Hydra : Resilient and Highly Available Remote Memory
summary
Hydra is a remote-memory system that encodes 4 KB pages on-the-fly and serves reads/writes with single-digit μs latency by asynchronous encode/write, k+r optimistic reads, and zero-copy paths. It implements CodingSets placement to spread k data + r parity splits across disjoint failure domains, reducing correlated-data-loss by an order of magnitude. On a 50-node testbed Hydra matches replicated performance while cutting overhead to 1.25× and outperforms prior remote-memory solutions by up to 4.35×.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:54Remote Memory is Vulnerable
- 2:11Remote Memory Performs Great!
- 3:21Performance Tradeoff for Resiliency
- 6:48Remote Memory w/ Hydra
- 8:01Data Plane
- 8:32Hydra Datapath - Asynchronously Encoded Write
- 9:33Hydra Datapath – Late-binding Read
- 12:53Replication-like Performance at Lower Overhead
- 13:50Performance Tradeoff w/ Hydra
related talks
16mHydra: Serialization-Free Network Ordering for Strongly Consistent Distributed Applications
27mHydra: a federated resource manager for data-center scale analytics
18mHermit: Low-Latency, High-Throughput, and Transparent Remote Memory...
20mUniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
15mOne-sided RDMA-Conscious Extendible Hashing for Disaggregated Memory