GraphX: Unifying Data- Parallel and Graph-Parallel Analytics Reynold Xin, Joseph Gonzalez AMPLab, UC Berkeley Spark User Meetup July 2, 2013 

Graphs are Essential to Data Mining and Machine Learning Identify influential people and information Find communities Understand people’s shared interests Model complex data dependencies 

Pregel Specialized Graph Systems 

B C D E F A Specialized Graph Systems 1.  APIs to capture complex graph dependencies

Specialized Graph Systems 1.5 423 0 100 200 300 400 500 GraphLab Hadoop Runtime (in minutes, counting 34.8 billion triangles) 

Limitations of Specialized Systems No support for ETL & Post Processing Not interactive Requires separate runtime to maintain both in engineering and operations Data-parallel systems excel at these! 

No content

Vertices 

Edges Edges 

username password location rxin ****** Berkeley jegonzal iJustRaised$6.7m Pittsburgh franklin ****** Shanghai istoica ****** Piedmont user1 user2 relationship rxin jegonzal friend franklin rxin advisor istoica franklin coworker 

How can we combine the advances in data- parallel analytics and graph-parallel analytics? 

How can we combine the advances in data- parallel analytics and graph-parallel analytics? 

Why the Unification? No disparate systems to maintain / learn Leverage data-parallel engines for task scheduling, distribution and dispatch Reuse existing engines for ETL and consumption of graph computation output 

Enable Joining Tables and Graphs Simplify ETL and graph structured analytics User Data Product Ratings Friend

The Research Challenge Expressive graph computation primitives implementable on existing data-parallel engines » Relational databases (with UDFs and UDAFs) » MapReduce-like frameworks, e.g. Hadoop, Spark Leveraging advanced properties and engine extensions to make these primitives fast » An optimizer for choosing execution strategies » Controlled data partitioning » New index-based access methods and operators 

Graph Primitives A graph consists » A table of vertices (vid Int, attr1 Int, attr2 …) » A table of edges (srcId Int, dstId Int, attr1 Int, attr2 String …) Relational operators: » E.g. filter vertices, edges, joining vertices with tables Graph computation primitives: »  aggregateNeighbors(mapUdf, reduceUdaf) 

aggregateNeighbors B C D E F A map reduce 

aggregateNeighbors B C D E F A map(F) map(D) map(C) map(B) map(E) 

aggregateNeighbors B C D E F A map(F) map(D) map(C) map(B) map(E) 

aggregateNeighbors B C D E F A map(F) map(D) map(C) map(B) map(E) reduce 

Example: Vertex Degree B C D E F A map: 1 reduce: sum 

Example: Vertex Degree B C D E F A 1 1 1 1 1 

Example: Vertex Degree B C D E F A sum: 5 A: 5 B: 1 C: 1 D: 2 E: 3 F: 2 

Logical Query Plan aggregateNeighbors(mapUdf, reduceUdaf) »  CREATE VIEW aggResults AS SELECT reduceUdaf(*) FROM (SELECT mapUdf(v.attr1, v.attr2, …, e.attr1, …) FROM vertices v RIGHT OUTER JOIN edges e on v.id=e.srcId) GROUP BY e.dstId 

GraphX operators Relational operators on edges and vertices: » Filter, join, projection… Graph computation operator: » aggregateNeighbors 

We can express both Pregel and GraphLab using aggregateNeighbors in 20 lines of code! 

Performance Optimizations Replicate & co-partition vertices with edges » GraphLab (PowerGraph) style vertex-cut partitioning » Minimize communication by avoiding edge data movement in JOINs In-memory hash index for vertices for fast joins Optimizer for choosing execution strategies » E.g. if mapUdf does not need edge data, we can rewrite the query to delay the join 

Current Implementation Pregel (20) PageRank (5) GraphX Spark (relational operators) Connected Comp. (10) Shortest Path (10) ALS (40) GraphLab (20) 

Demo 

vertices = spark.textFile("hdfs://path/pages.csv") edges = spark.textFile("hdfs://path/to/links.csv”) .map(line => new Edge(line.split(‘\t’)) g = new Graph(vertices, edges).cache println(g.vertices.count) println(g.edges.count) g1 = g.filterVertices(_.split('\t')(2) == "Berkeley") ranks = Analytics.pageRank(g1, numIter = 10) println(ranks.vertices.sum) 30   

ranks = Analytics.pageRank(g1, numIter = 10) println(ranks.vertices.sum) 31   

Early Performance 22 165 1340 0 200 400 600 800 1000 1200 1400 1600 GraphLab GraphX Hadoop Runtime (in seconds, PageRank for 10 iterations) 

GraphX 1.  Graph-parallel primitives implementable in data-parallel (or relational) engines. 2.  Currently slower than GraphLab, but » No need for specialized systems » Easier ETL, and easier consumption of output » Interactive graph data mining 3.  Future work will bring performance closer to specialized engines. 

Berkeley Data Analytics Stack Spark Shark BlinkDB SQL HDFS / Hadoop Storage Mesos / YARN Resource Manager Spark Streaming GraphX MLBase 

Backup slides 

Vertex Cut Partitioning B C D E F A 

Vertex Cut Partitioning B C D E F A Partition 1 Partition 2 Partition 3