Data Replication in Mobile Ad Hoc Networks

1
Data Replication in Mobile Ad Hoc Networks

• Project Group PaMANet
• December 18, 2003
• Marcel Tiemeyer

2
Overview

• Replication
• Broadcasting Invalid Reports (IR)
• Advanced broadcasting Hybrid Predicate Invalid
  Reports (HPIR)
• Performance of HPIR-Broadcasting
• Outlook

3
Replication Why?

• Advantages
• Local availability of data
• Fast local access
• No permanent connection to master-database
• Limited bandwidth and broken connections only
  smaller problem
• Less workload on server
• Disadvantages
• Transferring updated data from one database to
  all other databases

4
Symmetric/Asymmetric replication

• Symmetric
• All copies are equal in all respects
• Asymmetric
• Designated copy (master) might have very
  different role than other copies (replicas)

5
Single/Floating/Multi Master

• Single master
• Only one master can perform updates
• Replicas read-only
• Floating master
• Like single master
• Master goes down ? one replica chosen to be new
  master
• Multi master
• Multiple master copies with read/write access
• Conflict resolution strategies/transactions
  necessary

6
Classification

• If users have access to inconsistent data, they
  may possibly make wrong decisions ? classified
  by latency
• Synchronous replication data must be totally
  consistent at all times
• latency is zero
• Asynchronous replication data must have
  acceptable level of inconsistency for short
  period of time
• latency is greater than zero

7
Synchronous replication

• Can be achieved by two-phase-commit
• Coordinator issues update-statement to clients
• Clients begin updating
• Coordinator asks for commit
• Clients answer commit or rollback
• If all clients answer commit coordinator
  notifies clients of successful updateElse
  coordinator sends rollback to all clients
• Clients notify coordinator

8
Voting

• Can also be achieved by voting (quorum)
• coordinator asks all clients to vote OK or not
  OK for current operation
• if enough clients (e.g. half of the clients for
  write-operations, one third for
  read-operations) vote OK operation is
  performed
• weighted voting (e.g. master has five votes,
  clients only one) possible
• Make sure, that
• Qw(A) Qw(A) gt W(A) (only one write-operation)
• Qr(A) Qw(A) gt W(A) (no read- and
  write-operation)

9
Snapshots

• Snapshot is simple example for asynchronous
  replication
• Read-only-copy of master is taken
• Snapshot only consistent from moment of
  snapshot-taking to first update on master

10
Broadcasting IR-messages

• Ruling principles
• Server sends broadcast-messages to all clients
  (e.g. every 10 minutes) including all updated/new
  records (since last broadcast)
• Clients request and download changed records
• New disadvantages
• Many messages not necessary for all clients
• Clients do not know, if they can satisfy a query
  locally
• Waste of bandwidth

11
Hybrid Predicate Invalid Report (HPIR)

• Server still sends broadcast-messages containing
  changed records to all clients
• Clients still request and download changed records

• but
• HPIR contains more information
• Clients can decide whether they need updated
  records or not
• Clients know, if they can satisfy a query locally

12
Schematic diagram
Server
Clients
HPCD
PR-Tree
Cache



data

update

Update-queue

Broadcast-queue
HPIR-Item
13
(Non-)Location Dependent Attribute (LDA/NLDA)

• Strategy based on Location Dependent Data (LDD)
• Relation divided into
• LDAs Location Dependent Attributes (e.g.
  x-position, y-position)
• NLDAs Non-Location Dependent Attributes

14
Example traffic information system
somewhere the distance (e.g. at
x-position150 and y-position70)
free parking space??

• every parking space has the following attributes
• x-position
• y-position
• full
• name

Location Dependent Query (LDQ) Give me all free
parking lots within 10 km which are not full

represented as Q(fullfalse) and
(140x-position160) and (60y-position80)
LDAs
NLDAs
15
Hybrid Predicate Invalid Report (HPIR)

• Hybrid Predicate Invalid Report is broadcasted
  and contains
• Hybrid Predicate Representation (HPR)
• Operation (PU, RU)
• Timestamp
• RecordID (only for PU)

16
Hybrid Predicate Representation (HPR)

• Hashed LDAs
• Hashed by multi-attribute hashing functions (Ax ?
  ax, Ay ? ay )
• Each LDA hashed separately
• Additional string to indicate exact range value
  (ax, ay)
• NLDAs not hashed (they stay a3,,an)
• Predicate for LDQ axay,ax,ay,a3,,an

17
Example (continued)
Hybrid Predicate of LDQ Q(fullfalse) and
(140ltx-positionlt160) and (60lty-positionlt80)
denoted as 001000,4060,6080,0,
Hashing function

• 000 xlt100
• 001 100xlt200
• 010 200xlt300
• 011 300xlt400
• 100 400xlt500
• 101 500xlt600
• 110 600xlt700
• 111 700x

H(x)
18
Operations

• Point Update Operation (PU)
• Only one record is updated
• Range Update Operation (RU)
• Set of records (depending on range predicate) is
  updated

19
R-Tree

• Data is stored in a PR-Tree which is based on
  R-Trees
• Tree-structure designed for spatial searches
• Height balanced
• Dynamic index structure
• n-dimensional
• Leaf nodes contain index records which contain
  pointers to data objects
• Derivation (PR-Tree) is used on server

20
PR-Tree

• Leaf node stores
• Rectangle section depending on LDAs
• Update frequency
• Query frequency
• Timestamp of last update
• Timestamp of last query

21
Insertion algorithm (server)

• After Location Dependent Query (LDQ) is executed,
  server may insert LDAs into PR-Tree
• Compute set of hashing sections
• Update query frequency and timestamps
• If tree contains rectangle return result and
  attach code n_covered
• Else If update-(query-) frequency is too high
  (low), do not insert into tree but return result
  and code n_discarded
• Else insert rectangle into tree and return code
  n_cached

22
HPIR-Queue

• HPIR-Queue stores HPIR-items for records affected
  by update until next broadcast
• Construction
• Update values
• Find affected leaf nodes
• Construct new HPIR
• Insert into HPIR-queue

23
Storage on client - HPCD

• Hybrid Predicate Cache Description (HPCD)
  contains
• ID
• HPR of data records received from server
• Content (pointer to real data buffer)
• Update queue (records to be retrieved)
• Timestamp of last update
• Update-count

24
Updating HPCD (client)

• Listen to HPIR-Broadcasting, download HPIRs which
  correspond to local HPCD
• Point Update
• If recordID is in cache update
• If not cached and all attributes satisfy
  predicate of HPCD insert into update queue
• Range Update
• Update values which are in the cache
• Find records which would satisfy predicate of
  HPCD after the update. Insert conditions into
  update queue

25
Example (continued)
1. No relevant HPIR received
somewhere the distance
free parking space??

• Local cache is queried
• no records in update queue
• 4 free parking spaces nearby

26
Example (continued)
2. HPIR received
somewhere the distance
A parking lot
free parking space??

• Local cache is queried
• Records in update queue ? download data from
  server
• Only 2 free parking spaces nearby

27
Performance

• 200 clients
• 2000 records in database
• 10 minute broadcast interval
• 600 hot records
• Probability of accessing hot data 80

28
Benefits of this approach

• Client queries server only if query uses records
  stored in update queue
• Client can update some cache content locally
  instead of downloading modified data
• Decision whether query can be satisfied locally
  is easy
• ? HPIR can reduce transmission

29
Outlook

• No LDD but some kind of partitioning
• Broadcasting e.g. per access-point is not too
  expensive
• Intelligent broadcasting could be solution for us

30

• Thank you for your attention!
• Any questions?