Pregel: A System for Large-Scale Graph

ProcessingPresented by Dylan Davis

Authors: Grzegorz Malewicz, Matthew H. Austern, Aart J.C. Bik, James C. Dehnert, Ilan Horn, Naty Leiser, Grzegorz Czajkowski

(GOOGLE, INC.)

Overview•What is a graph?•Graph Problems• The Purpose of Pregel•Model of Computation•C++ API• Implementation•Applications• Experiments

What is a graph?G = (V, E)

Binary Tree

Graph Problems

Network Routing Social Network Connections

The Purpose of Pregel•Google was interested in applications that could perform internet-related graph algorithms, such as PageRank, so they designed Pregel to perform these tasks efficiently.• It is a scalable, general-purpose system for implementing graph algorithms in a distributed environment.•Focus on “Thinking Like a Vertex” and parallelism

Model of Computation

Model of Computation (Vertex)

Vertex ID

Vertex Value

Edge ValueVertex

ID

Vertex ID

Edge Value

Model of Computation (Superstep)Superstep 0 Superstep 1 Superstep 2

Execution Time

Compute()

Compute()

Compute() Compute()

Compute()

Compute() Compute()

Compute()

Compute()

Model of Computation (Vertex Actions)

A vertex can:

Vertex ID

Vertex Value

• Modify its values• Receive messages from

previous superstep• Send messages• Request topology changes

Model of Computation (State Machine)

C++ API

C++ API (Message Passing)Destination

Vertex IDMessage

Value

2 571 2

Message Buffer

C++ API (Combiners & Aggregators)

Combiner Aggregator

C++ API (Topology Mutations)V

Superstep

C++ API (Input and Output)0 1 2 3 4

0 0 0 1 1 01 0 0 0 1 12 1 1 0 1 13 0 1 1 0 14 1 1 1 0 0

Implementation

Implementation (Basic Architecture)

Implementation (Program Execution)

Flow:1. Copy user program – Master copy & worker copies2. Master assigns graph partitions3. Master takes user input data, assigns to workers –

load vertex data4. Supersteps (Compute() and send messages)5. Save output

Implementation (Fault Tolerance)Checkpoint

WorkerSave()

WorkerSave()

WorkerSave()

Recover

WorkerRecompute()

WorkerWorker

Recompute()X

Implementation (Worker)

Worker Worker

Implementation (Master)List of

WorkersMaster

Partitions

Applications

Applications (Shortest Path)2 1

5

3

Experiments

Experiments (Description)• Test the execution times of Pregel running the Single-

Source Shortest Path algorithm. •Use a cluster of 300 multicore commodity PCs.•Run Pregel with Binary Tree graphs, and with a more

realistic, randomly-distributed graph. •Results do not include initialization, graph generation,

and result verification times.• Failure Recovery is not included (reduces overhead)

Conclusion•Pregel is a model suitable for large-scale graph computing

with a production-quality, scalable and fault tolerant implementation.

•Programs are expressed as a sequence of iterations, in each of which a vertex can receive messages sent in the previous iteration, send messages to other vertices, and modify its own state and that of its outgoing edges.

•This implementation is flexible enough to express a broad set of algorithms.