SlideShare una empresa de Scribd logo
1 de 29
Descargar para leer sin conexión
Pregel: A System
For Large Scale
Graph Processing
Grzegorz Malewicz, Matthew H. Austern, Aart
J. C. Bik, James C. Dehnert, Ilan Horn,
Naty Leiser, and Grzegorz Czajkowski
Presented By
Riyad Parvez
Real World Graph Processing
● Web graph:
○ PageRank (influential vertices)
● Social graph:
○ Popularity rank, personalized rank, shortest paths, shared
connections, clustering (communities), propagation
● Advertisement:
○ Target ads
● Communication network:
○ Maximum flow, transportation routes
● Biology network
○ protein interactions
● Pathology network
○ find anomalies
Graph Processing is Different
● Poor locality of memory access.
● Very little work done per vertex.
● Changes degree of parallelism over the
course of execution.
Why not MapReduce (MR)
● Graph algorithms can be expressed as
series of MR jobs.
● Data must be reloaded and reprocessed at
each iteration, wasting I/O, network.
bandwidth, and processor resources.
● Needs an extra MR job for each iteration just
to detect termination condition.
● MR isn’t very good at dynamic dependency
graph.
Pregel
● Developed at Google
● Modeled after Bulk Synchronous Parallel (BSP)
computing model
● Distributed message passing system
● Computes in vertex-centric fashion
○ “Think like a vertex”
● Scalable and fault tolerant
● Influenced systems like Apache Giraph and
other BSP distributed systems
Bulk Synchronous Parallel
● Leslie Valiant introduced in 1990.
● Computations are consist of a sequence of iterations,
called superstep.
● During superstep, framework calls user-defined
computation function on every vertex.
● Computation function specifies behaviour at a single
vertex V and a single superstep S.
● Supersteps end with barrier synchronization.
● All communications are from superstep S to superstep
S+1.
● Performance of the model is predictable.
Bulk Synchronous Parallel
Source: https://cs.uwaterloo.ca/~kdaudjee/courses/cs848/slides/jenny.pdf
Bulk Synchronous Parallel
Source: https://cs.uwaterloo.ca/~kdaudjee/courses/cs848/slides/jenny.pdf
Pregel Computation Model
● Computation on locally stored data.
● Computations are in-memory.
● Terminates when all vertices are inactive or no
messages to be delivered.
● Vertices are distributed among workers using hash(ID)
mod N, where N is the number of partitions (default
partitioning)
● Barrier synchronization
➢ Wait and synchronize before the end of superstep
➢ Fast processors can be delayed by slow ones
● Persistent data is stored on a distributed storage system
(GFS/BigTable)
● Temporary data is stored in disk.
C++ API
Source: http://kowshik.github.io/JPregel/pregel_paper.pdf
Vertex
● Can mutate local value and value on outgoing edges.
● Can send arbitrary number of messages to any other
vertices.
● Receive messages from previous superstep.
● Can mutate local graph topology.
● All active vertices participate in the computation in a
superstep.
Vertex State Machine
● Initially, every vertices
are active.
● A vertice can deactivate
itself by vote to halt.
● Deactivated vertices
don't participate in
computation.
● Vertices are reactivated
upon receiving
message.
Example
Messages
● Consists of a message value and destination
vertex.
● Typically sent along outgoing edges.
● Can be sent to any vertex whose identifier is
known.
● Are only available to receiver at the beginning of
superstep.
● Guaranteed to be delivered.
● Guaranteed not to be duplicated.
● Can be out of order.
Combiner
● Sending messages incurs overhead.
● System calls Combine() for several messages intended
for a vertex V into a single message containing the
combined message.
● No guarantees which messages will be combined or the
order of combination.
● Should be enabled for commutative and associative
messages.
● Not enabled by default.
Combiner
Source: https://wiki.engr.illinois.edu/download/attachments/188588798/pregel.pdf?version=1
Aggregator
● Mechanism for global communication,
monitoring and global state.
○ Vertices provide value to aggregator in superstep S.
○ Values are combined using a reduction operator.
○ Resulting value is available to all vertices at
superstep S+1.
● New aggregator is defined by subclassing
"Aggregator" class.
● Reduction operator should be associative and
commutative.
Reduction (Aggregator)
Source: https://wiki.engr.illinois.edu/download/attachments/188588798/pregel.pdf?version=1
Topology Mutation
● Vertices can dynamically create/destroy vertices, edges.
● Mutations and conflict resolution take place at barrier.
● Except local mutation (self-edge) immediately takes
place.
● Order of mutations
○ Edge deletion
○ Vertex deletion
○ Vertex addition
○ Edge addition
Master
● Partitions the input and assigns one or more partitions
to each worker.
● Keeps list of
○ All alive workers
○ Worker's unique identifiers
○ Addressing informations
○ Partition of the graph is assigned to the worker.
● Coordinates barrier synchronization i.e., superstep.
● Fault tolerance by checkpoint, failure detection and
reassignment.
● Maintains statistics of the progress of computation and
the state of the graph.
● Doesn’t participate in computation.
● Not responsible for load-balancing.
Worker
● Responsible for computation of assigned
vertices.
● Keeps two copies of active vertices and
incoming messages
○ Current superstep
○ Next superstep
● Place local messages immediately in
message queue.
● Buffer remote messages.
○ Flush asynchronously in single message if threshold
reached.
Fault Tolerance
● Checkpoint at the beginning of superstep.
○ Master saves aggregators.
○ Workers save vertices, edges and incoming messages.
● Worker failure detected by ping messages.
● Recovery
○ Master reassigns failed worker partition to other available
workers.
○ All workers restart from superstep S by loading state from the
most recently available checkpoint.
● Confined recovery: recovery is only confined to lost
partitions
○ Workers also save outgoing messages.
○ Recomputes using logged messages from healthy partitions
and recalculated ones from recovering partitions.
PageRank
Source: http://kowshik.github.io/JPregel/pregel_paper.pdf
Performance
Experimental Setup:
● Hardware: A cluster of 300 multi-core commodity PCs.
● Algorithm: SSSP with unit weight edges.
• All-pairs shortest paths impractical b/c O(|V|2) storage.
● Measures scalability w.r.t. both the number of workers
and the number of vertices.
● Data collected for:
• Binary trees (to test scalability).
• log-normal random graphs (to study performance in a realistic
setting).
● No checkpointing.
Performance
Source: http://kowshik.github.io/JPregel/pregel_paper.pdf
Performance
Source: http://kowshik.github.io/JPregel/pregel_paper.pdf
Performance
Source: http://kowshik.github.io/JPregel/pregel_paper.pdf
Summary
● Distributed system for large scale graph
processing.
● Vertex-centric BSP model
○ Message passing API
○ A sequence of supersteps
○ Barrier synchronization
● Coarse grained parallelism
● Fault tolerance by checkpointing
● Runtime performance scales near linearly to the
size of the graph (CPU bound)
Discussion
● No fault tolerance for master is mentioned in
the paper (Probably Paxos or replication).
● Static partitioning! What happens if a worker
is too slow?
● Dynamic partitioning, network overhead for
reassigning vertices and state.
● Good for sparse graph. But communication
overhead for dense graph can bring the
system down to knees.

