TLDR Tech / Practical
Kernel migration of GPU server request loop's CPU impact studied
Scientists explored if relocating a GPU server's request loop into the kernel via eBPF and io_uring could decrease CPU usage and latency. Their prototype, bpf-gpu-server, performed comparably to a strong C++ epoll setup, showing no performance gains. The analysis indicated that the TCP/IP stack is the main CPU bottleneck per request, and moving the loop from user space to kernel space simply transfers time from user to system.
Read the full story