I am using Mlnx CX-5 and CX-6 Dx cards in my lab. I have enabled PFC with 8 TCs and configured QoS with DSCP mode. I am using DPDK pktgen to generate packets with 100 Gbps (CX-6 Dx) per PHY port. I can see the receiver card is dropping the packets by seeing rx_discard packets counter.
I am assuming the PFC should be lossless for UDP data and whenever the Rx buffer fills, it should generate PFC pause frames to stop or pause data on the sender side but I can’t see this effect.
Can someone guide me how to setup Mlnx cards to work as lossless traffic device for UDP packets.
Welcome to the NVIDIA Developer forums. I am going to move your topic to the Networking category so the support team has visibility.
Hi, Thanks for moving to the appropriate place. I would be highly thankful for the kind reply.
Someone from the Networking team will jump in here to help ASAP.
Can you please provide the output of the following command from the interfaces used in the test:
mlnx_qos -i ethernet_interface
Do you have a switch in between the nodes or they are connected in back to back?
With what priority are you marking the outgoing traffic (DSCP value in the IP header)? You can use tcpdump to review if your are not sure.
Here are the details:
No switch in between. Two ConnectX-6 Dx cards are connected back to back.
Different DSCP values but one at a time. For example 0, 8, 16, 24, 32
OS: CentOS 8.5
Latest Ofed version
Some more information:
After reboot, when QoS configurations are done, we see the PFC pause packets. But if we make any change to QoS configurations like UP-TC mapping, DSCP to UP or change buffer sharing, then we can’t receive any PFC pause packets.
But still, if we analyze the data at the time of PFC pause packets, we are unable to stop completely the sending data. Rather I can see the sender server contineosly sends the data. When it receives the PFC pause packets, it reduces the bandwidth from 100Gbps to 24Gbps but can’t reach to 0.
As I have made changes to the Rx driver to do empty the Rx Queue, that’s why I am assuming that the data should not be lost. But I can see the Rx buffer or Queue drops the data which is not a lossless effect.
Hope this will provide you enough information to analyze what happens.
In general we do not recommend touching the receive buffer on ConnectX-6 Dx devices. I can see how it might break the PFC.
I suggest you start simple with standard mlnx_qos config (in trust DSCP mdoe) and enabling one or more (NOT all) priorities with PFC. Do not change any buffer sizes and do not remap buffer and priorities. The (shared) buffer architecture in ConnectX-6 Dx should be able to handle that.
Your UP-TC mapping should be ok.
I can see the same behaviour with default QoS settings. Here are the QoS configurations
Here are the ethtool stats on both servers
See ethtool statistics in below attached image.
By seeing the PFC packets, it generates PFC packets with timeout 65355 and after some time generate the same packet with timeout 0. This sequence continues.
The question I have is, why the Rx buffer drops data in PFC mode? Is there a fix timeout after which the Rx buffer become empty in the PFC mode or the data remain in the Rx buffer untill it is read by the Rx Queue?
In the data you provided I do not see any drops on the RX buffer.
In general, the way it works is that when we reach some threshold we will issue an xoff pause and on a different threshold we will issue an xon pause
I suggest that you approach technical support to review further (as more debug data is required).
Can you tag the technical team here or assign them this ticket or guide me how to contact them please.
Please send an email to firstname.lastname@example.org with the details and they will guide you through the process.