hotring: a hotspot-aware in-memory key-value storelifeifei/papers/hotring-fast20... · 2020. 6....

1
HotRing: A Hotspot-Aware In-Memory Key-Value Store Database KVS Application A small portion of items that are frequently accessed 1.Hotspot issue 5.Design & Evaluation 3.Ideal & Challenge Memory accesses required to retrieve an item should be (negatively) correlated to this hotness Ensuring dynamic hotspot shift and lock-free access is a challenge #1: Ordered-ring structure #3: Lock-free rehash Safe termination of lookup processes #2: Identify hotspots and adjust head pointer Random Movement Statistical Sampling Head Pointer Format Item Format Presenter: Le Cai Authors: Jiqiang Chen, Liang Chen, Sheng Wang, Guoyun Zhu, Yuanyuan Sun, Huan Liu, Feifei Li #YCSB benchmarks Key size Value size Zipfian θ key-bucket ratio # of thread 8 bytes 8 bytes 1.22 8 64 Transition period Hash Table Initialization Deletion Split 2.Workloads analysis Daily: 1% data holds 50% accesses Extreme: 1% data holds 90% accesses Daily distribution Extreme distribution 4.HotRing: Ring-based hash index The head pointer can point to any items dynamically

Upload: others

Post on 12-May-2021

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: HotRing: A Hotspot-Aware In-Memory Key-Value Storelifeifei/papers/HotRing-fast20... · 2020. 6. 16. · HotRing: A Hotspot-Aware In-Memory Key-Value Store. Database. KVS. Application

HotRing: A Hotspot-Aware In-Memory Key-Value Store

Database

KVS

Application

A small portion of items that are frequently accessed

1.Hotspot issue 5.Design & Evaluation

3.Ideal & ChallengeMemory accesses required to retrieve an item

should be (negatively) correlated to this hotness

Ensuring dynamic hotspot shift and lock-free access is a challenge

#1: Ordered-ring structure

#3: Lock-free rehash

Safe termination of lookup processes

#2: Identify hotspots and adjust head pointer

Random Movement Statistical Sampling

Head Pointer Format Item Format

Presenter: Le Cai

Authors: Jiqiang Chen, Liang Chen, Sheng Wang, Guoyun Zhu, Yuanyuan Sun, Huan Liu, Feifei Li

#YCSB benchmarksKey size Value size Zipfian θ key-bucket ratio # of thread8 bytes 8 bytes 1.22 8 64

Transition period

② ③

Hash Table Initialization

DeletionSplit

2.Workloads analysis Daily: 1% data holds 50% accesses Extreme: 1% data holds 90% accesses

Daily distribution Extreme distribution

4.HotRing: Ring-based hash index The head pointer can point to any items

dynamically