Dissemination and Synchronization for Mobility (and Beyond)

Dissemination and Synchronization for Mobility (and Beyond)

Michael FranklinUC Berkeley

MDM Tutorial7 January 2001

© 2001 Michael J. Franklin MDM 2001 Tutorial 2

Outline1. Dissemination vs. Synchronization2. Architectural Concepts

– Types of nodes– Data Delivery Mechanisms– User Profiles

3. Data Dissemination– DBIS Toolkit, Xfilter, Continuous Queries

4. Synchronization– for PDAs: Palm HotSync, Edison, SyncML– Data Recharging– Consistency for weakly connected devices

5. Wrap Up


Intro: Data Dissemination• disseminate

– 1. To scatter widely, as in sowing seed.

– 2. To spread abroad, promulgate. disseminate information

• In a data management context, this refers to the proactive distribution of relevant data to users.

• Examples:– News feeds, stock tickers, event broadcasts,

SPAM, …


Intro: Data Synchronization

• synchronize – 1. To cause to occur or

operate with exact coincidence in time or rate.

– 2. To cause to occur or operate at the same time as something else.

• In a data management context this refers to making base data and device-cached data consistent.

• Examples:– Palm HotSync, Email (?), disconnected operation


Discussion• From the definitions, you might think that the two

concepts are completely unrelated, but are they?• Examples:

– Email Lists/On-line communities– Groupware apps such as shared calendars– AvantGo

• What are the essential characteristics that distinguish one from the other?

• How related? How different?


Tutorial Goals• To identify common infrastructure to support

large scale data distribution: dissemination and syncrhonization.

• To describe recent and on-going research in supporting dissemination.

• To describe existing synchronization protocols and future directions for them.

• To outline avenues for continuing research and infrastructure development.


2. Architectural Concepts• Dissemination and Sync are inherently distributed;

– Both require a Network architecture.• A key concept is that of an Overlay Network

1. “application-level” network built on top of Internet protocols; interacts with the “regular” internet.

2. May use both public and private communication links.3. Exploits “Data Centers” deployed around the world.4. Content Routing can be done at the application level so

can be based on application and data semantics.5. Caching, Prefetching, Staging, etc. can be done

transparently.6. E.g., CDNs such as Akami, FastForward Networks


Architecture (continued)

• We will focus on three key aspects of such architectures:

1. Types of nodes in the system.2. Options for data delivery mechanisms.3. Representation of data needs and

preferences through user profiles.


i) Types of Nodes

• Clients– Interact with end user, may cache data and updates

• Client Proxies – Deal with disconnection, provide network interface

• Data Sources– The ultimate repositories for data

• Intermediaries (“Information Brokers”)– Provide storage/caching, application level routing– value added data processing– communications transducing


Network Components

Internet

profile

query

response

profile

query

response

DataSources Information

Brokers

ClientProxies


ii) Data Delivery Options• There are many ways to move data between

sources and receivers:• Pull vs. Push

– Does the data move because the receiver asked for it or because the source decided to send it?

• Periodic vs. Aperiodic– Does the data move according to a predefined

schedule or is movement event/demand driven?• Unicast vs. 1 to N

– Does the data go to a single receiver or many?• Reliability Guarantees

– best effort, guaranteed once, transactional…


Data Delivery Mechanisms

PushPull

Aperiodic Periodic

Unicast 1-to-n Unicast 1-to-n

Aperiodic Periodic

Unicast 1-to-n Unicast 1-to-nrequest/response

on-demandbroad-cast

polling pollingw\snoop

Email lists

publish/subscribe

Person- alizedNews

Broad-castdisks

Dimensions are largely orthogonal – all combinations are potentially useful.

[Franklin & Zdonik, OOPSLA 97]


Network TransparencyClientsBrokersSources

A fundamental principle for systems design:Type of a link matters only to nodes on each end.


iii) User Profiles• An expression of a user’s (or group of users)

data interests and priorities.• Must be Declarative:

– Query languages enabled modern database systems.– Profile languages will enable next generation

information management.• Sources:

– users– learned (implicitly or through feedback)– hybrid– collaborative/clustering approaches


Why are Profiles Needed?

