Cranking Up Frankendancer
summary
Firedancer’s leader pipeline uses pinned CPU-core “tiles” (net, quick, verify, dup, pack, bank, poh, shred) and AVX-accelerated Ed25519 to double signature throughput; the pack tile and net tile become the main limits. With consensus limits raised, the setup processes 1.04M TPS of simple system transfers, pushes roughly ≈42 Gbps aggregate network traffic, and hits a pack-tile ceiling near 1.08M TPS while followers keep up by retransmitting shreds.
AI-written summary · about
chapters (9)
- 0:00Introduction to Firedancer
- 0:51Architecture and data flow
- 6:29Core performance principles
- 7:41Demo setup and methodology
- 12:40Cluster configuration
- 17:37First performance benchmark
- 21:04Increasing consensus limits
- 23:24Final benchmark results
- 26:34Conclusion and summary
related talks
33mReaching Ludicrous Speed
01h 00mExploring Traffic Jams in Your Data Flows
35mPushing the Limits of Prometheus at Etsy
13mLow-level wizardry in dav1d
26mAre You Sure About Your Mesh Performance?
43mRealtime performance at scale (...)
10mTezos on the Fast Track: 1M TPS Demo & Visualization
32mSolid Ground