a distributed feature composition a a and · 2009-07-01 · distributed feature composition is...
Post on 13-Mar-2020
5 Views
Preview:
TRANSCRIPT
A
AA
A
A
AA
A A
A A
A
MODULARITY IN
DISTRIBUTED FEATURE COMPOSITION
and
THIRTEEN YEARS OF LEARNING FROM MICHAEL JACKSON
Pamela Zave
AT&T Laboratories—Research
Florham Park, New Jersey USA
A
AA
A
A
AA
A A
A A
A
SOME THINGS I LEARNED FROM MICHAEL
WHAT TO SAY . . .
. . . after asking the staff of a restaurantto turn the music down, and being told,"We can't, the customers like it."
HOW TO MAKE A LEFT TURN, WHILEDRIVING IN NORTH AMERICA
WHY THE WORD "MULTIPARADIGM"IS BAD
"Multi paradigm" is a chimera.
an imaginarymonster compoundedof incongruous parts
Start using "polyparadigm".
(possibilities too numerous to list)
Latinorigin
Greekorigin
(despite his many efforts topersuade us, no one at AT&TResearch has ever had the
courage to attempt this)
you are here
path ofother carmaking left turn
your path gives you agood view ofopposingtraffic at alltimes
A
AA
A
A
AA
A A
A A
A
features are being added andchanged continually
are optional for each subscriber
can often be enabled/disableddynamically by their subscribers
Call Waiting
FEATURES INTERACT WHEN ONEFEATURE MODIFIES OR INFLUENCESANOTHER
FEATURE INTERACTIONS ARE VERYCOMMON
all features are modifying orenhancing the same basic service,which is real-time communicationamong people
the number of interactions iscombinatorial in the number offeatures
TELECOMMUNICATION SERVICES . . .
TransferConference
. . . ARE CONCEIVED AND BUILT INTERMS OF FEATURES SUCH AS:
FEATURE CHURN
FEATURE CHURN AND FEATUREINTERACTIONS HAVE CAUSEDSEVERE SOFTWARE PROBLEMS INTHE CIRCUIT-SWITCHED TELEPHONENETWORK
(ever since it becamesoftware-controlled in the 1960s)
Parallel RingingQuiet TimeVoice Mail
Speed DialingOutgoing Call BlockingVoice Dialing
A
AA
A
A
AA
A A
A A
A
usage: a dynamically assembled graphof boxes and internal calls
DISTRIBUTED FEATURE COMPOSITION (DFC) . . .
IB IBFB FB FB
box: a concurrent process, providingeither interface or feature functions
internal call: a featureless, point-to-point connection with a two-waysignaling channel and any number ofmedia channels
system boundary
router DFC router: routesinternal calls to boxes
featuredata
feature data:persistent, usuallypartitioned by feature,interface to Web services
transparency
autonomy
context independence
. . . IS AN ADAPTATION OF THE PIPES-AND-FILTERS ARCHITECTURE TO TELECOMMUNICATION SERVICES
BECAUSE A FEATURE BOX HAS . . . . . . THE ARCHITECTURE SUPPORTS
feature modularity
structured feature composition
management of feature interactions
A
AA
A
A
AA
A A
A A
A
signals to callee
avail, unavail from callee
SIGNALING INTERACTIONS
VM
VM
caller chooses whetherto interrupt targetedsubscriber
caller records voicemail in subscriber'smailbox
if no, QT sendsunavail upstream
if yes, QT continues the chainand becomes transparent
VM PR
if single branch succeeds,send avail upstream
if whole attempt times out,send unavail upstream
absorbsfailuresof singlebranches
QT
QT
QuietTime
VoiceMail
Para-llel
Ring
each feature works correctlyand makes sense by itself
one feature can trigger or cancelanother, even though neitherknows that the other is present
A
AA
A
A
AA
A A
A A
A
CW
CW
CW
CW CW
CW CW
CW
CW
ROUTING INTERACTIONS
MCM
MCMMCM
4
MCMMCM1
2
MCM
3
MCM
MCMMCM
Mid-Call Move Call Waiting
no feature box evermanages more thanthree internal calls
dynamic reconfiguration isperformedautomatically by theDFC routingalgorithm
preserves theinvariant that everyperson has hisfeatures in everyrelevant path, in thecorrect order
in general, asignaling path consists of segmentsset up in alternatingdirections
A
AA
A
A
AA
A A
A A
A
AT&T CALLVANTAGE SERVICESM
. . . is a consumer, broadband, voice-over-IP service.Its advanced features were built with DFC.
