PFE performance is lower compared to GMAC on S32g3 RDB3

[kamal_n]

We are trying to measure the performance of PFE and GMAC on S32G3 using udp communication.

Test setup:

Board 1 : RDB3 with BSP43 

Board 2: RDB3 with BSP43

Both RBD connected using ethernet cable and measured.

Buffer size: 64 bytes

Tried with simple UDP client server application. udp_client running in Board1 and ,udp_server running in Board2. It is simple ping pong udp test and measures the round-trip time of udp packets.

We tried four combinations and results were attached for those combinations.

ETH<->ETH, ETH<->PFE2, PFE2<->ETH, PFE2<->PFE2

From the results we could find that PFE to PFE communication is very less performance compared to other combination. We are expecting PFE communication should be higher compared to GMAC.

[alejandro_e]

Hello @kamal_n,

Thanks for reaching out to us. Could you give me more details regarding the test? Please share the following:

• The used configuration for each MAC interface (SGMII or RGMII)
• the output of the following command in u-boot:

  => printenv

• The exact  of the RDB3 ports you are using in each test for both boards.
• Any change related to the SerDes (in case of SGMII) in your setup
• Any change change related to the SJA1110 in your setup
• Any change related to network in Linux in your setup

Thanks in advance

[kamal_n]

Hi @alejandro_e ,

Please find below details

1. MAC configuration RGMII

2. Attached printenv results as file

3. 1000Base-T port used with RGMII for GMAC0 and PFE_MAC2

4. No changes done specific to SJA1110 and we are not using SJA1110 

5. No network specific changes done from our side.

We have tested using buffer size of 64 bytes and tried also 1024 bytes.

Can you suggest the size of buffer to note difference in performance in PFE.

Thanks in advance.

[alejandro_e]

Hello @kamal_n,

Thanks for all the details, please allow me some time to replicate your tests and confirm I see the same behavior, at the same time I will search more information regarding the latency of PFE in linux.

Thanks for your patience.

[alejandro_e]

Hello @kamal_n,

I was able to reproduce your results, I only tested with PFE-PFE and GMAC-GMAC, getting similar results, PFE has a lower performance compared to GMAC.

I was not able to find anything clear in the documentation, I will share this topic with the internal team and wait for their feedback.

Thanks in advance for your patience.

[alejandro_e]

Hello @kamal_n,

Thanks for your patience, the internal team has shared the following:

"

When not using L2-Bridge configuration (using PFE as a normal network interface) the latency of PFE is worse than GMAC theoretically. It is because PFE will try to check the configuration to see if the packets hit the rules and it takes time to do it.

The advantage of PFE is that it can be used as a switch or router, which offloads the CPU. The classic scenario is that you could send frames from PC1 to EMAC1 and PFE will forward it to EMAC0 or EMAC2 according to your configuration. In this case, CPU does not participate in forwarding and latency would be definitely short than the case that CPU handle the forwarding. 

" 

To configure PFE bridging, please check the PFE_S32G_A53_LNX_UserManual.pdf, which you can find in FlexNet packed in PFE-DRV_S32G_A53_LNX_1.9.0_DOC.zip. In particular check sections:

3 Build Procedure

2.10 FCI use cases for Linux

Please let me know if you encounter any problem while configuring the PFE bridging.