Más contenido relacionado

La actualidad más candente

Deep Dive into the New Features of Apache Spark 3.0
Deep Dive into the New Features of Apache Spark 3.0Deep Dive into the New Features of Apache Spark 3.0
Deep Dive into the New Features of Apache Spark 3.0Databricks
 
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...ScyllaDB
 
Migrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at FacebookMigrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at FacebookMariaDB plc
 
Databricks and Logging in Notebooks
Databricks and Logging in NotebooksDatabricks and Logging in Notebooks
Databricks and Logging in NotebooksKnoldus Inc.
 
Performance Schema for MySQL troubleshooting
Performance Schema for MySQL troubleshootingPerformance Schema for MySQL troubleshooting
Performance Schema for MySQL troubleshootingSveta Smirnova
 
Spark streaming
Spark streamingSpark streaming
Spark streamingWhiteklay
 
What's new in MariaDB TX 3.0
What's new in MariaDB TX 3.0What's new in MariaDB TX 3.0
What's new in MariaDB TX 3.0MariaDB plc
 
Spark 2.x Troubleshooting Guide
Spark 2.x Troubleshooting GuideSpark 2.x Troubleshooting Guide
Spark 2.x Troubleshooting GuideIBM
 
Introduction to PostgreSQL
Introduction to PostgreSQLIntroduction to PostgreSQL
Introduction to PostgreSQLJoel Brewer
 
