Intel E810 100Gbps very poor performance with DPDK

[marcof1]

Hello,

I’m experiencing very poor performances with Intel E810 100Gbps NIC. NIC performs really poor with DPDK driver (only a few kpps performing testpmd forwarding).

I’m using firmware 4.10, for some reason I can’t update it to 4.30, the inventory fails with “Access error” message. I’m using a server with Linux 5.4 kernel.

Detalis about the testbed:

The Intel E810 NIC is connected back to back with a NVIDIA Bluefield-2 100Gbps.
I’m using dpdk 22.07, NVM version 4.10, DDP version 1.3.30 as suggested by this guide: https://edc.intel.com/content/www/us/en/design/products/ethernet/config-guide-e810-dpdk/software-firmware-requirements/
I’d like to try also the NVM v4.20 but, as I told you, the update package fails. Maybe the kernel (v5.4) that I’m using is too old?
Here is a screenshot: the Intel NIC is performing testpmd and forwarding traffic which is sent from the NVIDIA NIC using Trex traffic generator.
Trex is sending roughly 2Mpps on port 0, but only few of them (roughly 235Kpps) are then received on port 1 (see attachment).
I noticed that the Intel NIC is actually not even receiving many of the packets it has to forward, this is very important as it could be the core problem in my case. See attachment, testpmd says port 0 only received 9’451’227 packets, while Trex has sent 76’348’324 packets. This is roughly a proportion of 1 to 10, the same that we can see between the packets that Trex sends and receives.
You can find in the attachments three screenshots: testpmd start, testpmd output, Trex traffic generator output.
Are you aware of any problem with these NICs? Or any other best practice. should use?

I’m attaching screenshots, any help is appreciated.

Thanks

[Srri]

Hi Marco,

We found duplicate of your request so we will proceed to close this community case. We will be working on 06034145 and will follow up on that case.

Best Regards,

Srri

Intel Customer Support

[evnech]

Hello. Are there any updates about this topic? We have some performance issues during testpmd test for different Intel CPU generation (4,5) and for any NVM versions of 810/830 cards. Every times we have met with TX drops on cards, especially for platforms with 3 and more cards in platform it shows more drops. We can not get line rates for cards without drops. For example 800 Gbps for 1000 bytes packets through all 8 ports of E830 on platform or 400Gbps for 1000 bytes packets through all 8 ports of E810 on platform (this is just 100 mpps - not so many for 8581 CPU). We have tested different platforms. And every time get the same result - TX drops. We use to investigate missed errors if our DPDK app can not handle packets. But for cards 800 series every time we get TX drops. Any proposals or recommendations for 800 series cards?

[Sazzy_Intel]

To ensure you receive a proper assistance from Intel, we recommend you to open new community thread post regarding this enquiry since this is old thread post. Once created, we will assist you from there.