Querying and Routing in NextGeneration Networks

1
Querying and Routing in Next-Generation Networks

• Boon Thau Loo
• Ph.D. Qualifying Exam Proposal
• 24 Aug 2004

2
Current Internet

• Current Internet Architecture
• Host-centric protocols defined in terms of IP
  addresses.
  
• Routing functionality is embedded in
  infrastructure.
  
• Two limitations
• Difficult to locate dynamic objects by names.
• Applications have little control over the path
  followed by their packets.
  

3
Data-centric Networks

• Data independence Allow users to name (and
  query) data regardless of location.
  
• E.g., Distributed Hash Tables (DHT).
• Addresses the first limitation, but not the
  second (inflexible routing).

4
Evolution of my work

• PIER Relational Query Processor on DHTs.
• Querying the Internet with PIER.
• Application
• Enhancing P2P File-Sharing with PIER.
• Network Monitoring
• Querying (Gathering) Network Topologies.
• Influence base network functionality
• Customizable Routing with Queries.

5
Customizable Routing

• Emerging topic in networking community.
• Why is customizing routing important?
• Flexibility support different application
  requirements.
  
• Evolvability of routing infrastructure.
• Some existing solutions
• Overlay networks (multicast, RON).
• Active Networks.
• Recent ideas i3 Routing Service, NIRA.
• Proposed solution Customizable Routing with
  Declarative (Recursive) Queries.

6
My Unifying Theme

• The synergy between Query Processing and Routing
  in networks.
  
• Three main contributions
• Compare P2P search performance for two P2P
  architectures.
  
• Querying Network Topologies with Declarative
  (Recursive) Queries.
  
• Customizable Routing with Declarative (Recursive)
  Queries.

7
Roadmap

• P2P Search A Comparative Study.
• Querying and Routing in Networks with Recursive
  Queries.
  
• Research Plans and Timeline.

8
P2P Search A Comparative Study

• Qualifying Exam Proposal
• Part I

9
Problem Statement

• P2P Search
• Flooding (Unstructured) vs DHTs (Structured) ?
• Lots of debate, lots of papers, little
  consensus.
  
• Why study P2P Search?
• Canonical P2P application. Live workloads.
• Good stress test on any P2P design.
• More robustness. No single point of failure.
• Social Issues
• More resistent than centralized systems to
  censoring and manipulated rankings.
  
• RIAA.

10
Distributed Hash Tables (DHT)

• Hash table Interface
• put(key,object), get(key)
• Properties
• If object exists in network, it can always be
  found.
  
• Scalable O(log n) hops and state.
• Robust self-configuring and resilient to
  failures and churn.
  
• DHT Search
• Inverted Lists indexed by Keyword.

11
Two Workloads

• P2P Web Search (IPTPS 03)
• P2P File Sharing (IPTPS 04, VLDB 04)
• Less demanding application
• Smaller dataset (millions of files).
• Index by filenames and metadata.
• Less stringent user requirements.
• Replicas of items follow a long-tailed
  distribution.
  
• Popular items at head of distribution.
• Rare items at tail of distribution.

12
Gnutella Measurements

• Main areas of study
• Gnutella Topology
• Crawl from multiple vantage points on PlanetLab.
• Search Quality
• Measure query results size and latency.
• Reissue Gnutella queries from 30 LimeWire
  Ultrapeers simultaneously on PlanetLab
  Approximate perfect answer.

13
Summary of Measurements

• Queries with few results are searching for rare
  items.
  
• Searching on Gnutella
• Highly effective for popular items.
• Less effective for rare items.
• Significant opportunity to do better.
• Large fraction of queries return few or no
  results even when they exist.
  
• Bad response times for queries on rare items.

14
Hybrid Solution
Flood-based Network (All items)
DHT (Index Rare Items)
15
PlanetLab Deployment
L
L
Horizon of P1
U2
L
L
P2
P1
U1
L
P3
L
L
L
L
L
L
Gnutella Leaf
Gnutella Ultrapeer
U
Gnutella links
Hybrid Ultrapeer (PIER Gnutella)
P
PIER links
16
Gnutella Measurement Study Important Lessons

• Gathering Gnutella Topology requires recursive
  link traversals.
  