• Necessary for push-based dissemination– how else to know what to send to user?

• Useful for optimizing data synchronization– can precompute data to be transferred to user– can identify potential hot spots

• Also can be used for data management– Caching– Staging at brokers and proxies– Prefetching– Precomputation of customized data views


Profile ContentsThree main components:

1) Domain Specification: content-based, declarative specifications of user interests (read “queries”).

2) Utility Specification: Specifications of user priorities and dependencies among data items and requirements for resolution, freshness, ordering, etc.

3) User Context information: where, when, who, what. Useful for tailoring data delivery to users based on their current and future needs.


Example ProfileWHERE

<article><subject> Database <\><title> $t <\><year> $y <\><conference> $c <\>

<\> ELEMENT AS $XIN (www.cs.*.edu/*/$S), $S conforms to “bib.dtd”CONSTRUCT $XUTILITY ( $X )

(10 * ( $c = “SIGMOD” OR $c = “VLDB”)) +(8 * ( $c = “EDBT” OR $c = “ICDE”) +(100 * ( $a = “Gray”)) -(2001 - $y)


Summary So Far• Despite initial impressions, Dissemination and

Synchronization are closely related.– A common infrastructure can support both.

• Basis is an overlay network with application-level routing, transparent caching, staging, etc.– Nodes are clients, proxies, brokers, and sources.– Various data delivery mechanisms combined via

network transparency.

• User profiles are the key to push-based delivery, precomputation, and network data management.


3. Data Dissemination

• Stock and sport tickers• Personalized news delivery• Traffic information systems• Software distribution

Selective Dissemination of Information (SDI)•the right data to the right people at the right time

Demand for timely dissemination

of data to a large set of consumers

Huge Amount of Dynamic

Data

Ubiquity of Information

Services

•Asymmetric (server to devices) data flow/usage dictates system architecture.


Dissemination Topics

1. The DBIS Toolkit

2. XFilter: efficient routing and filtering of XML documents.

3. Related Database technologies: triggers and continous queries.


Dissemination-Based Information Systems (DBIS)• Outgrowth of “Broadcast Disks” project.

SIGMOD 95 (Acharya et al.)• Framework proposed OOPSLA 97 (Franklin & Zdonik)• Toolkit description/demo SIGMOD 99 (Altinel et al.)• XML-based Profile system (Xfilter) in

VLDB 00 (Altinel & Franklin)• Profile learning techniques in

ICDE 00 (Cetintemel, Franklin, Giles)

• Now part of “Data Centers” NSF ITR Project with Stan Zdonik @ Brown & Mitch Cherniack @ Brandeis- focus on profile-based data management


DBIS FrameworkThe DBIS Framework is based on three fundamental

principles:

1) No one data delivery mechanism is best for all situations (e.g., apps, workloads, topologies).2) Network Transparency: Must allow different mechanisms for data delivery to be applied at different points in the system.3) Topology, routing, and delivery mechanism should vary adaptively in response to system changes.

Goal is to provide a library of components from which to construct dissemination apps.


DBIS Example

1-to-n pushServerDB

Proxy cache

An example:

Can vary dynamically

Unicast pull

Proxy cache

Proxy cache

Unicast pull

Unicast pull


DBIS Toolkit

• Data Source Library – wraps data sources to encapsulate communication and convert data.

• Client Library – encapsulates comm., converts queries and profiles, monitors and filters data.

• Information Broker – primary component of the DBIS. Handles communication transducing, caching, scheduling, profile management and matching.

• Catalog Manager (master)• Real-Time Performance Monitoring Tool and

Control Panel.


