Asynchronous and Fault-Tolerant Recursive Datalog Evalua9on in - - PowerPoint PPT Presentation

Asynchronous and Fault-Tolerant Recursive Datalog Evalua9on in Shared-Nothing Engines

Jingjing Wang, Magdalena Balazinska, Daniel Halperin University of Washington

Modern Analy>cs Requires Itera>on

• Graph applica>ons

– Graph reachability – Connected components – Shortest Path

• Machine learning

– Clustering algorithms – Logis>c regression

• Scien>fic analy>cs

– N-body simula>on

• …

Jingjing Wang - University of Washington 2

Galaxy Evolu>on: An Itera>ve Example

Jingjing Wang - University of Washington 3

A Simula9on of the Universe Present day Big Bang Millions of years ago … …

Picture from D. H. Stalder et. al. arXiv:1208.3444 [astro-ph.CO]

Galaxy Galaxy

Galaxy Evolu>on: Itera>ve Lineage Tracing

Present day

Jingjing Wang - University of Washington

Millions of years ago

…

Galaxy Par9cle

… …

Millions of years ago

Galaxy Evolu>on: Why It Is not Easy

• Large-scale data sizes

– Scalability

• Itera>ve is the core

– Support efficient itera>ve constructs

• Users are data scien>sts

– Provide an easy-to-use query interface

• Shared datasets and resources

– Within a data management system

Jingjing Wang - University of Washington 5

Itera>ve Analy>cs: Where to Do

• SQL Server

– Single-node, cannot handle huge scale

• MapReduce

– Rigid programming model – Write to disk, expensive itera>on

• In-memory systems such as Spark

– Synchronous opera>ons

• Graph engines such as GraphLab

– Think like a vertex

Jingjing Wang - University of Washington 6

No Exis>ng System Meets All Requirements

• Synchronous itera>ons only

– AsterixDB, HaLoop, Pregel, REX, Spark, PrIter, Glog, …

• Single-node

– LogicBlox, DatalogFS, …

• No declara>ve language

– Stratosphere, Naiad, Grace, GraphLab, …

• Specialized for graphs

– GraphLab, Grace, …

• Not a data management system

– SociaLite, …

• Theory on recursive queries

– DatalogFS, …

Jingjing Wang - University of Washington 7

Outline and Contribu>ons

• Full-stack solu>on for itera>ve processing

– Declara>ve rela>onal query language

• A subset of Datalog-with-Aggrega>on

– Scalable and easily implementable

• Small extensions to exis>ng shared-nothing systems

– Efficient itera>ve computa>on

• Execu>on models and op>miza>ons
• Implementa>on and empirical evalua>on using

Jingjing Wang - University of Washington 8

Outline and Contribu>ons

• Full-stack solu>on for itera>ve processing

– Declara9ve rela9onal query language

• A subset of Datalog-with-Aggrega9on

– Scalable and easily implementable

• Small extensions to exis9ng shared-nothing systems

– Efficient itera>ve computa>on

• Execu>on models and op>miza>ons
• Implementa>on and empirical evalua>on using

Jingjing Wang - University of Washington 9

From Datalog Programs to Asynchronous Query Plans

• Datalog: a rela>onal query language

– Nicely expresses recursions

• Two special operators

– IDBController

• Maintains state of “nonconstant” rela>ons

– Termina>onController – Easy extensions to an exis>ng engine

• Automa>c compila>on

Jingjing Wang - University of Washington 10

DECLARE @id AS INT, @lvl AS INT SET @id = 3 SET @lvl = 2 ;WITH cte (id, parent, child, lvl) AS ( SELECT id, parent, child, 0 FROM t WHERE id = 1 UNION ALL SELECT E.id, E.parent, E.child, M.lvl+1 FROM t AS E JOIN CTE AS M ON E.parent = M.child WHERE lvl < @lvl ) SELECT * FROM CTE --where lvl=@lvl

• -OPTION (MAXRECURSION 10)

CC(x,x) :- Edges(x, ) CC(y,$Min(v)) :- CC(x,v), Edges(x,y) :- CC(y,v)

Outline and Contribu>ons

• Full-stack solu>on for itera>ve processing

– Declara>ve rela>onal query language

• A subset of Datalog-with-Aggrega>on

– Scalable and easily implementable

• Small extensions to exis>ng shared-nothing systems

– Efficient itera9ve computa9on

• Execu9on models and op9miza9ons
• Implementa9on and empirical evalua9on using

Jingjing Wang - University of Washington 11

Itera>ve Computa>on: How Can We Do Beqer

Jingjing Wang - University of Washington 12

• Performance impact: # of intermediate tuples

– More tuples, more work, more resources

• Op>miza>on: recursive execu>on models

– Synchronous vs. asynchronous

• Op>miza>on: priori>zing tuples

– For asynchronous model, favor new tuples vs. base tuples

Op>miza>on: Recursive Execu>on Models

Jingjing Wang - University of Washington 13

• Synchronous

– Stop at the end of each itera>on

• Asynchronous

– No barrier, propagate updates when ready

• Galaxy Evolu>on

– Synchronous

• Find all galaxies at >mestep 1, then 2, …

– Asynchronous

• Galaxy A is a part of the evolu>on history
• A shares par>cles with galaxy B

Galaxy Evolu>on: Execu>on Model Does Not Maqer Much

Jingjing Wang - University of Washington 14

100 200 300 400 500 600 8 16 32 64

Time (seconds) # workers

80GB, 27 snapshots 16 machines

Another Applica>on: Least Common Ancestor

Jingjing Wang - University of Washington 15

2 1 4 3 5

dist:1 dist:2 dist:3 Cita>on Paper

LCA: Asynchronous Can Be Much Slower Than Synchronous

Jingjing Wang - University of Washington 16

20 40 60 80 100 120 140 160 8 16 32 64

Time (seconds) # workers

2 million papers 8 million cita>ons

• For asynchronous processing

– Choice: favor new tuples vs. base tuples

• Example: connected components

Op>miza>on: Priori>zing Tuples

Jingjing Wang - University of Washington 17

1 2 3 4 1 2 3 4

Connected Components: Pull Order Impacts Run Time

Jingjing Wang - University of Washington 18

500 1000 1500 2000 8 16 32 64

Time (seconds) # workers

21 million ver>ces 776 million edges

Async, new tuples first Async, base tuples first Sync

Conclusion

• Full-stack solu>on for itera>ve big-data analy>cs

– A declara>ve language – Small extensions to exis>ng shared-nothing engines – Efficient itera>ve execu>on – Failure handling methods – More details in the paper

• Empirical evalua>on of various models

– No single method outperforms others – Future work: an adap>ve cost-based op>mizer

Jingjing Wang - University of Washington 19