Hydra - Resilient and Highly Available Remote Memory

Explore a 15-minute conference talk from FAST '22 introducing Hydra, a groundbreaking resilient and highly available remote memory system. Discover how this low-latency, low-overhead mechanism accesses erasure-coded remote memory within microseconds, significantly improving performance-efficiency tradeoffs. Learn about CodingSets, a novel coding group placement algorithm that enhances load balancing and reduces data loss probability under correlated failures. Examine Hydra's impressive performance, achieving results close to fully in-memory scenarios even with only 50% local memory, and outperforming existing remote-memory solutions by up to 4.35 times. Delve into topics such as remote memory vulnerability, performance tradeoffs for resiliency, Hydra's data plane, asynchronous encoded writes, and late-binding reads.

Intro
Remote Memory is Vulnerable
Remote Memory Performs Great!
Performance Tradeoff for Resiliency
Remote Memory w/ Hydra
Data Plane
Hydra Datapath - Asynchronously Encoded Write
Hydra Datapath – Late-binding Read
Replication-like Performance at Lower Overhead
Performance Tradeoff w/ Hydra