DTIC ADA459468: Comparison of Floor Control Protocols for Collaborative Multimedia Environments

13 Pages·0.4 MB·English

by  Defense Technical Information Center

Comparison of (cid:13)oor control protocols
for collaborative multimedia environments
Hans-Peter Dommel and J. J. Garcia-Luna-Aceves
Baskin School of Engineering
Computer Engineering Department
University of California, Santa Cruz, CA 95064, USA
ABSTRACT
Improvementsin networkingallow for increasingly complex collaborationenvironments with regardto session scale,
rangeof shared tasks, and distance between remote parties. Floor control protocolsadd an access discipline to such
environments that allows to mitigate race conditions on shared resources and throttle media transmission. Primary
causes for resource competition among users may be the lack of mutual awarenessand formal session orchestration,
or network and host limitations.
Various, often proprietary and unscalable solutions for (cid:13)oor control have been implemented for telemedicine,
video conferencing, or distributed interactive simulation. To this date, an analytic comparison of the e(cid:14)cacy of
these solutionsislacking. With e(cid:14)cacy,wemeantheproportionoftime that aprotocoltakestoallocatearesource,
accountingforsocialandtechnicaloverheadfromuserbehavior,protocolcost,andnetworkconditions. Wepresenta
noveltaxonomyand comparativeperformance analysisof known classesof (cid:13)oor control protocols,including socially
drivenprotocols,collisionsensingonsharedresources,(cid:13)oortokenpassinginfully-connectedandringtopologies,and,
innovatively, across shared control trees. Accordingly, aggregated and selective transmission of control information
over a multicast control tree o(cid:11)ers the best scalability and e(cid:14)cacy. A novel hierarchical (cid:13)oor control protocol
correlating in its operation with tree-based reliable multicast is outlined.
Keywords: Floor Control, Multimedia Collaboration, Reliable Multicast
1. INTRODUCTION
1
With IP-multicasting more powerful collaborative multimedia applications (CMA) gradually enter mainstream
computing. Users of such applications can overcome their separation in time and space and share work e(cid:11)orts in
real-time,withthegoalofapproximatingthequalityofface-to-faceinteractions. WhileearlierCMAwereproprietary,
monolithic, and limited in media modality and session size, new applications support multi-party and multimodal
collaboration in large sessions, such as distributed interactive simulations, distance learning seminars, or special-
2
purpose MBonesessions. However,comparedto advancesin reliablemulticasting and multicast routing, there has
been little progress with regardto group coordination support for such systems.
3
Floor control is an access discipline for CMA, which may solve (cid:13)awed telepresence and coordination problems
4
as reported by Isaacs and Tang in studies on video conferencing. Typically deployed in the session or application
layer, (cid:13)oor control lets users attain exclusive control over a shared resource by attaining a (cid:13)oor, which is a short-
lived synchronization primitive for multimedia objects. The (cid:13)oor semantics is generalized to multimedia from its
5
traditional notion as the \right to speak". Deployment of (cid:13)oor control in CMA may be complex due to system
capabilities, in terms of the network, host, and communication subsystem, as well as user behavior. Early work on
6
(cid:13)oor controlled systems has many di(cid:11)erent faces, including text-based remote collaboration, electronic meeting
7 8 2 9
support, distributed teleconferencing, moderation of MBone seminars, or web-centric groupwork. Users may
pro(cid:12)t from (cid:13)oor control, because it de(cid:12)nes turn-taking rules, fosters interactivity, and prohibits unfairness. From a
system perspective, it can providequality-of-serviceinput regardingadmission control and bandwidth allocation for
10 11
bulky media streams, or serve as concurrency control for synchronization of multiple media (cid:13)ows.
Furtherauthorinformation: (Sendcorrespondence toH.-P.D.)