DBIS Components


Cache

Network Manager

Profile Manager

ClientManager

Acknowledgement (Tune information)

Data SourceManager

Profiles / Pull Requests

Decomposed Profiles /

Profile Updates

ForwardedProfiles

Data Source Registration

Pull Requests

Catalog Updates

Broker Manager

Data Sources

Data ItemsData Items

Dat

a So

urce

s

IB Master

ClientsBroadcast Medium

Other Information Brokers

Profiles / Pull Requests

HD

Filtered Data

Mapper

Scheduler

Broadcast Manager

Information Broker


More on Brokers• Brokers are middleware components that can

act as both clients and servers.• Must support data caching

– Needed to convert pushed-data to pulled-data– Also allows implementation of hierarchical caching

• Profile Management– Profiles needed for push– Allow informed data management: prefetch,

staging, etc.• Profile Matching

– No profile language sufficient for all applications.– Need an API for adding app-specific profiling


DBIS Toolkit


DBIS Research Issues• Each data delivery mechanism has unique

aspects– Broadcast Disks - scheduling., caching,

prefetching, updates, error handling,…– On-demand Broadcast - scheduling, data staging– Publish/Subscribe- large-scale filtering,

channelization• Security/Fault-tolerance/Reliability• End-to-End network design and control• Fundamental performance tradeoffs• Profile Languages and Processing


XFilter: XML Document Filtering• Provides efficient filtering (routing) of XML documents

against many XPath profiles by:– Representation of XPath queries as Finite State

Machines (FSMs)– Sophisticated FSM indexing and processing– Enhancements to avoid “query” skew

• Accepts any XML document (no DTDs needed)• Implemented in the DBIS-Toolkit and as a stand-alone

library• Developed by Mehmet Altinel for his Ph.D. work,

Published in [Altinel & Franklin, VLDB 2000]


Why XML-Based SDI?

• XML is becoming the dominant format for data exchange on the Internet

• XML provides structural and semantic cues

• Query languages for XML have been developed

• The combination of XML encoding and expressive query languages allows the creation of highly focused and accurate profiles


The challenge is to efficiently and quickly match incoming XML documents against the potentially huge set of user profiles.

An XML-Based SDI System

XML Conversion

XML Documen

ts Filter Engine

User Profiles

Users

Filtered Data

Data Sources


XPath as a Profile Language• W3C recommendation (used for path

expressions in XSLT and XPointer)• Has the right level of expressiveness for SDI

– Operates on a single document at a time– Can address any node in an XML document using

hierarchical relationships, wildcards and element node filters

• In XFilter, we use XPath to describe predicates over entire documents– If the result contains at least one element of a document,

then the document satisfies the XPath expression


Important XPath Features

• Parent/Child (‘/’) and Ancestor/Descendant (‘//’): /catalog/product//msrp

• Wildcards (match any single element):/catalog/*/msrp

• Element Node Filters to further refine the nodes:– Filters can contain nested path expressions

//product[price/msrp < 300]/nameFilter applied to

product element node


XFilter Architecture

XPath Parser

Filter Engine

Path Nodes Profile Info

XML Document

s

XML Parser(SAX Based)Element

Events

SuccessfulProfiles &

Filtered Data

ProfileBase

SuccessfulQueries

Query Index

User Profiles(XPath Queries)

/a//b/c//b/d/*/e/c/*/d//e

/a/b[c/d]/e//d/*/*/e/b/e


XML Parsing and Filtering

• Event-based XML Parsing using SAX API• XML documents are converted to a linear

sequence of events that drive the execution of the filter

• Callback functions are implemented to deal with the different events– Start Element– Element Data– End Element


Filter Engine• Tricky aspects of the XPath language:

– Checking the order of elements in the queries– Handling wildcards and descendent operators– Evaluating filters that are applied to element

nodes (Nested path expressions)• Solution:

– Convert each XPath query into a Finite State Machine (FSM)•A profile is considered to be satisfied when

its final state is reached– Index the states of FSMs for efficient

evaluation


FSM Representation• Each element node is a state• A state is represented using a Path Node

structure (Contains information to process current state):– Compare the level of element name in input document

with the level value of the path node– Evaluate the element node filter if there is any– Locate next path nodes for the state change in the FSM

representation– Calculate the level values of next states using relative

distance values (in terms of levels) stored in the path nodes

– Not generated for wildcard (“*”) nodes


Path Node Decomposition/ a / * / b // c[@att1 = ‘500’] / d

Path Node 1Rel Dist = NALevel = 1

Path Node 2Rel Dist = 2Level = ?

Path Node 3Rel Dist = NALevel = AnyFilter Expression

Path Node 4Rel Dist = 1Level = ?