ACM DBPL Keynote: The Graph Traversal Machine and Language
ACM DBPL Keynote: The Graph Traversal Machine and LanguageACM DBPL Keynote: The Graph Traversal Machine and Language
ACM DBPL Keynote: The Graph Traversal Machine and LanguageMarko Rodriguez
 
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013Introduction and Overview of Apache Kafka, TriHUG July 23, 2013
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013mumrah
 
Intro to MySQL Master Slave Replication
Intro to MySQL Master Slave ReplicationIntro to MySQL Master Slave Replication
Intro to MySQL Master Slave Replicationsatejsahu
 
Introduction to Spark Streaming
Introduction to Spark StreamingIntroduction to Spark Streaming
Introduction to Spark Streamingdatamantra
 
MariaDB AX: Analytics with MariaDB ColumnStore
MariaDB AX: Analytics with MariaDB ColumnStoreMariaDB AX: Analytics with MariaDB ColumnStore
MariaDB AX: Analytics with MariaDB ColumnStoreMariaDB plc
 
Introducing Apache Giraph for Large Scale Graph Processing
Introducing Apache Giraph for Large Scale Graph ProcessingIntroducing Apache Giraph for Large Scale Graph Processing
Introducing Apache Giraph for Large Scale Graph Processingsscdotopen
 
The Parquet Format and Performance Optimization Opportunities
The Parquet Format and Performance Optimization OpportunitiesThe Parquet Format and Performance Optimization Opportunities
The Parquet Format and Performance Optimization OpportunitiesDatabricks
 
MapReduce: Simplified Data Processing on Large Clusters
MapReduce: Simplified Data Processing on Large ClustersMapReduce: Simplified Data Processing on Large Clusters
MapReduce: Simplified Data Processing on Large ClustersAshraf Uddin
 

La actualidad más candente (20)

Deep Dive into the New Features of Apache Spark 3.0
Deep Dive into the New Features of Apache Spark 3.0Deep Dive into the New Features of Apache Spark 3.0
Deep Dive into the New Features of Apache Spark 3.0
 
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...
How We Reduced Performance Tuning Time by Orders of Magnitude with Database O...
 
Migrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at FacebookMigrating from InnoDB and HBase to MyRocks at Facebook
Migrating from InnoDB and HBase to MyRocks at Facebook
 
Databricks and Logging in Notebooks
Databricks and Logging in NotebooksDatabricks and Logging in Notebooks
Databricks and Logging in Notebooks
 
Performance Schema for MySQL troubleshooting
Performance Schema for MySQL troubleshootingPerformance Schema for MySQL troubleshooting
Performance Schema for MySQL troubleshooting
 
Spark streaming
Spark streamingSpark streaming
Spark streaming
 
What's new in MariaDB TX 3.0
What's new in MariaDB TX 3.0What's new in MariaDB TX 3.0
What's new in MariaDB TX 3.0
 
Spark 2.x Troubleshooting Guide
Spark 2.x Troubleshooting GuideSpark 2.x Troubleshooting Guide
Spark 2.x Troubleshooting Guide
 
Hadoop Map Reduce
Hadoop Map ReduceHadoop Map Reduce
Hadoop Map Reduce
 
Introduction to PostgreSQL
Introduction to PostgreSQLIntroduction to PostgreSQL
Introduction to PostgreSQL
 
PostgreSQL
PostgreSQLPostgreSQL
PostgreSQL
 
ACM DBPL Keynote: The Graph Traversal Machine and Language
ACM DBPL Keynote: The Graph Traversal Machine and LanguageACM DBPL Keynote: The Graph Traversal Machine and Language
ACM DBPL Keynote: The Graph Traversal Machine and Language
 
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013Introduction and Overview of Apache Kafka, TriHUG July 23, 2013
Introduction and Overview of Apache Kafka, TriHUG July 23, 2013
 
Intro to MySQL Master Slave Replication
Intro to MySQL Master Slave ReplicationIntro to MySQL Master Slave Replication
Intro to MySQL Master Slave Replication
 
Introduction to Spark Streaming
Introduction to Spark StreamingIntroduction to Spark Streaming
Introduction to Spark Streaming
 
MariaDB AX: Analytics with MariaDB ColumnStore
MariaDB AX: Analytics with MariaDB ColumnStoreMariaDB AX: Analytics with MariaDB ColumnStore
MariaDB AX: Analytics with MariaDB ColumnStore
 