supported many thousands of customersworld-wide (without two media-intensivefeatures)
easy feature evolution
the service won two industry awards,citing voice quality and advanced features
this is unprecedented speedDEPLOYMENT AND EVOLUTION
FEATURE DEVELOPMENT
a group of researchersdelivered eleven complexfeatures . . .
Mid-Call MoveTen-Way CallingParallel Find Me
. . . two months from theinception of the project
SYSTEM INTEGRATION
all the feature interactionswere successfully analyzedand managed during thesame two months
many integration problems with vendor-supplied components (IP routers,gateways, phone adaptors, media servers)
DFC modularity was extremely useful foradding adaptors to patch over integrationproblems
don't want to embed thesein the service
2003-2004
e.g.,
A
AA
A
A
AA
A A
A A
A
T-MEETING
. . . is a teleconferencing system for internal use at AT&T.It was built entirely with DFC.
PHASE ONE (2006) PHASE TWO (ongoing)
has mid-conference control fromboth phone and Web, recording,active speaker identification, userswitching between multipleconferences
during development, modularity supported functional prototyping,code reuse, deferred design decisions
as with CallVantage, there were veryfew bugs in the application code
supports millions of user minuteson a typical workday
a re-implementation with our newstandards-based software tools
A
AA
A
A
AA
A A
A A
A
SIP application server(running in endpoint or network)
StratoSIP(SIP at a very
high level)
ECharts forSIP Servlets
(E4SS)
compositionalmedia control
(dist. algorithm)
compilation
inclusion
JSR 289 SIP Servlet container
SIPapplications
in Java
SOME OF OUR NEW STANDARDS-BASED TOOLS
DFC router
Webapplications
SIP IS THE DOMINANT SIGNALINGPROTOCOL FOR IP-BASEDMULTIMEDIA APPLICATIONS
THESE TOOLS IMPLEMENTDFC ON TOP OF SIP
this language for boxprogramming is basedon Boxtalk
A
AA
A
A
AA
A A
A A
A
DISTRIBUTED FEATURE COMPOSITION REALLY WORKS
WITHIN THE DOMAIN OFTELECOMMUNICATION SERVICES,MODULARITY IS GENERAL-PURPOSE
feature modularity: a feature is anaddition or exception to the basicservice
module is an addition or exception toanother feature
modularity isolates probable change
module is a unit of re-usable code
module is an adaptor
an off-the-shelf component is packaged as a module for betterintegration
modularity opens services to applications from the public
the original purpose
MANAGEMENT OF FEATURE INTERACTIONS IS MORE THANANALYSIS
studying each class of interactiontends to yield deep domainknowledge . . .
. . . of how and why features interact in that way
. . . of which interactions are desirable and which are undesirable
. . . of how features should be managed to prevent the bad interactions and enable the good ones
without the architecture,there would be no limits
on how features could interact
now 25,000iPhone appsin the store
A
AA
A
A
AA
A A
A A
A
AFTER TWELVE YEARS,
DISTRIBUTED FEATURE COMPOSITION
IS STILL AHEAD OF ITS TIME
It is a challenge to explain DFC modularity, composition, and abstractions to the SIP community, . . .
. . . although the need for them is urgent.
The DFC routing algorithm fills anunmet need in the Internet architecture.
Being rich in functions, it may be asgeneral as will ever be needed.
Being expensive to implement in somecontexts, there are niches for otherversions with fewer functions.
A
AA
A
A
AA
A A
A A
A
HOW TO THINK LIKE MICHAEL JACKSON
?
top related