Unification of Temporary Storage in the NodeKernel Architecture
Talkat USENIX ATC '19 · on Thursday, 2019-07-11· at 17:35· USENIX ATC Track II: Grand Ballroom VII–IX·
summary
NodeKernel designs a unified temporary-storage layer and ships it as Crail, a metadata‑driven storage with hierarchical namespaces, storage classes (DRAM and NVMe‑Flash) and a bag abstraction that maps shuffle partitions to files. It uses RDMA/NVMe‑over‑Fabrics to cut metadata round trips to microseconds, hits ~10 µs DRAM get latency and ~200 Gbps read throughput, sustains ~8M IOPS per metadata shard, and reports up to 4.8× NoSQL and 3.4× Spark speedups while cutting storage cost up to 8×.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:19Temporary Intermediate Data
- 2:07Shortcomings of Temporary Data Storage
- 5:39Temporary data distribution
- 7:58Temporary Data Storage Requirements
- 11:38Nodekernel: Node Types
- 13:08Crail: System Architecture
- 17:49Evaluation
- 18:48Small and Large Data Sets
- 23:02Flexible Deployment
- 23:40DRAM / Flash Ratio
- 24:44Conclusions
- 25:36Open Source
- 26:43Crail Data Plane
related talks
35mSystems @Scale Summer Q&A Week 3
19mBeyond the Node: Scaling Inference with Cluster-Wide KVCache...
23mSize-aware Sharding For Improving Tail Latencies in In-memory Key-value Stores
18mEthane: An Asymmetric File System for Disaggregated Persistent Memory
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM
17mSingularFS: A Billion-Scale Distributed File System Using a Single Metadata Server
50mImmutable Values in the Big Wide World
15mReplicating Persistent Memory Key-Value Stores with Efficient RDMA Abstraction