234565

PN2

PN3

PN4<a><x><b><y> <c att1 = 500><d/> </c>…

PN1Level = 1

El = a

PN1

1Level

PN2Level = 3

El = b PN3Level = Any

El = cFilter Expression

PN4Level = 6

El = d

PN2Level = 3



Level = 1El = a

PN1

PN2Level = 3



PN4Level = 6

El = d

Query issatisfied


Handling Multiple Queries

• Hash table based on the element names in the queries

• Each node contains two lists of path nodes:– Candidate List: Stores the path nodes that represent

current state of each query– Wait List: Stores the path nodes that represent the

future states• State transition is represented by promoting a path

node from the Wait List to the Candidate List• Initial distribution of path nodes has a significant

impact on performance

Key insight for scalable SDI:Index the queries instead of the data


Examples

Q1 = / a / b // c

Q11

NA1

Q121?

Q13

NA-1

Q1-1 Q1-2 Q1-3

Q2 = // b / * / c / d

Q21

NA-1

Q222?

Q231?

Q2-3Q2-2Q2-1

Q3 = / * / a / c // d Q4 = b / d / e Q5 = / a / * / * / c // e

Q33

NA-1

Q321?

Q31

NA2

Q3-3Q3-2Q3-1

Q51

NA1

Q5-1

Q523?

Q5-2

Q53

NA-1

Q5-3

Q41

NA-1

Q4-1

Q421?

Q4-2

Q431?

Q4-3

Query IdPositionRel DistLevel


Query Index Construction

z

a

b

c

d

e

WL

CLQ2-1

Q2-2

Q2-3

Q3-1

Q3-2

Q3-3

Element Hash Table

CL : Candidate ListWL: Wait List

WL

Q1-1

Q1-2

Q1-3

WL CL

WL

CL

CL

WL CL

Q4-1

Q4-2

Q4-3

Q5-1

Q5-2

Q5-3


Enhanced Algorithms

• Drawbacks of the “Basic” approach:– Query skew: hot elements are likely to have

very long Candidate Lists – Unnecessary evaluations of queries for

which the input document contains only a subset of the required element names

• Two enhancement strategies:– List Balance– Prefiltering


List Balance Algorithm

• When adding an FSM to the Query Index, select a “pivot” Path Node whose element has the shortest Candidate List length

• Treat the pivot node as the initial state of the FSM– Attach the portion of FSM that precedes the

pivot node as a prefix– Evaluate the prefix as a precondition by using

a stack of traversed element nodes in the XML document


FSMs in List BalanceQ1 = / a / b // c

Q11

NA1

Q121?

Q13

NA-1

Q1-1 Q1-2 Q1-3

Q2 = // b / * / c / d

Q21

NA-1

Q222?

Q231?

Q2-3Q2-2Q2-1

Q3 = / * / a / c // d

Q32

NA-1

Q311?a

XX

NA2

Q3-2Q3-1

Query IdPositionRel DistLevel

Q4 = b / d / e

XX

NA-1

Q4-1

Q421?

Q4-2

Q411?b

Q5 = / a / * / * / c // e

XX

NA1

XX3?

Q51

NA-1

a, c

Q5-1Prefix


Query Index in List Balance

CL : Candidate ListWL: Wait List

Q1-1

Q1-2

Q1-3

WL CL

WL CL

WL CL

WL CL

WL CL

Q2-1

Q2-2

Q2-3

Q3-1

Q3-2

Element Hash Table

Q4-1

Q4-2 Q5-1

z

a

b

c

d

e


Prefiltering• Implemented as an initial pass that is

performed before processing the queries• Based on Yan’s [Yan 94] Key Based

algorithm• Each input XML document is parsed twice

– In the first pass:• Match the element names for each query with the

document– In the second pass:

• Consider only the queries that passed the first step• Selectivity of the Prefiltering step

determines its benefit.


Nested Path Expressions• Element Node Filters may contain other

XPath queries• Nested query is treated like a separate

query • For relative execution, initial state of nested

query is activated after parent element node is satisfied.