H.-P.D.: E-mail: [email protected]
J.J.G.-L.-A.: E-mail: [email protected]
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
1998 2. REPORT TYPE 00-00-1998 to 00-00-1998
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Comparison of floor control protocols for collaborative multimedia
5b. GRANT NUMBER
environments
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
University of California at Santa Cruz,Department of Computer REPORT NUMBER
Engineering,Santa Cruz,CA,95064
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
The original document contains color images.
14. ABSTRACT
15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON

a. REPORT b. ABSTRACT c. THIS PAGE 12
unclassified unclassified unclassified
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
12
A(cid:13)oorcontrolmethodologyischaracterizedbyseveraldichotomies,centralizedvs. distributedimplementation,
13,14 13,15
mechanism vs. policy, and explicit vs. implicit (cid:13)oor hand-over. In centralized systems, one host controls
(cid:13)oorassignmentforanentiresession,whichiseasiertodeploy,but su(cid:11)ers fromoverloadand resilienceproblems. In
distributed systems, eachhostin asessioncontributestothe globalcontrol processas much asits own resourcesare
concerned. Consistencymanagementande(cid:14)ciencyofcommunicationaremajorconcernswiththisapproach. Hybrid
solutions take the best of both worlds and centralize tracking of single (cid:13)oors, while distributing the administrative
loadoverthehostsinacollaborativesession. Hybridsolutionsarealsowell-suitedfornetworkcomputers,wherehosts
havelimitedcapabilitiesandcommunicatewitheachotherviasessionservers. Amechanism,e.g.,activitysensingor
token passing, is the physical propagation and synchronization apparatus for control information. It is instantiated
withpoliciesregardingserviceorder,priorities,orfairness, suchas\free-for-all",\moderated",\prioritized",\voice-
activated", or \high-resolution". Policies allow to adjust (cid:13)oor control to the session style, e.g., a lecture, panel, or
laboratory. The service order of queued (cid:13)oor requests, or the signaling on how the (cid:13)oor is attained has also been
13
considered a policy. In explicit (cid:13)oor passing, a user must submit a signal speci(cid:12)cally to attain the (cid:13)oor, e.g., by
raising the hand or pressing a button, in contrast to implicit passing, as it is the case with voice activation.
Despitethisbackgroundofworkon(cid:13)oorcontrol,adetailedanalysisontheoperationalprinciplesandperformance
of (cid:13)oorcontrol protocolsis still missing. This paper presents a novel taxonomyand e(cid:14)cacy analysisof (cid:13)oorcontrol
protocols, based on their operational principles. Hierarchical (cid:13)oor control is described, operating on a control tree,
whichmatchesthe backbonereliablemulticasttreeandmulticastroutingtree. Thegoalistoshowhowmechanisms
for group coordination, that is (cid:13)oor control, (cid:12)t in with the emerging large-scaleInternet conferencing.
In Section 2 we present our system model and taxonomy, as the foundation for our comparativee(cid:14)cacyanalysis
presentedinSection3. Section4outlinestheoperationofatree-based(cid:13)oorcontrolprotocol. Thepaperisconcluded
in Section 5.
2. SYSTEM MODEL AND TAXONOMY
Wede(cid:12)neimportanttermstolaythefoundationforacommonmethodologyandtaxonomyof(cid:13)oorcontrolprotocols.
