https://lists.sr.ht/~azz/constellation-announce/%3CwiH9sNBxB03ch-HkDoHRVyQp0gWT7iyWf1dIdP9n-xxfEEcC_yb4p9uQW_NWjZgqvE-SD7lvGfPvtQ8bFgMOXEaNivnEEjdYqPKP6WMbnKM=@aciedo.com%3E
quoting note17hp…zhejJust “finished” first implementation of a simple ish memory allocator for Starlight (Constellation’s server implementation). It’s designed to allocate packets directly as they come off the network card. Context: Constellation is a Bitcoin L3 I’m working on.
Fun fact: these packets are only copied in memory once in the entire transaction pipeline (raw pointers are involved 🥸).
29ns/packet means Slabocator9000 could allocate 400gbps of packets per CPU core. That’s only running allocation, so obv full throughput won’t hit that per core but it’s pretty damn fast.
It also reuses all memory and therefore will only use as much memory as it possibly needs.