• Knowledge of topology can improve the quality of
  search.
  
• Challenge
• Can we perform topology discovery efficiently and
  accurately?
  
• If so, what other functionality can we provide?

17
Querying and Routing in Networks with Recursive
Queries

• Qualifying Exam Proposal
• Part II

18
Introduction

• The Internet is made up of distributed graphs!
• IP Routers.
• Overlay networks.
• WWW Hypertext structures.
• Recursive queries could be used to discover
  topologies.
  
• A recursive query engine is an attractive routing
  infrastructure!
  
• Well see routing protocols (DV, DSR) are just
  recursive queries.
  
• Customizable End-hosts can express their own
  route desires.
  
• Efficient? Query Optimization Techniques.

19
Outline

• Intro to Recursive Queries.
• Applications.
• Recursive Query Processing and Optimization
  Techniques.
  
• Current Status.
• Research Agenda and Timeline.

20
Background Datalog Program
R1 reachable(S,D) - link(S,D)
R2 reachable(S,D) - link(S,Z), reachable(Z,D)
Query reachable(M,N)

• R1 1-hop reachable
• R2 ? 2 hop
• reachable(a,N) reachable from node a

L(S,Z)
R(Z,D)
S
Z
D
R(S,Z), R(S,D)
21
Datalog Facts

• Base Facts
• Supplied to query processor.
• node(nodeID, load, ), nodeID is either an IP
  address or DHT identifier.
  
• link(source, destination, cost, )
• Derived Facts
• Intermediate data
• reachable(source, destination)
• path(source, destination, path, cost)
• Result Facts
• Sent back as query results, or stored in
  network.
  
• E.g., ShortestPath, NextHop

22
Execution Model

• Distributed Query Processing
• Each node embeds query processing functionality.
  PIER is an example system but this model is not
  constrained to DHTs.
  
• Each query processor has access to local base
  facts.
  
• Query Execution
• Recursive Query is issued by one of the nodes.
• Disseminated to all or subset of other nodes for
  execution.
  
• Wrapper for external networks
• PIER runs a recursive crawl query to gather
  information on external network.
  
• Each node is responsible monitoring a subset of
  external network nodes.
  

23
New Challenges

• Different Metrics
• Centralized I/O, CPU, Number of facts.
• Distributed Communication Overhead, Latency.
• Network is dynamic and soft-state.
• Long running, concurrent queries.

24
Parallel and Distributed Deductive Databases

• Parallel
• Hash-based partitioning (data fragmentation).
• Direct (matrix methods).
• Distributed
• Semantic Fragmentation (Disconnection Set).
• Main differences
• Multi-hop environment
• Relationship to routing algorithms.
• Setting up network state.
• Dynamic graphs.
• Long-running, concurrent queries.
• Smaller scale compared to ours.

25
App 1 Network Topology Monitoring

• Gnutella Monitoring Service
• Search horizon statistics (number of nodes,
  files)
  
• Diameter of the network.
• Robustness of the network.
• Direct search query towards high degree nodes.
• Study a DHT under churn
• Dynamic Resilience How many possible live
  paths are there between any two nodes?
  
• Average Path Length Given routing algorithm,
  what is the average number of hops between any
  two nodes?
  
• Check for invariants.

26
App 2 Customizable Routing Infrastructure

• Best-Path routing
• Shortest Path (Distance Vector)
• Shortest-k-paths
• Least-loaded path
• Disjoint-Paths greedy routing
• Disjoint-k-paths (edge and node disjoint)
• Dynamic Source Routing (DSR)
• Policy Decision
• Paths that include/exclude certain nodes
• Do not carry/trust traffic from certain nodes.
• What are Datalogs limitations?