AcollaborationenvironmentisatupleCE=(S;U;R;F)consistingofhostsV connectedbynetworklinksE (cid:26)V(cid:2)V,
users U, shared resources R and (cid:13)oors F. Communication among hosts is solely via message exchange, and not in
shared memory. Links are assumed to be reliable for (cid:13)oor control messages. A collaborative session is a tuple CS
=(sid;(cid:1);U;R;F),instantiatingCE,andischaracterizedbyauniquesessionidenti(cid:12)ersid,sessionduration(cid:1),users
U,resourcesRandtheircorresponding(cid:13)oorsF. Thesessionsizem=jUjcanvary,dependingonwhetherthesession
is open or closed, and in our terms m > 100 denotes a \large" session. Sessions may have di(cid:11)erent organizational
styles,suchaslecture,businessmeeting, paneldiscussion,or hearing. Users u2U, whichmaybe humansorsystem
agents, can assume the control rolesof (cid:13)oor originator (FO), owning a resource, (cid:13)oor coordinator (FC), moderating
the (cid:13)oor of a resource, or (cid:13)oor holder (FH), having exclusive access to a resource. Depending on the session style,
FH and FO may be identical. Shared resources r 2 R, e.g., a robotic device, surgical instrument, video and audio
channel, or graphical data object in the user interface, can be replicated or located at a speci(cid:12)c host. Finally, we
de(cid:12)ne a (cid:13)oor as a tuple
f(r;T;FC;FH;ta;(cid:14);QoS)=st2F;
where r denotes the resource, T is the media type of r (video, audio, graphics, text), FC and FH identify the (cid:13)oor
coordinator and holder, ta is a timestamp marking the start of the (cid:13)oor holding time (cid:14), and QoS is a quality-of-
servicedirectiveindicatingdeliverypropertiesforr. The(cid:13)oorstatest=[free jbusy j idlejrequestedj nil]records
the current state of (cid:13)oor f at a host. For simplicity, we assume that there is one user per host, and one (cid:13)oor f per
resource r, without (cid:12)ner granularity.
Floor control protocols rest on three operational pillars: (1) the random or scheduled access to resources, which
is characterized by the mechanism and host topology, relegating directives among hosts; (2) the centralized or
distributed control over (cid:13)oors in the session; and (3) the (cid:13)oor policies established among hosts. These properties
are re(cid:13)ected in the taxonomyin Figure 1 (only paradigms indicated with solid lines are analyticallycompared.) We
divide known (cid:13)oor control paradigms in two classes:
Random-access Group Coordination (RGC) lets users contend for a shared resource, either mediated by social
protocols,orbyremotelysensingitsstatusbeforeorduringresourceaccess. Sensingisaccomplishedeitherbyusers,
tracking each other’s activities through the user interface, or by the system, sensing the state of an application,
host, or network for local and remote activities. The global (cid:13)oor state is marked with assertions on local variables
and notokenentity is explicitly exchangedasa placeholder. RGC schemesareinherently contention-based,because
Group Coordination
Random-Access Scheduled
Social Activity Token Polling Reservation
Mediation Sensing Passing
(RAS)
Incoordi- with Directly Ring- Tree- Implicit
nation Feedback Connected based based
(RSI) (RSF) (STD) (STR) (STT)
Figure 1. Taxonomy of group coordination.
hosts must actively compete for a (cid:13)oor. Remote sensing can be costly and ine(cid:14)cient. In RGC, (cid:13)oor acquisition is
characterized by \sensing", \busy", \active", and \idle" states.
ScheduledGroupCoordination(SGC)uses(cid:13)oortokenpassing,reservationorpollingforpendingcontroldirectives
as resourceaccess mechanism. The token is a unique placeholder, which is used to request, deny, reserve,or granta
(cid:13)oor. Regulated (cid:13)oortoken capturedispersesraceconditions on resources. Hosts can \ask"for the (cid:13)oor, ora token
circulates among hosts and is \o(cid:11)ered" to hosts. Token tracking and ensuring authenticity and consistency can be
costly. SGC schemes typically operate on a logical host topology such as a ring or tree. In SGC, (cid:13)oor acquisition
is characterized by \request", \deny", \grant", and \release"control messages. Predominant session geometries are
shown in Figure 2. Resource ri in the shared workspaceis hosted by user ui.
u1 r1 u1 r1 u1 r1 u3 r3
u1 u2 u3 u4
u2 u4 u2 u4 u2 u4 u1 r1
r2 r4 r2 r4 r2 FC r4
r
(FC) u3 r3 u3 r3 u3 r3 u2r2 u4r4
(a) (b) (c) (d) (e)
Figure 2. Session geometries.