Introducing Apache Giraph for Large Scale Graph Processing
Introducing Apache Giraph for Large Scale Graph ProcessingIntroducing Apache Giraph for Large Scale Graph Processing
Introducing Apache Giraph for Large Scale Graph Processing
 
The Parquet Format and Performance Optimization Opportunities
The Parquet Format and Performance Optimization OpportunitiesThe Parquet Format and Performance Optimization Opportunities
The Parquet Format and Performance Optimization Opportunities
 
Catalyst optimizer
Catalyst optimizerCatalyst optimizer
Catalyst optimizer
 
MapReduce: Simplified Data Processing on Large Clusters
MapReduce: Simplified Data Processing on Large ClustersMapReduce: Simplified Data Processing on Large Clusters
MapReduce: Simplified Data Processing on Large Clusters
 

Similar a Pregel: A System For Large Scale Graph Processing

Cassandra - A Decentralized Structured Storage System
Cassandra - A Decentralized Structured Storage SystemCassandra - A Decentralized Structured Storage System
Cassandra - A Decentralized Structured Storage SystemVarad Meru
 
Distributed implementation of a lstm on spark and tensorflow
Distributed implementation of a lstm on spark and tensorflowDistributed implementation of a lstm on spark and tensorflow
Distributed implementation of a lstm on spark and tensorflowEmanuel Di Nardo
 
Taskerman - a distributed cluster task manager
Taskerman - a distributed cluster task managerTaskerman - a distributed cluster task manager
Taskerman - a distributed cluster task managerRaghavendra Prabhu
 
Netflix machine learning
Netflix machine learningNetflix machine learning
Netflix machine learningAmer Ather
 
Advanced memory allocation
Advanced memory allocationAdvanced memory allocation
Advanced memory allocationJoris Bonnefoy
 
Graph processing
Graph processingGraph processing
Graph processingyeahjs
 
Ledingkart Meetup #2: Scaling Search @Lendingkart
Ledingkart Meetup #2: Scaling Search @LendingkartLedingkart Meetup #2: Scaling Search @Lendingkart
Ledingkart Meetup #2: Scaling Search @LendingkartMukesh Singh
 
Reactive by example - at Reversim Summit 2015
Reactive by example - at Reversim Summit 2015Reactive by example - at Reversim Summit 2015
Reactive by example - at Reversim Summit 2015Eran Harel
 
Apache spark - Spark's distributed programming model
Apache spark - Spark's distributed programming modelApache spark - Spark's distributed programming model
Apache spark - Spark's distributed programming modelMartin Zapletal
 
Software Design Practices for Large-Scale Automation
Software Design Practices for Large-Scale AutomationSoftware Design Practices for Large-Scale Automation
Software Design Practices for Large-Scale AutomationHao Xu
 
Computer Graphics - Lecture 01 - 3D Programming I
Computer Graphics - Lecture 01 - 3D Programming IComputer Graphics - Lecture 01 - 3D Programming I
Computer Graphics - Lecture 01 - 3D Programming I💻 Anton Gerdelan
 
Concurrency - Why it's hard ?
Concurrency - Why it's hard ?Concurrency - Why it's hard ?
Concurrency - Why it's hard ?Ramith Jayasinghe
 
Scalability broad strokes
Scalability   broad strokesScalability   broad strokes
Scalability broad strokesGagan Bajpai
 
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018Greenplum Overview for Postgres Hackers - Greenplum Summit 2018
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018VMware Tanzu
 
BlaBlaCar Elastic Search Feedback
BlaBlaCar Elastic Search FeedbackBlaBlaCar Elastic Search Feedback
BlaBlaCar Elastic Search Feedbacksinfomicien
 
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...Thien Q. Tran
 
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...PingCAP
 

Similar a Pregel: A System For Large Scale Graph Processing (20)

Apache Giraph
Apache GiraphApache Giraph
Apache Giraph
 
Cassandra - A Decentralized Structured Storage System
Cassandra - A Decentralized Structured Storage SystemCassandra - A Decentralized Structured Storage System
Cassandra - A Decentralized Structured Storage System
 
Distributed implementation of a lstm on spark and tensorflow
Distributed implementation of a lstm on spark and tensorflowDistributed implementation of a lstm on spark and tensorflow
Distributed implementation of a lstm on spark and tensorflow
 
