Lock-free Linearizable 1-Dimensional Range Queriessalaa/wttm2016/html/... · Bapi Chatterjee...
20
Lock-free Linearizable 1-Dimensional Range Queries Bapi Chatterjee Distributed Computing and Systems Group Department of Computer Science and Engineering Chalmers University of Technology, Sweden 8th WTTM 2016 July 25, 2016 1/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Transcript of Lock-free Linearizable 1-Dimensional Range Queriessalaa/wttm2016/html/... · Bapi Chatterjee...
Lock-free Linearizable 1-Dimensional RangeQueries
Bapi Chatterjee
Distributed Computing and Systems GroupDepartment of Computer Science and Engineering
Chalmers University of Technology, Sweden
8th WTTM 2016July 25, 2016
1/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Introduction
2/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Introduction
The problem statement
• Given: Set of data-points on the real line and an intervalthereon.
• Setup: Dynamic and Concurrent addition and removal ofdata-points.
• Output: The data-points in the interval.
• Requirement:• Lock-freedom.
• Linearizability.
3/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Introduction
Problem relevance
• Ubiquitous multi-core systems.
• The ordered Set ADT provides size(T a, T b) that isimplemented in sequential setup.
• Increasing availability of efficient lock-free ordered sets.
• Linearizability provides an aligned view with respect to areal-time line.
• Growing popularity of In-memory Databases. Some withlock-free index structures e.g. MemSQL uses lock-freeskip-lists.
4/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Introduction
Existing linearizable concurrent range search implementations
1 Brown and Avni [OPODIS ’12]: Lock-free k-ary search treebased; Lock-free; Double-scan and validation.
2 Avni et al. [PODC ’13]: STM based approach on a variant ofSkip-list with “fat” nodes; Blocking.
3 Sagonas and Winblad [LCPC Workshop ’15]: Locking treestructure based; Blocking.
5/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Presented Method
6/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Motivation for an alternate approach
• Progress guarantee.
• Independence from the performance of a particular STM.
• Avoiding multi-scan and restart.
7/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Overview
Afek et al. ’93: Snapshot
↓
Attiya et al. ’08, Imbs etal. ’09: Partial Snapshot
↓
↓
Brown and Avni ’12: kstbased Range Search
Jayanti ’05: Snapshot
↓
↓
Petrank et al. ’13:Lock-free Iterators
↓
Our method
8/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Snap-collectorAn Ordered Set Data Structure
∗sc
isActive: bool
A Snap-Collector Object
L: node-list
R1: report-list of p1 Rn: report-list of pn
Remove-report
Add-report
9/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Range-collectorAn Ordered Set Data Structure
∗rc
isActive: bool
A Range-Collector Object
L: node-list
R1: report-list of p1 Rn: report-list of pn
Remove-report
Add-report
x : lower limit
x ′: upper limit
10/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Lock-free data structure with Range-collector-listAn Ordered Set Data Structure
head
tail
11/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Lock-free data structure with Range-collector-listAn Ordered Set Data Structure
head
tail
12/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Algorithm
Lock-free data structure with Range-collector-list
An Active set
An Ordered Set Data Structure
head
tail
13/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Performance
14/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Performance
Performance
2 4 8 16 28 32
2,000
3,000
4,000
5,000
6,000
7,000
#threads
Ops
/ms
LLBSTSL
KST 64
Figure: 5% Add, 5% Rem, 89%Con, 1% RSearch, 0≤x<102
Machine: Intel(R) Xeon(R) E5-2695 v3 processor, dual chip with 14 hardware processes per chip (28 threads with HT) @ 2.30 GHz. Average of 10 repetitions of 5 seconds runs. Integer type keys.
15/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Performance
Performance
2 4 8 16 28 320
2,000
4,000
6,000
#threads
Ops
/ms
LLBSTSL
KST 64
Figure: 25% Add, 25% Rem, 49%Con, 1% RSearch, 0≤x<102
Machine: Intel(R) Xeon(R) E5-2695 v3 processor, dual chip with 14 hardware processes per chip (28 threads with HT) @ 2.30 GHz. Average of 10 repetitions of 5 seconds runs. Integer type keys.
16/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Performance
Performance
2 4 8 16 28 32
1,000
2,000
3,000
#threads
Ops
/ms
LLBSTSL
KST 64
Figure: 5% Add, 5% Rem, 89%Con, 1% RSearch, 0≤x<103
Machine: Intel(R) Xeon(R) E5-2695 v3 processor, dual chip with 14 hardware processes per chip (28 threads with HT) @ 2.30 GHz. Average of 10 repetitions of 5 seconds runs. Integer type keys.
17/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Performance
Performance
2 4 8 16 28 32
400
600
800
1,000
#threads
Ops
/ms
LLBSTSL
KST 64
Figure: 25% Add, 25% Rem, 49%Con, 1% RSearch, 0≤x<103
Machine: Intel(R) Xeon(R) E5-2695 v3 processor, dual chip with 14 hardware processes per chip (28 threads with HT) @ 2.30 GHz. Average of 10 repetitions of 5 seconds runs. Integer type keys.
18/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Summary
19/20 Lock-free Linearizable Range Queries Bapi Chatterjee
Summary
Summary
Summary
• A generic linearizable lock-free 1-dimensional range search.
• No multiple-scan, no invalidation.
• Scalability does not deteriorate with high number of updates.
20/20 Lock-free Linearizable Range Queries Bapi Chatterjee
![Bapi jco[1]](https://static.fdocuments.us/doc/165x107/55587609d8b42aaa7e8b5447/bapi-jco1.jpg)