In Fig. 2(a), users ui attempt to gain control of a centralized resource r with limited or no coordination. In
Fig.2(b), participantsui communicatedirectlywith oneanothertoattain the(cid:13)oorin the mostdirect manner. This
schemeisalsosuitedforachievingdistributedconsensus,e.g.,withvoting. Fig.2(c)depictsring-basedcontrol,based
on the idea of passing a (cid:13)oor token in a linear fashion along the hosts in the ring. In Fig. 2(d), a star-topology is
used to concentrate (cid:13)oor requests onto a dedicated FC host, which is a subcase of Fig. 2(e). In this tree structure,
control directives are propagatedalong the branches of a multi-level control tree towards the current FH.
3. EFFICACY OF FLOOR CONTROL PROTOCOLS
The following comparative analysis of known classes of (cid:13)oor control protocols is a (cid:12)rst attempt to characterize the
e(cid:14)cacy of interactive behavior of people and processes from a resource contention perspective. The intention is not
to predict exactly how a protocol performs for a given CE; accomplishing this would require far more host- and
network-speci(cid:12)cdetailsandstatisticsonuserbehavior. Ourgoalissimplytoassesstheoverheadofvariousprotocols
16
with regard to control state management. The basic methodology is derived from multiaccess communication,
based on the analogy between access mitigation for the data channel and shared resources.
We state the following assumptions to make the analysis tractable: the individual processing cost for control
packets, including protocol overhead and user-interface speci(cid:12)cs, is the same for all hosts; control message delivery
between hosts is reliable and no failures in hosts or the network occur; information from a host reaches any other
host with the same average system-wide propagation delay; and the interarrival rate of (cid:13)oor requests is Poisson,
given that there is no indication for cross-correlations between subsequent (cid:13)oor requests. We take the interarrival
rate of (cid:13)oor requests, the task length, and the network propagation delay into account. An important aspect of our
comparative analysis is the impact that multicasting at the network layer has on the e(cid:14)cacy of the (cid:13)oor control
protocols. When network multicasting is not available, the user hosts are forced to contact one another explicitly,
which substantiallyincreasesthe processingoverheadat hostsandthe possibilityof disagreementonwhich host has
the (cid:13)oor. We model the existence of multicasting at the network layer by assuming that a user host needs only to
pass information once to the network (during an activity period and to gain the (cid:13)oor) for the information to reach
all other hosts with the same averagedelay. The notation used in our analysis is summarized in Table 1.
(cid:12)1 average\think" timebefore(cid:13)oortoken arrival
(cid:12)2 average\think" timeat(cid:13)oortokenpresence
(cid:13) averageprocessingtimefora(cid:13)oordirective
(cid:14) durationofaverageactivityperiod
n(cid:15) processingandunicastingoverheadfornhosts
(cid:17) e(cid:14)cacyofa(cid:13)oorcontrolprotocol
G averageo(cid:11)ered(cid:13)oorrequestload
(cid:19) averagedurationofidletime
(cid:21) (cid:13)oorrequestinterarrivalrate
m averagenumberofhostsinsession
n averagenumberofactivehostsinsession
(cid:23) averagevulnerabilityperiod
(cid:28) averagepropagationdelay
Table 1. Analysis parameters.
The \think time", denoted with (cid:12)1 (the time until the expected (cid:13)oor arrives at the local host) and (cid:12)2 (the time
once the (cid:13)oor token is present at the local host and o(cid:11)ered to the user), re(cid:13)ects user choices to grab the (cid:13)oor in
tokenrings;(cid:13) istheaveragetimetoprocessandcommunicatea(cid:13)oordirective,includingpacketizationdelay,generic
queueing transmission times, and local processing overhead; (cid:14) is the duration of an activity period; we use n(cid:15) > (cid:28)
as the additional delay to account for the communication and processing overhead in n actively collaborating hosts;
G = (cid:14) (cid:2)(cid:21) is the normalized o(cid:11)ered request load on (cid:13)oors, including new and previously denied and resubmitted
(cid:13)oor requests; (cid:19) sums up the expected idle time for a resource during and after (cid:13)oor holding time; (cid:21) represents the
relative frequency of demand for resource access and thus indicates the contention level; m < 1 denotes all hosts
in a collaborative session; n < m is the number of active hosts in the multicast group for the current (cid:13)oor and
transmission; (cid:23) is the time during which a host’s attempt to access a resource can be intercepted by another host;
and (cid:28) denotes the averageend-to-end delaybetween hosts, coalescingmultiple routinghops into one (a packetmust
hence traverse on the average the same number of hosts on the path from the sender to a group of receivers.) The
activity period and time to process and communicate (cid:13)oor directives are assumed to include the time incurred in
providing feedback among hosts.
