Parallel Input Fetching in Bitcoin Core
summary
Claim: merged PR 35295 adds a lockfree parallel fetcher in the coins view overlay to read prevouts into an MPSC list before connectblock, avoiding per-input round trips to the DB. Implementation uses an atomic multi‑producer single‑consumer queue and 8 worker threads by default (configurable to 16). No protocol change. Benchmarks show ~3x faster on a laptop, 46% faster on Raspberry Pi 5, and about 5x faster on AWS, cutting the long tail of the last ~100k blocks during IBD.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:30PR 35295 impact
- 1:55UTXOs and UTXO set
- 3:44UTXO cache optimizations
- 8:58Block validation with ephemeral cache
- 12:25Block validation bottlenecks
- 13:29Parallel input fetching implementation
- 16:19Benchmark results
- 18:27IBD and spam mitigation
- 19:12Q&A