27
Recursive Query Processing in Networks

• Introduction to Recursive Queries
• Applications
• Recursive Query Processing and Optimization
• The Basics
• Datalog ? Query Plan
• Query Execution
• Query Optimization Techniques
• Work-Sharing
• Queries over Dynamic Graphs
• Current Status
• Research Plan

28
Datalog ? Query Plan

• R1 reachable(S,D) - link(S,D)
• R2 reachable(S,D) - link(S,Z), reachable(Z,D)
• Query reachable(M,N)

R2
Ship tuples to table.field
R1
29
Query Execution
l(a,b), l(a,c)
a
r(a,b), r(a,c)
0th Iteration
l(a,g), l(a,d), l(b,e)

• R1 reachable(S,D) - link(S,D)
• R2 reachable(S,D) - link(S,Z), reachable(Z,D)
• Query reachable(M,N)

l(c,e)
c
b
r(c,e)
r(a,g), r(a,d), r(b,e)
l(e,f)
d
e
l(d,f)
r(e,f)
r(d,f)
l(f,h)
f
g
r(f,h)
l(g,f)
r(g,f)
l(h,i)
h
r(h,i)
i
30
Query Execution
a
r(a,b), r(a,c)
1st Iteration
l(a,c)
l(a,b)
r(b,d), r(b,e), r(b,g)

• R1 reachable(S,D) - link(S,D)
• R2 reachable(S,D) - link(S,Z), reachable(Z,D)
• Query reachable(M,N)

c
r(c,e)
b
l(b,e)
l(c,e)
l(b,d)
d
e
r(e,f)
r(d,f)
l(b,g)
l(e,f)
l(d,f)
l(g,f)
f
r(f,h)
g
l(f,h)
r(g,f)
h
r(h,i)
l(h,i)
i
31
Network-Reachability Query
r(a,b), r(a,c), r(a,e), r(a,d), r(a,g)
a
2nd Iteration
r(b,d), r(b,e), r(b,g), r(b,f)
l(a,b)
r(c,e), r(c,f)

• R1 reachable(S,D) - link(S,D)
• R2 reachable(S,D) - link(S,Z), reachable(Z,D)
• Query reachable(M,N)

c
b
l(a,c)
r(e,f), r(e,h)
r(d,f), r(d,h)
d
e
l(b,e), l(c,e)
l(b,d)
l(d,f), l(g,f), l(e,f)
r(f,h), r(f,i)
f
g
r(g,f), r(g,h)
l(b,g)
r(h,i)
h
l(f,h)
i
l(h,i)
32
Network-Reachability Query
r(a,b), r(a,c), r(a,e), r(a,d), r(a,g), r(a,f)
a
3rd Iteration
l(a,b)
r(b,d), r(b,e), r(b,g), r(b,f), r(b,h)
r(c,e), r(c,f), r(c,h)

• R1 reachable(S,D) - link(S,D)
• R2 reachable(S,D) - link(S,Z), reachable(Z,D)
• Query reachable(M,N)

c
b
l(a,c)
r(d,f), r(d,h), r(d,i)
l(b,e), l(c,e)
r(e,f), r(e,h), r(e,i)
d
e
l(b,d)
l(d,f), l(g,f), l(e,f)
f
r(f,h),r(f,i)
g
r(g,f), r(g,h), r(g,i)
l(b,g)
h
r(h,i)
l(f,h)
i
l(h,i)
33
Remarks

• Resembles Distance Vector Protocol
• Computation begins with initial reachable set and
  shipping it to all neighbors. (R1)
  
• Neighbors update reachable set with its own
  neighborhood set, and forward resulting reachable
  set to neighbors. (R2)
  
• Computes all-pairs paths Work is shared by all
  nodes.
  
• Converges after Network-Diameter rounds of
  communication.

R1 reachable(S,D) - link(S,D)
R2 reachable(S,D) - link(S,Z), reachable(Z,D)
34
Distance Vector Routing

• Routing Table Formation
• An entry in the routing table nextHop(src,dst,nex
  thop,cost)
  
• R1 path(S,D,D,C) - link(S,D,C)
• R2 path(S,D,Z,C) - link(S,Z,C1),
  path(Z,D,W,C2), CC1C2
  
• R3 bestPathLength(S,D,min) -
  path(S,D,Z,C)
  
• R4 nextHop(S,D,Z,C) - nextHop(S,D,Z,C),
• bestPathLength(S,D,C)

• Changes
• New rules R3 and R4.
• Path stores the next hop in path.

