Yeah that doesn’t look good. Another thing that profiling might surface is that network device drivers typically perform better on multicore when they can run rx and tx in parallel on separate cores. That OCaml uses a single CPU lock on the memory collector might be relevant to its comparison with other GC languages that do not. Not at all sure why the Haskell implementation is so slow, but without looking at it my first guess would be to look for whether it uses explicit strictness appropriately.