PlanetServe: A Decentralized, Scalable, and Privacy-Preserving Overlay for...
summary
PlanetServe builds a peer-to-peer overlay that uses Theta secure information dispersal (Rabin IDA + Shamir shares) for anonymous prompts, an HR tree hash registry for decentralized KV-cache discovery and load-aware forwarding, and sampling-based verification with a small BFT committee to detect dishonest model nodes. A prototype on thousands of nodes and A100 GPUs shows under-millisecond routing overhead and achieves over 50% latency reduction (avg and TTF token) versus a non-forwarding baseline while preserving privacy and flagging misbehaving nodes.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:04Decentralized LM serving challenges
- 2:08Roles in PlanetServe
- 3:12Theta algorithm
- 4:54Anonymous routing with onion
- 8:06HR tree overlay forwarding
- 10:59Model verification via sampling
- 12:26Anonymity and overhead evaluation
- 13:30Reputation score evolution
- 13:50Latency improvements with HR tree
- 14:52Summary and future
related talks
23mOblivP2P: An Oblivious Peer-to-Peer Content Sharing System
42mWhy Privacy Matters
27mThe Four Pillars of Large-Scale Decentralized Pre-Training: Methods, Bottlenecks...
25mInternet of Peers
01h 01mToward Provably Private Federated Learning
32mGetting started with p2p
27mDecentralized Storage with IPFS: How does it work under the hood?
17mDistributed Intelligence: The Alignment Protocol