Taskerman - a distributed cluster task manager
Taskerman - a distributed cluster task managerTaskerman - a distributed cluster task manager
Taskerman - a distributed cluster task manager
 
Netflix machine learning
Netflix machine learningNetflix machine learning
Netflix machine learning
 
Advanced memory allocation
Advanced memory allocationAdvanced memory allocation
Advanced memory allocation
 
Graph processing
Graph processingGraph processing
Graph processing
 
Ledingkart Meetup #2: Scaling Search @Lendingkart
Ledingkart Meetup #2: Scaling Search @LendingkartLedingkart Meetup #2: Scaling Search @Lendingkart
Ledingkart Meetup #2: Scaling Search @Lendingkart
 
Reactive by example - at Reversim Summit 2015
Reactive by example - at Reversim Summit 2015Reactive by example - at Reversim Summit 2015
Reactive by example - at Reversim Summit 2015
 
Apache spark - Spark's distributed programming model
Apache spark - Spark's distributed programming modelApache spark - Spark's distributed programming model
Apache spark - Spark's distributed programming model
 
Software Design Practices for Large-Scale Automation
Software Design Practices for Large-Scale AutomationSoftware Design Practices for Large-Scale Automation
Software Design Practices for Large-Scale Automation
 
Computer Graphics - Lecture 01 - 3D Programming I
Computer Graphics - Lecture 01 - 3D Programming IComputer Graphics - Lecture 01 - 3D Programming I
Computer Graphics - Lecture 01 - 3D Programming I
 
Concurrency - Why it's hard ?
Concurrency - Why it's hard ?Concurrency - Why it's hard ?
Concurrency - Why it's hard ?
 
Scalability broad strokes
Scalability   broad strokesScalability   broad strokes
Scalability broad strokes
 
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018Greenplum Overview for Postgres Hackers - Greenplum Summit 2018
Greenplum Overview for Postgres Hackers - Greenplum Summit 2018
 
BlaBlaCar Elastic Search Feedback
BlaBlaCar Elastic Search FeedbackBlaBlaCar Elastic Search Feedback
BlaBlaCar Elastic Search Feedback
 
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...
Set Transfomer: A Framework for Attention-based Permutaion-Invariant Neural N...
 
Why Concurrency is hard ?
Why Concurrency is hard ?Why Concurrency is hard ?
Why Concurrency is hard ?
 
FrackingPaper
FrackingPaperFrackingPaper
FrackingPaper
 
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...
[Paper reading] Interleaving with Coroutines: A Practical Approach for Robust...
 

Último

08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking Men08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking MenDelhi Call girls
 
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Miguel Araújo
 
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...Drew Madelung
 
Understanding Discord NSFW Servers A Guide for Responsible Users.pdf
Understanding Discord NSFW Servers A Guide for Responsible Users.pdfUnderstanding Discord NSFW Servers A Guide for Responsible Users.pdf
Understanding Discord NSFW Servers A Guide for Responsible Users.pdfUK Journal
 
Histor y of HAM Radio presentation slide
Histor y of HAM Radio presentation slideHistor y of HAM Radio presentation slide
Histor y of HAM Radio presentation slidevu2urc
 
IAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsIAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsEnterprise Knowledge
 
A Call to Action for Generative AI in 2024
A Call to Action for Generative AI in 2024A Call to Action for Generative AI in 2024
A Call to Action for Generative AI in 2024Results
 
Slack Application Development 101 Slides
Slack Application Development 101 SlidesSlack Application Development 101 Slides
Slack Application Development 101 Slidespraypatel2
 
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdfThe Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdfEnterprise Knowledge
 
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking MenDelhi Call girls
 
CNv6 Instructor Chapter 6 Quality of Service
CNv6 Instructor Chapter 6 Quality of ServiceCNv6 Instructor Chapter 6 Quality of Service
CNv6 Instructor Chapter 6 Quality of Servicegiselly40
 
Artificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and MythsArtificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and MythsJoaquim Jorge
 
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptxHampshireHUG
 
Data Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonData Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonAnna Loughnan Colquhoun
 
A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?Igalia
 
2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...Martijn de Jong
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerThousandEyes
 
The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024Rafal Los
 
[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdf[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdfhans926745
 
Exploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone ProcessorsExploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone Processorsdebabhi2
 

Último (20)

08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking Men08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking Men
 
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
Mastering MySQL Database Architecture: Deep Dive into MySQL Shell and MySQL R...
 