• If result not available, assume true and “mark” for later re-evaluation.a b e

c dQ1 = / a // b[ c / d = 100] / e

Q2 Q2

Q1


Performance Evaluation

• Experimental Environment– NITF DTD is used to generate input documents and

queries (Contains 158 elements organized in 7 levels with 588 attributes)

– IBM’s XML Generator is used to create input documents

– We implemented a similar XPath query generator • Workload Parameters to Examine

– Scalability of the algorithms– Different document and query settings


Scalability Experiments

Uniform Dist.

0

500

1000

1500

2000

2500

3000

0 20 40 60 80 100 120Number of Profiles (x1,000)

Filte

r Tim

e (m

sec)

BasicPrefilter + BasicList BalancePrefilter + List Balance

(Max. Depth = 5, No Wildcards, No filters)

Skewed Dist.

0

2000

4000

6000

8000

10000

12000

14000

16000

18000

0 20 40 60 80 100 120Number of Profiles (x1,000)

Filte

r Tim

e (m

sec)



Document Depth Experiments

Uniform Dist.

0

500

1000

1500

2000

2500

0 2 4 6 8 10 12Maximum Depth

Filte

r Tim

e (m

sec)


Skewed Dist.

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

0 2 4 6 8 10 12Maximum Depth

Filte

r Tim

e (m

sec)


(# of Profiles = 50,000, No Wildcards, No filters)


Filter Experiments

0

200

400

600

800

1000

1200

1400

1600

1800

2000

0 1 2 3 4 5Element Node Filter Level

Filte

r Tim

e (m

sec)


0

200

400

600

800

1000

1200

1400

1600

1800

2000

1 10 100Element Node Filter Selectivity (%)

Filte

r Tim

e (m

sec)


(# of Profiles=50,000, Max. Depth = 6, No Wildcards)


XFilter Summary• XFilter was designed for scalable filtering

and content-based routing of XML documents– Sophisticated indexing mechanisms and a modified Finite

Sate Machine approach to filter XML documents efficiently

• XPath query language is used in the profile model to define expressive user profiles– Avoids overwhelming the users with irrelevant data

• Performance – XFilter is suitable for Internet-scale applications – Enhancements provide substantial improvements


Related Approaches• SIFT [Yan & Garcia-Molina, TODS 2000]

– Explored both Boolean and Similarity-based matching models (more of an IR perspective).

– Also examined aspects of distributed filtering.• Database Continuous Queries

– Xerox Parc [Terry et al., SIGMOD 92]– NiagraCQ (Wisconsin) [Chen et al., SIGMOD 00]– OpenCQ (OGI/Georgia Tech) [Liu et al., TKDE 99]– Employ database queries and grouping (common-

subexpression) approaches.


Related Approaches(continued)• Trigger Systems for “Active Databases”

– [Widom & Finklestein, SIGMOD 90]– [Stonebraker et al., SIGMOD 90]– Not focused on filtering and routing per se, so

more general, complex and less scalable.• More recent work on triggers has focused

on scalability [Hanson et al., ICDE 99] • Change Detection in Semi-structured data

[Chawathe et al, ICDE 98]• Interesting issues are a) how much functionality

is needed and b) how much history is used.


Approaches in ContextAmount of Structure

Low High

FilterInputFilter

Method

Plain TextDocuments

IR-BasedMethods

XMLDocuments

DatabaseTuples

XFilter ContinuousQueries (CQ)

Usage of DBTechniques

Low High


Summary of Data Dissemination

• Described the DBIS architecture for deploying dissemination-based applications.– User profiles play a key role here.

• XFilter is an information filtering/routing system aimed at very large-scale systems.

• Similar problems (to varying degrees of scale) have been addressed in databases in the context of continuous queries and active databases.– Xfilter can exploit these approaches, esp. for

common sub-expr., and history-based filtering.


4. Synchronization

• Question: why/when is synchronization needed?

– i.e., What is wrong with good old ACID transaction models?


Why Synchronize?• Needed primarily due to disconnection.

– If always connected, then could something stricter.

– Efficiency concerns and long-running transactions also may be motivations.

