[dpdk-users] TX-dropped is high while sending custom packet via testpmd app

[dpdk-users] TX-dropped is high while sending custom packet via testpmd app

[Nilesh]

Hello,
     we are trying to send packets from the testpmd to another machine 
with DPDK running.
     We are building custom packets before sending it on the wire.
     After running the application the TX-dropped is quite high (can be 
inferred from following logs : )


$ sudo ./testpmd -c f -w 01:00.1  --   --nb-cores=2 
--eth-peer=0,A4:BF:01:37:23:AB --rxq=1 --txq=1


Port statistics ====================================
  ######################## NIC statistics for port 0 
########################
  RX-packets: 1445150    RX-missed: 0          RX-bytes:  86709064
  RX-errors: 0
  RX-nombuf:  0
  TX-packets: 1602045    TX-errors: 0          TX-bytes:  134571780

  Throughput (since last show)
  Rx-pps:            0
  Tx-pps:            0
  ############################################################################

  ---------------------- Forward statistics for port 0 
----------------------
  RX-packets: 1445160        RX-dropped: 0             RX-total: 1445160
  TX-packets: 1602045        TX-dropped: 694971472     TX-total: 696573517
  ----------------------------------------------------------------------------

  +++++++++++++++ Accumulated forward statistics for all 
ports+++++++++++++++
  RX-packets: 1445160        RX-dropped: 0             RX-total: 1445160
  TX-packets: 1602045        TX-dropped: 694971472     TX-total: 696573517
  ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++


What could be the reason for such packet drop? What tuning of 
buffer/queue size affects this ?

How to set optimal setting for high performance?


System specification :
2 machines with
Distributor ID:    Ubuntu
Description:    Ubuntu 16.04.5 LTS
Release:    16.04
Codename:    xenial

Kernel: 4.15.0-29-generic
DPDK: 19.05.0

Hardware:
CPU: Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz

NIC 1: Intel Corporation Ethernet Controller XL710 for 40GbE QSFP+ (i40e)
NIC 2: Intel Corporation I350 Gigabit Network Connection (igb)

Regards,
Nilesh

Re: [dpdk-users] [dpdk-dev] TX-dropped is high while sending custom packet via testpmd app

[Bruce Richardson]

On Tue, Jul 16, 2019 at 03:01:12PM +0530, Nilesh wrote:
> Hello,
>     we are trying to send packets from the testpmd to another machine with
> DPDK running.
>     We are building custom packets before sending it on the wire.
>     After running the application the TX-dropped is quite high (can be
> inferred from following logs : )
> 
> 
> $ sudo ./testpmd -c f -w 01:00.1  --   --nb-cores=2
> --eth-peer=0,A4:BF:01:37:23:AB --rxq=1 --txq=1
> 
> 
> Port statistics ====================================
>  ######################## NIC statistics for port 0 ########################
>  RX-packets: 1445150    RX-missed: 0          RX-bytes:  86709064
>  RX-errors: 0
>  RX-nombuf:  0
>  TX-packets: 1602045    TX-errors: 0          TX-bytes:  134571780
> 
>  Throughput (since last show)
>  Rx-pps:            0
>  Tx-pps:            0
>  ############################################################################
> 
>  ---------------------- Forward statistics for port 0 ----------------------
>  RX-packets: 1445160        RX-dropped: 0             RX-total: 1445160
>  TX-packets: 1602045        TX-dropped: 694971472     TX-total: 696573517
>  ----------------------------------------------------------------------------
> 
>  +++++++++++++++ Accumulated forward statistics for all ports+++++++++++++++
>  RX-packets: 1445160        RX-dropped: 0             RX-total: 1445160
>  TX-packets: 1602045        TX-dropped: 694971472     TX-total: 696573517
>  ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> 
> 
> What could be the reason for such packet drop? What tuning of buffer/queue
> size affects this ?
> 

Those stats for TX-dropped just look wrong to me. Given that you received
only 1.45 million packets, having dropped nearly 700 million doesn't make
sense. Even the Tx packet counts - though more reasonable - are higher than
the received count.

In terms of testpmd settings, the rxq=1 and txq=1 settings are the default
so aren't needed, and setting the number of forwarding cores to 2 shouldn't
do anything as there is only a single receive queue which can't be shared
among cores.

Regards,
/Bruce