F1.4xlarge P2P bandwidth

Question

Hi,

I'm new to AWS FPGA development. I notice that AWS F1.4xlarge (and F1.16xlarge) instances support effective P2P communication between FPGAs, without the help of the host. But the example from Xilinx (https://github.com/Xilinx/Vitis_Accel_Examples/tree/f61637e9263ecc1be3df34182ea6c53a0ca10447/performance/p2p_fpga2fpga_bandwidth) seems not able to work on AWS F1 instances.

I have read the code at https://github.com/awslabs/aws-fpga-app-notes/tree/master/Using-PCIe-Peer2Peer. But **how can I evaluate the bandwidth of P2P communication bandwidth concisely?** In the code provided, the program just uses `sleep(1)` to wait until the write is finished. Are there some signals that are triggered when the write is finished?

The result I hope to obtain is a table like the one at https://github.com/awslabs/aws-fpga-app-notes/tree/master/Using-PCIe-Write-Combining. Could you please provide the corresponding script or any useful advice?

Thanks a lot.

Answer

Dear customer

Thank you so much for your interest in AWS. We dont provide any perf numbers measurement for P2P as it is CL design dependent. However a customer can actually devise PCIM write/read counters in their design along with timers and measure the performance. This requires counter implementation in CL. Please feel free to let us know if you have any additional questions or if your original question had additional points not addressed.

Thanks

F1.4xlarge P2P bandwidth

Relevanter Inhalt