• The basic idea:– Clients cache secondary copies of data– Servers retain “copies of record”– Updates happen without two-phase commit.– Synchronization process attempts to make

these mutually consistent.• Upates on server sent to client, and vice-

versa• Run conflict resolution when a problem arises


Synchronization Topics

1. Device Synchronization• PalmOS HotSync• “Edison” Database-supported extensions• SyncML Industry Standard

2. Data Recharging

3. Consistency in Weakly Connected Environments


Palm HotSync Background• Metadata kept both at Device and on the Desktop.• Data on device is stored as records in PalmDBs.

– Each PalmDB is associated with an application– Each record has a set of status bits.

• Indicate if record has been created, modified, or deleted since last synchronization.

• The HotSync Manager runs on the desktop machine– Contains conduits, downloadable code for synchronizing with

specific applications.• Desktop maintains it’s own copies of the palmDBs,

including it’s own versions of the status bits.– Also maintains a snapshot of each palmDB taken immediately after

most recent synchronization.


HotSync Protocol• Device initiates synchronization protocol.

Can run in one of two modes:• Fast Sync

– Was device last synced with this desktop?– If so, then hand held sends data and status only for those

records whose status bits are set.– Conduit can do efficient comparison of bits, update its

copy of palmDB and send updates to the device.• Slow Sync

– Else, can’t compare bits – device sends entire palmDB to the conduit, which does a field by field comparison to figure out what changed.


Protocol (continued)• By comparing status bits (and possibly

palmDB snapshots) the synchronization logic determines what actions to perform.

• Examples:– Created at desktop send to device.– Deleted at device, not changed on desktop

delete from desktop.– Updated on desktop, not on device send to

device.– Updated on both raise an exception (invoke

manual conflict resolution).


Problems with HotSync• Can’t Sync with desktops not previously

configured for syncing with that device and applications.– some limited solutions are emerging for this, but restricted by

(lack of) reliability of the desktop.• Large performance penalty for synchronizing

with multiple desktops (home/office problem).– can be a problem at server too, if it has to handle lots of

devices.• No support for synchronizing with multi-user

concurrent data sources.– No notion of “interest” in a subset of the records in a database.– Many slow syncs and lots of snapshots to keep


Database-Supported Sync• The EDISON project at UC Berkeley is addressing

these shortcomings with an ORDB-backed server.– driven by Matt Denny, [Denny & Franklin, ?? 01]

• Server maintains shared data and synchronization metadata for all devices.

• Users interests in subsets of shared data sources are expressed using predicates (a.k.a. profiles?).

• Utilizes proxies (“sync nodes”) running on network access points.– Sync nodes communicate with the server to obtain necessary

metadata and data records.– They also execute synchronization logic.


EDISON (continued)• Eliminates the one snapshot per

device/data source pair by:– Incremental logging of metadata at the data source – Either push or pull network protocols to get the

correct updates to the synchronization points• Devices always use Fast Sync

– Use of shared server regardless of access point used results in always consistent metadata.

• Caching techniques can be used to stage metadata at sync nodes, but not needed for even fairly large systems (1000’s of devices).


EDISON Topology

• Device (HH): contains its own sync metadata & data• Sync Node (SN): sync access point on network,

caches sync vectors (SV) containing metadata• Data Source (DS): data repository and metadata log • A variant of the dissemination architecture

discussed earlier!

DS

SN SN SN

HH HH

network SNSV SV

…

DBID PalmID SyncInfo


EDISON Protocol (1st Phase)

• 1st phase: Compare HH data to SV1. HH initiates sync by connecting to an SN.2. SN sends INIT message to DS3. DS sends its copy of the SV and all the updates

which occurred since last sync for this HH4. Gets all modified records from the HH, and

creates an appropriate action to send to the DS • Uses Palm’s synchronization logic, but instead

of acting directly on the data, Edison creates actions for the DS

HH SN DS4

2

3

1


EDISON (Phase 2)

• Phase 2: Commit changes to data to DS1. SN batches actions into 1 network message and

sends to DS2. The DS performs the following as one transaction:

• Logs metadata changes for all sync vectors interested in same data

• Sends updated data values back to sender SN• Could also send updates to other “interested”

SNs if caching was being done.