35
Query Optimization Techniques

• The Basics
• Query Optimization
• Reduce communication overhead
• Magic Sets Rewrite
• Left-Right Recursion Rewrite
• Aggregate Selections
• Reduce latency
• Squaring Algorithm
• Work Sharing
• Queries over Dynamic Graphs

36
Opt 1 Magic Sets Rewrite

• R1 magicSource(D) - magicSource(S), link(S,D)
• R2 reachable(S,D) - magicSource(S), link(S,D)
• R3 reachable(S,D) - magicSource(S), link(S,Z),
  reachable(Z,D)
  
• R4 magicSource(b)
• R5 magicSource(e)
• Query reachable(M,N)

What if we do not need the recursive query to be
computed on a portion of the graph?
37
Combine common expressions

• R1 magicSource(D) - s(S,D)
• R2 reachable(S,D) - s(S,D)
• R3 reachable(S,D) - s(S,Z), reachable(Z,D)
• R4 magicSource(b)
• R5 magicSource(e)
• R6 s(S,D) - magicSource(S), link(S,D)
• Query reachable(M,N)

38
Optimized Magic Rewrite Plan I
R1 magicNodes(D) - s(S,D) R2 reachable(S,D) -
s(S,D) R3 reachable(S,D) - s(S,Z), reachable(Z
,D) R4 magicSource(b) R5 magicSource(e) R6 s
(S,D) - magicSource(S), link(S,D)
Query ?-reachable(M,N)
R3
R2
R1
R6
39
Optimized Magic Rewrite Plan II
R1 magicNodes(D) - s(S,D) R2 reachable(S,D) -
s(S,D) R3 reachable(S,D) - s(S,Z), reachable(Z
,D) R4 magicSource(b) R5 magicSource(e) R6 s
(S,D) - magicSource(S), link(S,D)
Query ?-reachable(M,N)
R3
R1
R2
R6
40
Executing Magic Reachable
0th Iteration
a
l(a,b), l(a,c)
l(b,d), l(b,e), l(b,g)
l(c,e)
c
b
r(b,d), r(b,e), r(b,g)
l(e,f)
l(d,f)
d
e
r(e,f)
f
l(f,h)
g
l(g,f)
l(h,i)
h
i
41
Executing Magic Reachable
1st Iteration
a
r(b,d), r(b,e), r(b,g)
c
b
l(b,e)
l(b,d)
r(d,f)
d
e
l(b,g)
l(e,f)
f
r(f,h)
g
r(g,f)
h
i
42
Executing Magic Reachable
2nd Iteration
r(b,d), r(b,e), r(b,g), r(b,f)
r(e,f), r(e,h)
r(d,f)
l(b,d)
l(b,e)
l(d,f)
l(g,f)
r(f,h)
l(e,f)
l(f,h)
r(g,f)
l(b,g)
r(h,i)
43
Remarks on Magic Rewrite
a
c
b

• Magic sets limits computation to portion of
  graph.
  
• Limit by source and/or destination.
• Few problems
• Does not help for undirected graphs.
• Still some redundant fact generation.
• Increased number of iterations

d
e
f
g
h
i
44
Opt 2 Left-Right Recursion Rewrite

• R1 reachable(S,D) - magicSource(S), link(S,D)
• R2 reachable(S,D) - reachable(S,Z), link(Z,D)
• R3 magicSource(b)
• R4 magicSource(e)
• Query ?-reachable(M,N)

R2
R1
45
Executing Left-Recursion Plan
a
1st Iteration
r(b,d), r(b,e), r(b,g)
c
b
r(b,e)
r(b,d)
d
e
r(e,f)
l(d,f)
r(b,g)
r(e,f)
f
l(f,h)
g
l(g,f)
h
l(h,i)
i
46
Executing Left-Recursion Plan
a
2nd Iteration
r(b,d), r(b,e), r(b,g), r(b,f)
c
b
r(e,f), r(e,h)
d
e
l(d,f)
r(b,f)
r(b,f)
r(b,f)
f
l(f,h)
g
r(e,h)
l(g,f)
h
i
47
Executing Left-Recursion Plan
a
3rd Iteration
r(b,d), r(b,e), r(b,g), r(b,f), r(b,h)
c
b
r(e,f), r(e,h), r(e,i)
d
e
l(d,f)
f
l(f,h)
g
r(b,h)
l(g,f)
l(h,i)
h
r(e,i)
i
48
Remarks on Left Rewrite
a

• Resembles Dynamic Source Routing (DSR). Same
  query, different routing protocol !!
  
