Anti-Entropy Bandits for Geo-Replicated Consistency

In ICDCS, July 2018.

Benjamin Bengfort and Konstantinos Xirogiannopoulos and Pete Keleher



Abstract:
We address the challenge of large, geographically distributed eventually consistent systems by improving synchronization using reinforcement learning techniques. Anti-entropy uses gossip and rumor spreading to propagate updates deterministically without saturating the network even in the face of network outages. These protocols use uniform random selection to choose synchronization peers, which means that a write occurring at one replica is not efficiently propagated across the network. In this paper we explore the use of multi-armed bandit algorithms to optimize for fast, successful synchronizations by modifying peer selection probabilities. The result is a synchronization topology that emerges according to access patterns and network latency which produces efficient synchronization, localizes most data exchange, lowers visibility latency and increases consistency.
@inProceedings{icdcs018,
	title = "Anti-Entropy Bandits for Geo-Replicated Consistency",
	author = "Benjamin Bengfort and Konstantinos Xirogiannopoulos and Pete Keleher",
	booktitle = {ICDCS},
	month = {July},
	year = {2018},
}


Available: bibtex, abstract,
Edit