Graphlab under the hood

12/10/12 1

GraphLab under the hood

Zuhair Khayyat

12/10/12 2

GraphLab overview: GraphLab 1.0

● GraphLab: A New Framework For Parallel Machine Learning

– high-level abstractions for machine learning problems

– Shared-memory multiprocessor

– Assume no fault tolerance needed

– Concurrent access precessing models with sequential-consistency guarantees

12/10/12 3

GraphLab overview: GraphLab 1.0

● How GraphLab 1.0 works?– Represent the user's data by a directed graph

– Each block of data is represented by a vertex and a directed edge

– Shared data table

– User functions:● Update: modify the vertex and edges state,

read only to shared table● Fold: sequential aggregation to a key entry in

the shared table, modify vertex data● Merge: Parallelize Fold function● Apply: Finalize the key entry in the shared table

12/10/12 4

GraphLab overview: GraphLab 1.0

12/10/12 5

GraphLab overview: Distributed GraphLab 1.0

● Distributed GraphLab: A Framework for Machine Learning and Data Mining in the Cloud

– Fault tolerance using snapshot algorithm

– Improved distributed parallel processing

– Two stage partitioning:● Atoms generated by ParMetis● Ghosts generated by the intersection of the

atoms

– Finalize() function for vertex synchronization

12/10/12 6

GraphLab overview: Distributed GraphLab 1.0

12/10/12 7

GraphLab overview: Distributed GraphLab 1.0

GHostsWorker 1 Worker 2

12/10/12 8

PowerGraph: Introduction

● GraphLab 2.1● Problems of highly skewed power-law graphs:

– Workload imbalance ==> performance degradations

– Limiting Scalability

– Hard to partition if the graph is too large

– Storage

– Non-parallel computation

12/10/12 9

PowerGraph: New Abstraction

● Original Functions:– Update

– Finalize

– Fold

– Merge

– Apply: The synchronization apply

● Introduce GAS model:– Gather: in, out or all neighbors

– Apply: The GAS model apply

– Scatter

12/10/12 10

PowerGraph: Gather

Worker 1 Worker 2

12/10/12 11

PowerGraph: Apply

Worker 1 Worker 2

12/10/12 12

PowerGraph: Scatter

Worker 1 Worker 2

12/10/12 13

PowerGraph: Vertex Cut

A

B

H

F

C

G

I

E D

A B

C I

E

D

A H

A G

B H

B C

C

CH

D E D I

E F I

F H F G

12/10/12 14

PowerGraph: Vertex Cut

A B

C I

E

D

A H

A G

B H

B C

C

CH

D E D I

E F I

F H F G

A B

A H

A G

C

B

DC

H

IC

D

I

E

E I

F

F G

12/10/12 15

PowerGraph: Vertex Cut (Greedy)

A B

C I

E

D

A H

A G

B H

B C

C

CH

D E D I

E F I

F H F G

A B

HG

B C

H

DC

C

I EE

F G

12/10/12 16

PowerGraph: Experiment

12/10/12 17

PowerGraph: Experiment

12/10/12 18

PowerGraph: Discussion

● Isn't it similar to Pregel Mode?– Partially process the vertex if a message exists

● Gather, Apply and Scatter are commutative and associative operations. What if the computation is not commutative!

– Sum up the message values in a specific order to get the same floating point rounding error.

12/10/12 19

PowerGraph and Mizan

● In Mizan we use partial replication:

a

b

c

d

e

f

g

W0 W1

a'a

b

c

d

e

f

g

W0 W1

Compute Phase Communication Phase

12/10/12 20

GraphChi: Introduction

● Asynchronous Disk-based version of GraphLab

● Utilizing parallel sliding window– Very small number of non-sequential accesses

to the disk

● Support for graph updates– Based on Kineograph, a distributed system for

processing a continuous in-flow of graph updates, while simultaneously running advanced graph mining algorithms.

12/10/12 21

GraphChi: Graph Constrains

● Graph does not fit in memory● A vertex, its edges and values fits in memory

12/10/12 22

GraphChi: Disk storage

● Compressed sparse row (CSR):– Compressed adjacency list with indexes of the

edges.

– Fast access to the out-degree vertices.

● Compressed Sparse Column (CSC):– CSR for the transpose graph

– Fast access to the in-degree vertices

● Shard: Store the edges' data

12/10/12 23

GraphChi: Loading the graph

● Input graph is split into P disjoint intervals to balance edges, each associated with a shard

● A shard contains data of the edges of an interval

● The sub graph is constructed as reading its interval

12/10/12 24

GraphChi: Parallel Sliding Windows

● Each interval is processed in parallel

● P sequential disk access are required to process each interval

● The length of intervals vary with graph distribution

● P * P disk access required for one superstep

12/10/12 25

GraphChi: Example

(1,2) (3,4) (5,6)

Executing interval (1,2):

12/10/12 26

GraphChi: Example

(1,2) (3,4) (5,6)

Executing interval (3,4):

12/10/12 27

GraphChi: Example

12/10/12 28

GraphChi: Evolving Graphs

● Adding an edge is reflected on the intervals and shards if read

● Deleting an edge causes that edge to be ignored

● Adding and deleting edges are handled after processing the current interval.

12/10/12 29

GraphChi: Preprocessing

12/10/12 30

Thank you

12/10/12 31

thegraphsblog.wordpress.com/

The Blog wants YOU