�ݺ�ߣ

SEARCH, NETWORK AND ANALYTICS
Using Set Cover to Optimize a Large-
Scale Low Latency Distributed Graph
LinkedIn Presentation Template
©2013 LinkedIn Corporation. All Rights Reserved. 1
Rui Wang
Staff Software Engineer

Outline
 LinkedIn’s Distributed Graph Infrastructure
 The Scaling Problem
 Set Cover by Example
 Evaluation
 Related Work
 Conclusions and Future Work

Outline
 Evaluation
 Related Work

LinkedIn Products Backed By Social Graph

LinkedIn’s Distributed Graph APIs
 Graph APIs
– Get Connections
 “Who does member X know?”
– Get Shared Connections
 “Who do I know in common with member Y”?
– Get Graph Distances
 “What’s the graph distances for each member returned from a search
query?”
 “Is member Z within my second degree network so that I can view her
profile?”
 Over 50% queries is to get graph distances

Distributed Graph Architecture Diagram

LinkedIn’s Distributed Graph Infrastructure
 Graph Database (Graph DB)
– Partitioned and replicated graph database
– Distributed adjacency list
 Distributed Network Cache (NCS)
– LRU cache stores second degree network for active members
– Graph traversals are converted to set intersections
 Graph APIs
– Get Connections
– Get Shared Connections
– Get Graph Distances

Outline
 Evaluation
 Related Work

Graph Distance Queries and Second Degree Creation
get graph distances
cache lookup
[exists=false] retrieve
1st degree connections
[exists=true] set
intersections
Partition IDs
partial merges
K-Way
merges
return
return

The Scaling Problem Illustrated
Graph DBNCS
• Increased Query Distribution
• Merging and De-duping shift to single NCS node

Outline
 Set Cover and Example
 Evaluation
 Related Work

Set Cover Problem
 Definition
– Given a set of elements U = {1, 2, …, m } (called the universe) and a
family S of n sets whose union equals the universe, the set cover
problem is to identify the smallest subset of S the union of which
contains all elements in the universe.
 Greedy Algorithm
– Rule to select sets
 At each stage, choose the set that contains the largest number of
uncovered elements.
– Upper bound
 O(log s), where s is the size of the largest set from S

Modified Set Cover Algorithm Example Cont.
 Build a map of partition ID -> Graph DB nodes for the input graph

 Randomly pick an element
from U = { 1,2,5,6 }
– e = 5
 Retrieve from map
– nodes = { R21, R13 }
 Intersect
– R21 = {1,5 } with U,
– R13 = { 5,6 } with U
 Select R21
– U = {2,6}, C = { R21}

 Randomly pick an element
from U = { 2,6 }
– e = 2
– nodes = { R11, R22 }
 Intersect
– R11 = {1,2 } with U,
– R22 = { 2,4 } with U
 Select R22
– U = {6}, C = { R21, R22 }

 Pick the final element from
U = { 6 }
– e = 6
– nodes = { R23, R13 }
 Intersect
– R23 = { 3,6 } with U,
– R13 = { 5,6 } with U
 Select R23
– U = {}, C = {R21, R22, R23 }

Modified Set Cover Algorithm Example Solution
 Solution Compared to Optimal Result for U , where U = {1,2,5,6}

Outline
 Evaluation
 Related Work

Evaluation
 Production Results
– Second degree cache creation drops 38% on 99th percentile
– Graph distance calculation drops 25% on 99th percentile
– Outbound traffic drops 40%; Inbound traffic drops 10%
95th quantile 99th quantile
0
1
2
3
control setcover control setcover
latency(normalized)
network cache building
95th quantile 99th quantile
0
1
2
3
control setcover control setcover
latency(normalized)
getdistances

Outline
 Evaluation
 Related Work

Related Work
 Scaling through Replications
– Collocating neighbors [Pujol2010]
– Replication based on read/write frequencies [Mondal2012]
– Replication based on locality [Carrasco2011]
 Multi-Core Implementations
– Parallel graph exploration [Hong2011]
 Offline Graph Systems
– Google’s Pregel [Malewicz2010]
– Distributed GraphLab [Low2012]

Outline
 Evaluation
 Related Work

Conclusions and Future Work
 Future Work
– Incremental cache updates
– Replication on GraphDB nodes through LRU caching
– New graph traversal algorithms
 Conclusions
– Key challenges tackled
 Work distribution balancing
 Communication Bandwidth
– Set cover optimized latency for distributed query by
 Identifying a much smaller set of GraphDB nodes serving queries
 Shifting workload to GraphDB to utilize parallel powers
 Alleviating the K-way merging costs on NCS by reducing K
– Available at
https://github.com/linkedin-
sna/norbert/blob/master/network/src/main/scala/com/linkedin/norbert/network/partitioned
/loadbalancer/DefaultPartitionedLoadBalancerFactory.scala

�ݺ�ߣ

Using Set Cover to Optimize a Large-Scale Low Latency Distributed Graph

Recommended

More Related Content

Similar to Using Set Cover to Optimize a Large-Scale Low Latency Distributed Graph (20)

Recently uploaded (20)

Using Set Cover to Optimize a Large-Scale Low Latency Distributed Graph