[dpdk-users] performance difference between ixgbe and i40e

[dpdk-users] performance difference between ixgbe and i40e

[János Pauer]

Hi all,

I am using OVS with dpdk v16.11. I have two setups for performance measurement, which are almost identical, the only difference is that one setup has i40e driven 10G NICs and the other one has ixgbe driven 10G NICs.
I have MoonGen for packet generating, and testPMD to reflect the traffic. The measurement results are much higher with the ixgbe-driven card. It's about +20% with smaller packet sizes, and more than +40% with bigger packet sizes.
Does anyone have any idea what can cause this? Is there any significant difference between the mentioned drivers?

Thanks and Regards,
János

Re: [dpdk-users] performance difference between ixgbe and i40e

[tom.barbette]

Hi János,

We actually saw the reverse behavior on our end. Does dmesg complains about your card not being on a PCI Express 3 x8 port? To achieve full 40G it is a requirement (assuming you have a 4X10G setup).

Tom

----- Mail original -----
> De: "János Pauer" <janos.pauer@ericsson.com>
> À: users@dpdk.org
> Envoyé: Mardi 16 Janvier 2018 16:01:31
> Objet: [dpdk-users] performance difference between ixgbe and i40e

> Hi all,
> 
> I am using OVS with dpdk v16.11. I have two setups for performance measurement,
> which are almost identical, the only difference is that one setup has i40e
> driven 10G NICs and the other one has ixgbe driven 10G NICs.
> I have MoonGen for packet generating, and testPMD to reflect the traffic. The
> measurement results are much higher with the ixgbe-driven card. It's about +20%
> with smaller packet sizes, and more than +40% with bigger packet sizes.
> Does anyone have any idea what can cause this? Is there any significant
> difference between the mentioned drivers?
> 
> Thanks and Regards,
> János

Re: [dpdk-users] performance difference between ixgbe and i40e

[János Pauer]

Hi Tom,

Thank you for your reply. All the cards are in a PCIe 3x slot, but my setup is overall 10G. I have 1x 10G on the generator side, and 2x10G on the echo side, using bond mode. Did you do some meadurement in 1X10G environment?

Thanks and Regards,
János

-----Original Message-----
From: tom.barbette@uliege.be [mailto:tom.barbette@uliege.be] 
Sent: 2018. január 16., kedd 17:28
To: János Pauer <janos.pauer@ericsson.com>
Cc: users@dpdk.org
Subject: Re: [dpdk-users] performance difference between ixgbe and i40e

Hi János,

We actually saw the reverse behavior on our end. Does dmesg complains about your card not being on a PCI Express 3 x8 port? To achieve full 40G it is a requirement (assuming you have a 4X10G setup).

Tom

----- Mail original -----
> De: "János Pauer" <janos.pauer@ericsson.com>
> À: users@dpdk.org
> Envoyé: Mardi 16 Janvier 2018 16:01:31
> Objet: [dpdk-users] performance difference between ixgbe and i40e

> Hi all,
> 
> I am using OVS with dpdk v16.11. I have two setups for performance 
> measurement, which are almost identical, the only difference is that 
> one setup has i40e driven 10G NICs and the other one has ixgbe driven 10G NICs.
> I have MoonGen for packet generating, and testPMD to reflect the 
> traffic. The measurement results are much higher with the ixgbe-driven 
> card. It's about +20% with smaller packet sizes, and more than +40% with bigger packet sizes.
> Does anyone have any idea what can cause this? Is there any 
> significant difference between the mentioned drivers?
> 
> Thanks and Regards,
> János

[dpdk-users] performance difference between ixgbe and i40e

[longtb5]

Hi Tom,

How do one calculate the theoretical limit of a 4x10G on a PCIe 3 x8 vs a
PCIe 2 x8?
Currently in our test server we have two 2x10G NICs on loopback with each
other and I couldn't reach
4x10G throughput (total of 80G for both direction). 

NIC 1 ----------------------------- NIC 2
port 0 <-------------------------> port 0
port 1 <-------------------------> port 1

Regards,
BL.