com>]

>> Here instead of recreating each packet separately, we use clones counter
>> to resend the same mbuf to the line multiple times.
>>
> Do you have any numbers on how much performance improvement gained?

Hi Ferruh,

Yes, I'll post that in v1. In general, on our 100G device in default
configuration I had to use 8-16 queues (read cores) to reach line rate on 1400
packet size. Will give exact numbers.

With clones I was able to see linerate on 1-2 tx queues. That of course
depends on if HW itself allows to handle that amount of traffic.

Reaching max PPS on small packet sizes obviously helps here as well.

>> +					rte_exit(EXIT_FAILURE,
>> +						 "clones must be >= 0 and <= %d (burst)\n",
>> +						 nb_pkt_per_burst);
> 
> Do you need to enforce the "n <= nb_pkt_per_burst", burst value can be 
> changed
> later and trying to keep 'clones' values in sync with it is additional work.
> 
> In the flowgen logic, with each burst a new packet is created anyway. So 
> instead
> of enforcing the 'clones' number range, in documentation you can say the 
> clone
> number can't exceed the burst number whatever it is set.

Looks reasonable, thanks.

>>   uint16_t nb_pkt_per_burst = DEF_PKT_BURST; /**< Number of packets per 
>> burst. */
>> +uint16_t nb_pkt_clones; /**< Number of tx packet clones to send. */
>>   uint16_t mb_mempool_cache = DEF_MBUF_CACHE; /**< Size of mbuf mempool 
>> cache. */
>>
> 
> Both the parameter name, 'clones', and the variable name 'nb_pkt_clones' are 
> too
> generic, and may mislead users. Please remember that testpms usage is very 
> wide.
> You are updating very specifically the flowgen forwarding engine, can you 
> please
> prefix the 'flowgen', like:
> 'flowgen-clones' & 'nb_pkt_flowgen_clones'.
> Also explicitly mention in the description that this is for flowgen clones.

Reasonable as well. Initially I was thinking about applying that to "tx-only"
as well, but the decided to stay with flowgen as more generic mode.

Thanks,
  Igor