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
Strategies for Unlocking Knowledge Management in Microsoft 365 in the Copilot...
 
Understanding Discord NSFW Servers A Guide for Responsible Users.pdf
Understanding Discord NSFW Servers A Guide for Responsible Users.pdfUnderstanding Discord NSFW Servers A Guide for Responsible Users.pdf
Understanding Discord NSFW Servers A Guide for Responsible Users.pdf
 
Histor y of HAM Radio presentation slide
Histor y of HAM Radio presentation slideHistor y of HAM Radio presentation slide
Histor y of HAM Radio presentation slide
 
IAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsIAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI Solutions
 
A Call to Action for Generative AI in 2024
A Call to Action for Generative AI in 2024A Call to Action for Generative AI in 2024
A Call to Action for Generative AI in 2024
 
Slack Application Development 101 Slides
Slack Application Development 101 SlidesSlack Application Development 101 Slides
Slack Application Development 101 Slides
 
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdfThe Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
The Role of Taxonomy and Ontology in Semantic Layers - Heather Hedden.pdf
 
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
 
CNv6 Instructor Chapter 6 Quality of Service
CNv6 Instructor Chapter 6 Quality of ServiceCNv6 Instructor Chapter 6 Quality of Service
CNv6 Instructor Chapter 6 Quality of Service
 
Artificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and MythsArtificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and Myths
 
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
 
Data Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt RobisonData Cloud, More than a CDP by Matt Robison
Data Cloud, More than a CDP by Matt Robison
 
A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?A Year of the Servo Reboot: Where Are We Now?
A Year of the Servo Reboot: Where Are We Now?
 
2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024The 7 Things I Know About Cyber Security After 25 Years | April 2024
The 7 Things I Know About Cyber Security After 25 Years | April 2024
 
[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdf[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdf
 
Exploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone ProcessorsExploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone Processors
 

