IRN's packet processing logic synthesized using Xilinx Vivado HLS
This project is maintained by NetSys
This repository contains the Xilinx Vivado HLS code for synthesizing IRN’s packet processing logic, as a proof-of-concept for its implementation feasibility. It reproduces the results presented in Sec 6.2 of IRN’s paper, titled “Revisiting Network Support for RDMA”, that appeared in SIGCOMM’18.
There are two folders in the repository:
irn_receiver: This folder contains the code for the receiver side module at the responder, namely receiveData.
irn_sender: This folder contains the code for sender side modules at the requester, namely txFree, receiveAck and timeout.
Each folder contains scripts to run synthesis, RTL verification and exporting for each module. A sample trace for input to the respective modules and the output, generated via simulations, has also been provided for the verification.
The code assumes availability of Xilinx Vivado HLS 2017.2 (or any other HLS version that supports the target device Kintex Ultrascale XCKU060).
The code can also be used with the free Xilinx Vivado Webpack edition, by changing the target device specified in the run_hls*.tcl
files inside the irn_receiver and the irn_sender folders to one of those that are supported by the Webpack edition, and editing the following line to match the clock rate supported by the device. However, the synthesis result would then differ from what is reported in the paper.
Go to irn_receiver folder. All path names in this section are specified relative to this folder.
Update the second line in file ‘megarun.sh’ to set the correct path to the licence file.
Run ‘megarun.sh’.
A new folder called rdma_receiver, that contains the synthesis results, is created within the irn_receiver directory.
The synthesis report can be found in rdma_receiver/solution-128/syn/report/receiveData_csynth.rpt. It provides the performance and resource usage results that were used to generate the first row of Table 2, in Section 6.2 of the paper.
A trace output, generated by the testbench, can be found at rdma_receiver/solution-128/sim/wrapc/output.txt and can be compared against traces/recvOutput-40Gbps-incast.txt. The latter is the output trace generated by the receiver nodes in the simulator used in Section 4 of the paper for the same input packet trace that was supplied to the Vivado HLS code. Generation of an identical output file by the HLS code indicates that it follows the same transport logic as the simulator code. More details about these traces can be found in traces/description.txt.
Go to irn_sender folder. All path names in this section are specified relative to this folder.
Update the second line in file ‘megarun.sh’ to set the correct path to the licence file.
Run ‘megarun.sh’.
For each $module in (txFree, receiveAck, timeout), new folders that contain the respective synthesis results are created within irn_sender. These folders are named rdma_sender_$module (i.e. rdma_sender_txFree, rdma_sender_receiveAck, and rdma_sender_timeout respectively).
For each $module in (txFree, receiveAck, timeout), the synthesis report can be found in rdma_sender_$module/solution-128/syn/report/$module_csynth.rpt. They provide the performance and resource usage results that were used to generate the second, third and fourth rows of Table 2, in Section 6.2 of the paper.
For each $module in (txFree, receiveAck, timeout), a trace output, generated by the testbench, can be found at rdma_sender_$module/solution-128/sim/wrapc/output.txt and can be compared against traces/sendOutput-40Gbps-incast.txt. The file traces/sendOutput-40Gbps-incast.txt needs to be created by concatenating traces/sendOutput-40Gbps-incast-part1.txt with traces/sendOutput-40Gbps-incast-part2.txt. It is the output trace generated by the sender nodes in the simulator used in Section 4 of the paper for the same input trace that was supplied to the Vivado HLS code. More details about these traces can be found in traces/description.txt.