Testing XDP vs DPDK

Question

I do have some experience with DPDK but currently I'm reading many blogs about XDP. I am trying to compare both technologies and understand the differences between DPDK and XDP. This raises some questions. I hope someone can help me with the following questions:

1. With DPDK I can map workers to CPU cores and isolate CPU cores which will be used by DPDK. In case of eBPF / XDP, which CPU cores are used? Are all available CPU cores used? Would it be possible to isolate CPU cores meant for eBPF / XDP programs?
2. When I test the throughput from a DPDK application, I'm able to check whether ring buffers (mempools) are full so packets will be lost. But how can I check whether an eBPF / XDP program causes packet drops because the throughput is too high? I assume when an eBPF / XDP program takes too much time to process a packet, eventually you will see packet drops? (especially when sending 64B packets on a high rate to find the maximum number of packets that can be send)

Thank you in advance for your help!

Worriment

Answer 1

With DPDK I can map workers to CPU cores and isolate CPU cores which will be used by DPDK. In case of eBPF / XDP, which CPU cores are used?

Answer: XDP with eBPF runs in kernel space, unlike DPDK user space.

Are all available CPU cores used?

Answer: Yes, but normally irqbalance or interrupt pinning will put the RX queue of the port on to specific core.

Would it be possible to isolate CPU cores meant for eBPF / XDP programs?

Answer: You are referring to KERNEL_CMD_LINE option isol, the understanding is incorrect. As mentioned above you can pin the interrupt of RX queue forcing to run eBPF XDP on that core.

When I test the throughput from a DPDK application, I'm able to check whether ring buffers (mempools) are full so packets will be lost. But how can I check whether an eBPF / XDP program causes packet drops because the throughput is too high?

Answer: you have use a mix of NIC and eBPF counters to achieve the same

I assume when an eBPF / XDP program takes too much time to process a packet, eventually, you will see packet drops? (especially when sending 64B packets on a high rate to find the maximum number of packets that can be send)

Answer: not necessary true, best performance of XDP is with zero-copy Driver to user space. Running application thread on a separate core gives an almost comparable performance as DPDK (tested with 2 * 10Gbps - 95% of DPDK performance).

Broadfaced

Great explanation! You have explained it clearly and in a way that is easy to understand. Thank you for sharing your knowledge. – Dissident