contention activity idle
period period period
Resource / X A1 I X A2 I X A3 ...
Floor g d ne t i g d ne t i g d
busy period B
turn period T
Host 1 / X A1
FC
STATE REQUEST GRANT GRANT
X A2
Host 2
REQUEST DENY REQUEST
X X A3
Host 3
Figure 3. Turn taking periods for a resource and three hosts.
A turn taking model re(cid:13)ecting the switching of control over a resource by users serves as the blueprint for our
analysis. A typical communication pattern between three users is depicted in Figure 3. Host 2 and 3 contend here
for the (cid:13)oor from host 1, and (cid:12)rst host 2 acquires the (cid:13)oor, then host 3. A turn consists accordingly of a resource
contention time X, an activity ((cid:13)oor holding) time A, and an optional idle time I. Based on this abstraction, we
de(cid:12)nethee(cid:14)cacyofa(cid:13)oorcontrolprotocol,denotedby(cid:17),astheproportionoftimethataprotocolneedstoallocate
a resource,including overheadfrom the protocol itself, the network, and user behavior. Formally, the e(cid:14)cacy is the
ratio of the average (cid:13)oor usage time U(cid:22) vs. the overall average turn length T(cid:22) = X(cid:22) +A(cid:22)+I(cid:22), given by Eq. 1. The
averagecontention period X(cid:22) and activity period A(cid:22) together are called the average busy period, B(cid:22) =X(cid:22) +A(cid:22). These
turnperiodsserveasthebuildingblocksforoure(cid:14)cacyanalysis,consideringbothpoint-to-pointandbroadcaststyle
communication.
U(cid:22) U(cid:22)
(cid:17) = = (1)
T(cid:22) B(cid:22)+I(cid:22)
3.1. Random-Access Group Coordination Schemes
Incoordinated Social Mediation (RSI) re(cid:13)ectspurely random resource accessin self-moderatingsessions. Thereis no
system support to mediate con(cid:13)icts and ensure system-wide resource consistency. Assuming that all information is
sent reliably to all user hosts, the latency introduced by the system can lead to inconsistent views of the (cid:13)oors at
di(cid:11)erent hosts and corrupt user cooperation. When this occurs, we assume that all users involved in the con(cid:13)ict
must restart their activities.
Theorem 1. The e(cid:14)cacy of RSI without multicast support is
(cid:0)2(cid:21)((cid:14)+n(cid:15))
(cid:17)RSI =(cid:21)((cid:14)+n(cid:15))e (2)
T
X A X I
t
d + ne
collides with collides with
start of A end of A
Figure 4. Typical RSI timeline.
Proof: Figure 4 shows a prototypical RSI turn. The proof is the same as for medium access in an ALOHA
16
channel. In the point-to-point model, we assume that n hosts are actively monitoring each other, and it takes an
additional time n(cid:15) for all hosts to perceive the activity from a given host. All the information is exchanged reliably,
and the averagevulnerability interval is twice the total of the task length and the added overhead incurred in serial
communication of the task information to all hosts (i.e., (cid:14)+n(cid:15)), becausemessagescan intercept activities any time.
(cid:0)2(cid:21)((cid:14)+n(cid:15))
Because request arrivals are Poisson, the probability that a task is successful is e . The success probability
2
times the number of arrivals in one activity period results in Eq. 2.
