benchmarks.md

Benchmarks

Linux native 5.19
AMD Ryzen 5800X @ 4.50 GHz

All echo servers pinned to CPU 0
Echo bench threads pinned to CPU 1-4
Each echo bench thread uses epoll

Comparisons

• xe
• l4cpp (aka liburing4cpp)
• photon
• frevib
• epoll (native, no framework)

Bottom of chart cut off to exaggerate differences

Notes

• xe is compiled without -flto

• xe, l4cpp, photon, and epoll use 512 byte buffers

• xe, photon, and epoll align buffers to 512 byte boundary

• frevib uses io_uring_prep_provide_buffers

• frevib did not complete for 2000 and 4000 connections (sqring size too small)

• xe uses a submission ring size of 256 (faster than 512)

• l4cpp uses a submission ring size of 512 (which is slower when connections > 1000)