Buffer Allocation Failed on U280 when Running on Larger Graphs

[yhzhoucs]

Description
Hi! I cloned the code from branch 'v_HBM' and attempted to execute BFS on U280 FPGA. The execution went smoothly with the provided dataset (rmat-14-32.txt). However, when I tried to run it with larger graphs——specifically, using soc-LiveJournal1.txt downloaded from the Stanford Network Dataset Collection——the program crashed during buffer allocation on the U280. Below is the error log:

[XRT] ERROR: std::bad_alloc
./main.cpp:86 Error calling edge_buffers.emplace_back(cl::Buffer(context, flags, edgesHeadArrays[i].size, &(edgesHeadArrays[i].ext_attr), &status)), error code is: -5

I've made a few adjustments to the host code by rewriting it using OpenCL2, but it's not the cause of the error. It appears that the issue lies in XRT's inability to allocate a buffer larger than 256MB with HBM.

Referrence:

• https://support.xilinx.com/s/question/0D52E00006hpYHiSAM/when-i-use-clcreatebuffer-for-physical-memory-but-this-function-return-error-code6?language=en_US
• https://support.xilinx.com/s/question/0D52E00007ECLNySAP/how-to-configure-data-layout-across-multiple-hbm-channels?language=en_US

So, I'm wondering it may be hard for the accelerator to support large graph processing as long as the size of edgesHeadArray exceeds 256 MB.

To Reproduce
Steps to reproduce the behavior:

1. Clone the v_HBM code.
2. Compile the project.
3. Download graph soc-LiveJournal from SNAP, extract it to dataset folder.
4. Run command ./host_graph_fpga_bfs xclbin_hw_bfs/graph_fpga.hw.xilinx_u280_gen3x16_xdma_1_202211_1.xclbin dataset/soc-LiveJournal1.txt.
5. See error.

Platform:

• Device: U280 FPGA
• OS: Ubuntu 18.04.6 LTS

Your assistance in resolving this issue is greatly appreciated!