Corollary 1. The e(cid:14)cacy of RSI with multicast support is
MC (cid:0)2(cid:21)(cid:14)
(cid:17)RSI =(cid:21)(cid:14)e (3)
This follows under the assumption that every update to and from a host requires only one transmission.
Social Mediation with Feedback (RSF) assumes that users cooperate based on social protocols. Feedback on
remote activities is gathered through the user interface. If a user contends for a (cid:13)oor and perceives remote activity,
she would back o(cid:11) for a random period and attempt to reclaim the (cid:13)oor, after remote activity subsided. RSF
15
in video conferencing is often realized as \voluntary distributed control", where cooperative users switch video
streamsmanuallyonando(cid:11),dependingonwhethertheyprefertoreceiveorsendspeci(cid:12)cvideotransmissions. POTS
conferencing also relies on RSF, which works well for very small groups.
Theorem 2. The e(cid:14)cacy of RSF without multicast support is
(cid:14)
(cid:17)RSF = (cid:14)+e2(cid:21)((cid:13)0+n(cid:15)) (cid:21)e((cid:21)(cid:13)(0(cid:13)+0+nn(cid:15))(cid:15)()1(cid:0)(cid:0)1e(cid:0)(cid:0)(cid:21)(cid:21)(((cid:13)(cid:13)00++nn(cid:15)(cid:15)))) +(cid:13)0+n(cid:15)+(cid:28) +(cid:19)+ (cid:21)1 (4)
h i
T
Xf Xs A I
t
Lf +g t g t d +ne t i+1/l
Figure 5. Typical RSF timeline.
Proof: A prototypical timeline of RSF is depicted in Figure 5. n active hosts need to sense and process remote
0
activity through the network interface. (cid:13) is the user time to remotely sense the resource state. The success
pofro2b(a(cid:13)b0i+litny,(cid:15))desenco.t,eid.ea.,sPPss,=eqPu[a0lspathckeeptrsoibnab(cid:23)i]li=tyet(cid:0)h2a(cid:21)t((cid:13)n0o+na(cid:15)c)t.ivTithyepaavcekreatgaeruritvileiszaintioannpavereiroadgelavsutslnUe(cid:22)ra=bi(cid:14)liPtys,paenrdiodth(cid:23)e
length ofthe averagebusyperiodB(cid:22) =X(cid:22)+A(cid:22)isdetermined bythetime neededtohandleunsuccessful(cid:13)oorrequests
in the failed contention period Xf and successful requests in Xs, with B(cid:22) = (1(cid:0)Ps)Xf +PsXs. An average failed
turn attempt consistsofageometrically-distributedinde(cid:12)nite number (L)ofinterarrivaltimesof(cid:13)oorrequestswith
0
duration f sec (average time between failed (cid:13)oor-request arrivals), plus the duration observing a request ((cid:13) ). The
17
values for L a(cid:21)n((cid:13)d0+fn(cid:15)h)ave been deriv0ed by T(cid:0)a1kagi(cid:0)a(cid:21)n(d(cid:13)0+Knl(cid:15)e)inrock. (cid:0)(cid:21)((cid:13)S0u+bns(cid:15)t)ituting our notation in these results, we
obtain L = e and f = ((cid:21)((cid:13) +n(cid:15))) (cid:0)e =(1(cid:0)e ), respectively. Accordingly, the average
e(cid:21)((cid:13)0+n(cid:15))(cid:0)1(cid:0)(cid:21)((cid:13)0+n(cid:15)) 0
time of a failed turn attempt equals Xf = (cid:21)((cid:13)0+n(cid:15))(1(cid:0)e(cid:0)(cid:21)((cid:13)0+n(cid:15))) +(cid:13) +n(cid:15)+(cid:28). An average successful turn lasts
(cid:20) (cid:21)
0
Xs =(cid:14)+(cid:13) +n(cid:15)+(cid:28). Finally,basedonthePoissonassumption,anidleperiodconsistsofanaverageidletimeinterval
plus the time until the next (cid:13)oorrequest arriveson the average,I(cid:22)=(cid:19)+(cid:21)1. Substituting intoEq.1,weobtain Eq.4.