Pregel: A System For Large Scale Graph Processing

  • 1. Pregel: A System For Large Scale Graph Processing Grzegorz Malewicz, Matthew H. Austern, Aart J. C. Bik, James C. Dehnert, Ilan Horn, Naty Leiser, and Grzegorz Czajkowski Presented By Riyad Parvez
  • 2. Real World Graph Processing ● Web graph: ○ PageRank (influential vertices) ● Social graph: ○ Popularity rank, personalized rank, shortest paths, shared connections, clustering (communities), propagation ● Advertisement: ○ Target ads ● Communication network: ○ Maximum flow, transportation routes ● Biology network ○ protein interactions ● Pathology network ○ find anomalies
  • 3. Graph Processing is Different ● Poor locality of memory access. ● Very little work done per vertex. ● Changes degree of parallelism over the course of execution.
  • 4. Why not MapReduce (MR) ● Graph algorithms can be expressed as series of MR jobs. ● Data must be reloaded and reprocessed at each iteration, wasting I/O, network. bandwidth, and processor resources. ● Needs an extra MR job for each iteration just to detect termination condition. ● MR isn’t very good at dynamic dependency graph.
  • 5. Pregel ● Developed at Google ● Modeled after Bulk Synchronous Parallel (BSP) computing model ● Distributed message passing system ● Computes in vertex-centric fashion ○ “Think like a vertex” ● Scalable and fault tolerant ● Influenced systems like Apache Giraph and other BSP distributed systems
  • 6. Bulk Synchronous Parallel ● Leslie Valiant introduced in 1990. ● Computations are consist of a sequence of iterations, called superstep. ● During superstep, framework calls user-defined computation function on every vertex. ● Computation function specifies behaviour at a single vertex V and a single superstep S. ● Supersteps end with barrier synchronization. ● All communications are from superstep S to superstep S+1. ● Performance of the model is predictable.
  • 7. Bulk Synchronous Parallel Source: https://cs.uwaterloo.ca/~kdaudjee/courses/cs848/slides/jenny.pdf
  • 8. Bulk Synchronous Parallel Source: https://cs.uwaterloo.ca/~kdaudjee/courses/cs848/slides/jenny.pdf
  • 9. Pregel Computation Model ● Computation on locally stored data. ● Computations are in-memory. ● Terminates when all vertices are inactive or no messages to be delivered. ● Vertices are distributed among workers using hash(ID) mod N, where N is the number of partitions (default partitioning) ● Barrier synchronization ➢ Wait and synchronize before the end of superstep ➢ Fast processors can be delayed by slow ones ● Persistent data is stored on a distributed storage system (GFS/BigTable) ● Temporary data is stored in disk.
  • 11. Vertex ● Can mutate local value and value on outgoing edges. ● Can send arbitrary number of messages to any other vertices. ● Receive messages from previous superstep. ● Can mutate local graph topology. ● All active vertices participate in the computation in a superstep.
  • 12. Vertex State Machine ● Initially, every vertices are active. ● A vertice can deactivate itself by vote to halt. ● Deactivated vertices don't participate in computation. ● Vertices are reactivated upon receiving message.
  • 14. Messages ● Consists of a message value and destination vertex. ● Typically sent along outgoing edges. ● Can be sent to any vertex whose identifier is known. ● Are only available to receiver at the beginning of superstep. ● Guaranteed to be delivered. ● Guaranteed not to be duplicated. ● Can be out of order.
  • 15. Combiner ● Sending messages incurs overhead. ● System calls Combine() for several messages intended for a vertex V into a single message containing the combined message. ● No guarantees which messages will be combined or the order of combination. ● Should be enabled for commutative and associative messages. ● Not enabled by default.
  • 17. Aggregator ● Mechanism for global communication, monitoring and global state. ○ Vertices provide value to aggregator in superstep S. ○ Values are combined using a reduction operator. ○ Resulting value is available to all vertices at superstep S+1. ● New aggregator is defined by subclassing "Aggregator" class. ● Reduction operator should be associative and commutative.
  • 19. Topology Mutation ● Vertices can dynamically create/destroy vertices, edges. ● Mutations and conflict resolution take place at barrier. ● Except local mutation (self-edge) immediately takes place. ● Order of mutations ○ Edge deletion ○ Vertex deletion ○ Vertex addition ○ Edge addition
  • 20. Master ● Partitions the input and assigns one or more partitions to each worker. ● Keeps list of ○ All alive workers ○ Worker's unique identifiers ○ Addressing informations ○ Partition of the graph is assigned to the worker. ● Coordinates barrier synchronization i.e., superstep. ● Fault tolerance by checkpoint, failure detection and reassignment. ● Maintains statistics of the progress of computation and the state of the graph. ● Doesn’t participate in computation. ● Not responsible for load-balancing.
  • 21. Worker ● Responsible for computation of assigned vertices. ● Keeps two copies of active vertices and incoming messages ○ Current superstep ○ Next superstep ● Place local messages immediately in message queue. ● Buffer remote messages. ○ Flush asynchronously in single message if threshold reached.
  • 22. Fault Tolerance ● Checkpoint at the beginning of superstep. ○ Master saves aggregators. ○ Workers save vertices, edges and incoming messages. ● Worker failure detected by ping messages. ● Recovery ○ Master reassigns failed worker partition to other available workers. ○ All workers restart from superstep S by loading state from the most recently available checkpoint. ● Confined recovery: recovery is only confined to lost partitions ○ Workers also save outgoing messages. ○ Recomputes using logged messages from healthy partitions and recalculated ones from recovering partitions.
  • 24. Performance Experimental Setup: ● Hardware: A cluster of 300 multi-core commodity PCs. ● Algorithm: SSSP with unit weight edges. • All-pairs shortest paths impractical b/c O(|V|2) storage. ● Measures scalability w.r.t. both the number of workers and the number of vertices. ● Data collected for: • Binary trees (to test scalability). • log-normal random graphs (to study performance in a realistic setting). ● No checkpointing.
  • 28. Summary ● Distributed system for large scale graph processing. ● Vertex-centric BSP model ○ Message passing API ○ A sequence of supersteps ○ Barrier synchronization ● Coarse grained parallelism ● Fault tolerance by checkpointing ● Runtime performance scales near linearly to the size of the graph (CPU bound)
  • 29. Discussion ● No fault tolerance for master is mentioned in the paper (Probably Paxos or replication). ● Static partitioning! What happens if a worker is too slow? ● Dynamic partitioning, network overhead for reassigning vertices and state. ● Good for sparse graph. But communication overhead for dense graph can bring the system down to knees.