[dpdk-users] Possibly wrong RX statistics reported by virtio driver (TRex generator)

[dpdk-users] Possibly wrong RX statistics reported by virtio driver (TRex generator)

[Nikos Anastopoulos]

Hi list,



The TRex traffic generator's folks recommended us to contact you for an issue that might relate to DPDK virtio driver (https://groups.google.com/forum/#!topic/trex-tgn/LxtTV7G43AE).



Specifically, our scenario involves running TRex inside a KVM instance with 2 vhost-user ports, which are connected in loopback mode via an OVS-DPDK (i.e. whatever sent through port1 goes to port2, and vice versa). We observe that the received traffic volume reported by TRex is much less than the transmitted (it is constantly 4B per packet), whereas:

1) the count of the received packets is the same

2) the contents of the received packets, as captured _inside_ the VM, are the same with the transmitted



Specifically, we experimented with different packet sizes being transmitted, i.e. 64B, 150B, 1000B, yet in the receive path the reported incoming volume corresponded constantly to 4B-per-packet, for all cases.



The TRex engineers suggested that it seems like an RX counters issues in virtio driver.



Experimental setup:

- TRex v2.45 (uses DPDK 17.11 AFAIK), within a QEMU v2.12

- on the host: OVS 2.10.0 with DPDK 17.11.3



Thanks,

Nikos

Re: [dpdk-users] Possibly wrong RX statistics reported by virtio driver (TRex generator)

[Stephen Hemminger]

On Wed, 31 Oct 2018 10:29:38 +0200
"Nikos Anastopoulos" <nanastop@intracom-telecom.com> wrote:

> Hi list, 
> 
>  
> 
> The TRex traffic generator's folks recommended us to contact you for an issue that might relate to DPDK virtio driver (https://groups.google.com/forum/#!topic/trex-tgn/LxtTV7G43AE). 
> 
>  
> 
> Specifically, our scenario involves running TRex inside a KVM instance with 2 vhost-user ports, which are connected in loopback mode via an OVS-DPDK (i.e. whatever sent through port1 goes to port2, and vice versa). We observe that the received traffic volume reported by TRex is much less than the transmitted (it is constantly 4B per packet), whereas:
> 
> 1) the count of the received packets is the same
> 
> 2) the contents of the received packets, as captured _inside_ the VM, are the same with the transmitted
> 
>  
> 
> Specifically, we experimented with different packet sizes being transmitted, i.e. 64B, 150B, 1000B, yet in the receive path the reported incoming volume corresponded constantly to 4B-per-packet, for all cases. 
> 
>  
> 
> The TRex engineers suggested that it seems like an RX counters issues in virtio driver. 
> 
>  
> 
> Experimental setup: 
> 
> - TRex v2.45 (uses DPDK 17.11 AFAIK), within a QEMU v2.12
> 
> - on the host: OVS 2.10.0 with DPDK 17.11.3
> 
>  
> 
> Thanks,
> 
> Niko

No idea, try dumping a sample of packets in virtio_update_packet_stats