HH SN DS1

2


EDISON (Phase 3)

• Phase 3: Apply changes from DS to HH1. Once SN gets updated values, SN applies

updates to HH and SN’s sync vector as needed; 2. sends DS an END message so that DS can clean

up logs

HH SN DS1 2


EDISON Summary• Initial Performance Studies show that system

scales quite well.– Assuming synch frequencies of several times a day, server

is very lightly loaded even with many thousands of clients.– Amount of data transferred over the wire is small.

• In future, synching may be much more frequent and devices will hold more and more data.

• In such cases, protocol can be extended to cache metadata and data at Sync Nodes– this offloads data access from servers, as well as

synchronization processing.


SyncML Standard• Industry Consortium with most major players:

Ericsson, Nokia, Motorola, Palm, Psion, IBM, …• Goal is to enable cross-format, cross-system

synchronization.• Simple architecture:

– Client: PDA, Phone or PC; intermittently connected.– Server: typically PC or Server; continuously available.

• Consists of a standard set of message types, each represented as an XML document.

• Supports different interaction models including “request/response” and “blind push”


SyncML Sync Types1. Two-way – “normal (fast) sync”, client sends first.2. Slow-sync – client sends all data3. One-way, client only – client sends only modified

records to server; server does not send to client4. Refresh, client only – client sends entire DB

to server5. One-way, server only6. Refresh, server only7. Server Alerted – Sync initiated by server (push?)


SyncML (continued)

• Standard requires servers to maintain mappings between its own record IDs and the IDs of records as kept by the client.

• Conflict Resolution logic is (of course) dealt with abstractly by the standard. It provides standard status codes that can be used to implement typical policies.

• Contains support for authentication of clients and servers.

• www.syncml.org


“Data Recharging”• An alternative approach to data

synchronization…

• A merger of dissemination and synchronization approaches.

• Joint work w/ Mitch Cherniack and Stan Zdonik as part of the Data Centers project.


Data Recharging - Motivation• Mobile devices require 2 resources: power and

data– It is impractical to be continuously connected to

fixed sources of these.• Devices cope with disconnection using caching:

– Power cached in rechargeable batteries– Data cached in hot-synched memory

• Recharging the power is easy…– Anywhere, Anytime, “Hands-off” operation, Flexible

connection duration• Recharging the data, well, we just covered that.


Data Recharging (continued)“Make recharging data as simple as recharging power”

• Anywhere – no need to connect to your home machine,

• Anytime – no prior arrangements necessary,

• “Hands-off” operation – system knows what you need

• Flexible connection duration – the longer you stay connected, the better your device-resident data gets.


Some Questions

• How to know where the user will be?– and do we care?

(for context – yes, for staging -??)• How to know what the user wants?• How to prioritize data delivery?

• The answer is User Profiles


“Data Recharging” Profiles• Recall, the three main components:

1) Content-based specifications of user interests(read “queries”)

2) Specifications of user priorities/requirements,priority ordering, resolution, freshness,

dependencies

3) User Context information – where, when, who, what

This info is available in the user’s PIM data!


First cut at Profile Model• Items of Interest Defined Explicitly (URLs)

– Dependencies and alternatives expressed in a tree– “Values” assigned to individual items

• Tree is built with special operators– Choose (n) – Value obtained for up to any n children– First (n) – Value obtained for up to n children in order

(e.g., for progressive resolution).– And – Value obtained only if all children are delivered.

• Total value of a “data charge” can be computed bottom up using simple formulas.

• (based on M.S. work by Danny Tom @ UC Berkeley)


Profile Example

Choose(2)

First (2) First (2)

First (2)

ANDChoose(1)

Time of Quote

Addendumto

News Story

Stock Name Current Price

News Story

Graph(low res)

Graph(hi res)


Exploiting Profiles• Need to use profiles to choose contents of a data

charge (not just evaluate them)• Want to maximize value delivered in a charge

without having to spend too much time choosing.• Two optimization problems:

– Bounded (known) sync time– Unknown sync time

• Bounded case is an instance of the “precedence-constrained knapsack problem”

• Can be implemented using approximations or various types of heuristics.

• Initial results indicate that approximations of the PCKP approach work best.


On-going Profile Work

• Current work on recharging profiles has taken on more of a database query processing approach.

• The idea is to separate the specification of interests from the calculation of “utility”.