• Lower communication overhead when there are few
  source nodes.
  
• Sharing is not implicit in query. Each node
  computes its facts independently of other nodes.

c
b
e
d
R1 reachable(S,D) - magicSource(S), link(S,D)
R2 reachable(S,D) - reachable(S,Z), link(Z,D)
R3 magicSource(b) R4 magicSource(e) Query ?-r
eachable(M,N)
f
g
h
i
49
Work Sharing

• Sharing among queries with identical rules
• If all nodes running the same query, use
  right-recursion.
  
• If only a few nodes running the same query, use
  left-recursion.
  
• Switching from left-to-right recursion can be
  expressed as a rewrite

a
c
b
e
d
R1 reachable(S,D) - magicSource(S), link(S,D)
R2 reachable(S,D) - ?magicSource(Z),
reachable(S,Z), link(Z,D)
R3 reachable(S,D) - magicSource(Z),
reacjable(S,Z), reachable(Z,D)
f
g
h
i
50
Work Sharing

• Merge common expressions for queries with
  different rules
  

R1 path(S,D,P,C,L) - link(S,D,C2,L2),
path(S,D,P1,C1,L1),
PconcatPath(link(X,Z), P1),
CFUN1(C1,C2), LFUN2(L1, L2)
R2 path(S,D,P,C) - link(X,Y,C),
PconcatPath((link(X,Y), nil) R3 bestPath(S,D,AG
G1(,AGG2()) - path(S,D,P,C,L)
51
Queries over Dynamic Graphs

• In practice, queries (both routing and
  monitoring) are long running.
  
• Timestamp derived facts
• Each base fact is maintained as soft-state with
  timestamp.
  
• Timestamp derived facts based on oldest base fact
  used in computation.
  
• Maintain state of long-running queries
• Incremental computation when new base facts are
  added.

52
Current Status

• PIER supports cycles in query plans.
• Naïve implementations of web and gnutella
  crawlers. Test-runs on PlanetLab.
  
• PIER hand-optimized dataflows
• Left vs Right recursion.
• Magic Sets.
• Semi-naïve vs Squaring algorithm.
• Aggregate Selections.
• Reachable, Shortest Path, Diameter Queries.
• Static graphs, no sharing.
• Initial PIER simulation results detailed in tech
  report
  
• Querying Network Graphs with Recursive Queries.
  UC Berkeley Technical Report UCB//CSD-4-1332, Jun
  2004.

53
Datalog ? Optimized Plan

• Bulk of research work.
• Data placement.
• Magic Sets
• Directed vs Undirect.
• How many participating nodes?
• Left vs Right recursion
• How many participating nodes?
• Aggregate Selections
• Density of network?
• Monotonic aggregate?
• Semi-naïve vs Squaring
• Density of network.

54
Datalog ? Optimized Plan

• Work-sharing / Materialization
• How many participating nodes? Sources?
  Destinations?
  
• Presence of similar overlapping queries?
• Presence of queries with common rules or
  correlated metrics?
  
• Presence of hub nodes?
• Queries restricted to a domain?
• Good vs Best plan?
• What if conditions change during query execution?
  Optimal plan may become sub-optimal Adaptive
  optimization techniques.
  

55
Research Plan (22 month timeline)
Infrastructure
Applications

• Phase I (Sept 04 Dec 04)
• Hand-Optimized dataflows
• Tradeoffs of different query plans
• Basic sharing
• Long-running queries
• Phase II (Jan 05 Dec 05)
• Datalog parser
• Automatic plan generation
• Enhanced Sharing
• Other advanced features

• Phase I (Nov 05 - Aug 05)
• Gnutella Monitoring Service
• Routing Infrastructure Service
• Phase II (Sept 05 - Dec 05)
• Decentralized Focused Web Crawler
• Study of Different DHTs under churn

Jan-Jun 2006 Wrap up