Accelerating Skewed Workloads With Performance Multipliers in the TurboDB Distributed...
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Build replaces one CockroachDB server with a Cicada turbo node and promotes the 40 million hottest keys to it, implementing Hybrid Concurrency Control to assign client timestamps and avoid cross‑node coordination and Phalanx Replication to replicate the turbo without blocking execution. Implemented on CockroachDB+Cicada, the system shows up to an order‑of‑magnitude higher throughput and lower latency on highly skewed YCSB/TPCC workloads, turning a Cockroach bottleneck with ~1 in 10 transactions aborting into sustained throughput until Cicada’s single‑node cap.
AI-written summary · about
chapters (9)
- 0:00TurboDB hybrid architecture
- 1:20Skewed workloads in distributed DBs
- 2:25Single machine outperforms under skew
- 3:51Bridging single and distributed DBs
- 5:17Hybrid concurrency control (HCC)
- 7:24Atomicity in multi-shot transactions
- 10:14FenX replication without latency
- 13:02Performance evaluation (YCSB & TPC-C)
- 15:34Summary and future
related talks
22mBuilding CockroachDB for performance at scale
48mBuilding mission-critical applications with CockroachDB
31mFuture-proofing operational data at scale
14mBuilding CockroachDB for performance at scale
18mFrom Today to 10X: Why Groww chose
28mGlobal Payments: Subsecond Performance Worldwide
16mScaleDB: A Scalable, Asynchronous In-Memory Database
29mHow City Storage Systems designed its Data Infrastructure to minimize toil