• Like database query languages, these profile languages should be declarative.

• Then, calculating the contents of a charge becomes more like a query optimization and execution proceedure.

• Watch this space for more details…


Recharging - Research Agenda• Profile Definition and Maintenance• Update Storage and Preparation• Efficient integration of "recharge" updates with

existing cached data.– Recharge, Trickle Charge, Jump Start...

• Consistency Guarantees• Global Data Staging

• More generally, Data Recharging blurs the line between synchronization and dissemination, can it be used for both?– How to exploit improved connectivity?


Data Caching and Consistency

• Synchronization in Peer-to-peer environments is more complicated than in the less symmetric PDA-based approaches.

• Centralized algorithms require connectivity at specific times.

• Alternative: Epidemic Algorithms • Conflict detection: timestamps, version vectors,…

– Conflict Handling (update commitment):• Optimistic (resolution) - Manual except in limited domains,• Pessimistic (avoidance) - primary copy, write-all or

voting-based.• Previous work: Bayou, Ficus, Coda, …


Epidemic Protocol Illustration

(Picture is by way of Ugur Cetintemel)


Deno - Cetintemel and Keleher

Pessimistic, Asynchronous (epidemic), voting-based“Bounded” weighted-voting:

– Each replica is assigned a currency ci s.t. 0 ci 1.0– Total currency in the system is bounded, i.e., ci=1.0– Currency can be re-distributed for optimization or planned

disconnection.An update’s life:

– Sites issue tentative updates– Updates and votes are propagated in a pair-wise fashion– Updates gather votes as they pass through sites– An update commits when it gathers plurality of votes


Decentralized Commitment• An update u wins an election

with plurality• A site s maintains:

– votes(u): the sum of votes u gained so far

– unknown: the sum of votes unknown to s

(i.e., 1.0 – votes(u), for u)• u commits iff for all u’ <> u,

votes(u) > votes(u') + unknown

Issues: time to commit; abort rates

s1 Oi

(s1, 0.20, u1)

votes(u1) = 0.20unknown = 0.80

(s1, 0.20, u1)(s5, 0.20, u1)

votes(u1) = 0.40

unknown = 0.60

(s1, 0.20, u1)(s5, 0.20, u1)(s6, 0.15, u2)

votes(u1) = 0.40votes(u2) = 0.15unknown = 0.45

(s1, 0.20, u1)(s5, 0.20, u1)(s6, 0.15, u2)(s2, 0.15, u1)votes(u1) = 0.55votes(u2) = 0.15unknown = 0.30

u1 commits!

s1 Oi

(s1, 0.20, u1)

votes(u1) = 0.20

unknown = 0.80

(s1, 0.20, u1)

(s4, 0.20, u2)votes(u1) = 0.20votes(u2) = 0.20

unknown = 0.60

(s1, 0.20, u1)

(s4, 0.20, u2)(s6, 0.25, u3)votes(u1) = 0.20votes(u2) = 0.20votes(u3) = 0.25unknown = 0.35

(s1, 0.20, u1)

(s4, 0.20, u2)(s6, 0.25, u3)(s2, 0.25, u2)

votes(u1) = 0.20votes(u2) = 0.45votes(u3) = 0.25unknown = 0.10u2 commits!


Wrap Up• Data Dissemination and Synchronization are indeed,

closely related.• A common set of architectural concepts can and

should be used.– These can be deployed as an overlay network.

• The key is to tackle these problems as data management issues, not only as networking problems.

• At the heart of all of these systems is an expressive, highly-functional user profile management system.

Profiling languages and evaluation algorithms will enable the next generation of data intensive applications.


Acknowledgements

• Mehmet Altinel – XFilter, DBIS Toolkit• Ugur Cetintemel – Deno• Mitch Cherniack – Data Recharging• Matt Denny – EDISON, Data Recharging • Pete Keleher - Deno• Danny Tom – Data Recharging• Stan Zdonik – DBIS and Data Recharging

Dissemination and Synchronization for Mobility (and Beyond)

Documents

franklin mdm

data processingcommunications

base data

data semantics

data disseminationdisseminate

data management context

exploits data centers

representation of data