2
Corollary 2. The e(cid:14)cacy of RSF with multicast support is
MC (cid:14)
(cid:17)RSF = (cid:14)+e2(cid:21)(cid:13)0 (cid:21)e(cid:13)(cid:21)0(cid:13)(01(cid:0)(cid:0)1e(cid:0)(cid:0)(cid:21)(cid:21)(cid:13)(cid:13)00) +(cid:13)0+(cid:28) +(cid:19)+ (cid:21)1 (5)
h i
0
With multicast support, the vulnerability period reduces to 2(cid:13) , and (cid:15) becomes negligible.
14,18
In Activity Sensing (RAS), activities on shared resources are monitored by a background process at the
sessionlayer(withouttheuserhavingtodoso), inordertosensewhichhostcurrentlyoperatesontheresource. The
16
RAS concept is related to collision sensing on a multiaccess channel. In principle, no changes to a collaboration-
unawareapplicationarerequired,becauseamodi(cid:12)edX-serverwouldinterceptand(cid:12)ltercallstosharedapplications.
The RSF system agent would back o(cid:11) for the user, when perceiving remote activity, or allow immediate access to
the resourceotherwise. Distributed activitysensingagentscollectivelymonitor resourcestatesmoreaccuratelythan
humans could.
Theorem 3. The e(cid:14)cacy of RAS without multicast support is
(cid:14)
(cid:17)RAS = (cid:14)+e2(cid:21)((cid:13)+n(cid:15)) (cid:21)e((cid:21)(cid:13)(+(cid:13)+nn(cid:15))(cid:15)()1(cid:0)(cid:0)1e(cid:0)(cid:0)(cid:21)(cid:21)(((cid:13)(cid:13)++nn(cid:15)(cid:15)))) +(cid:13)+n(cid:15)+(cid:28) +(cid:19)+ (cid:21)1 (6)
h i
T
Xf Xs A I
t
g t+Y g t d +ne t i+1/l
Figure 6. Typical RAS timeline.
Proof: The timeline is shown in Figure 6. (cid:13) is the system time to sense the resource state. Without multicast
support,ahosthastosend(cid:13)oordirectivesindividuallytoeveryotherhost,whichaccordingtoourassumptiontakes
(cid:13) +n(cid:15). Because multiple unicast messages are used by each host, (cid:13)oor-state inconsistencies can arise during the
entire time. The host is exchanging (cid:13)oor directives, i.e., the vulnerability period of the protocol is (cid:23) = 2((cid:13) +n(cid:15)).
Using this vulnerability period, Eq. 7 follows using the same approach as described in the proof of Theorem 4.
Theorem 4. The e(cid:14)cacy of RAS with multicast support is
MC (cid:14)
(cid:17)RAS = 1 (cid:21)(cid:28) (7)
(cid:14)+(cid:28) + (cid:21) +e ((cid:13)+2(cid:28) +(cid:19))
Proof: The accessstrategyisassumed asnonpersistent, i.e., ahost backso(cid:11)immediatelyfromattemptingtoaccess
a resource and claims the resource once it appears free again. The vulnerability period for accessing an unused
resourceisonepropagationdelay,(cid:23) =(cid:28),withinwhichotherhostscancauseacon(cid:13)ict(oppositetoRSF,wheretwice
(cid:0)(cid:21)(cid:28)
the length of the contention interval is the average vulnerability period); therefore, Ps =P[0 packets in (cid:23)]=e .
TheaverageutilizationperiodisU(cid:22) =Ps(cid:14). Theaveragelengthofasuccessfulbusyperiodissimply(cid:13)+(cid:14)+2(cid:28),which
accounts for the delivery and processing of (cid:13)oor directives, the activity period, and associated network latencies.
The length of an average unsuccessful activity period consists of one truncated activity lasting (cid:13) sec, followed by
17
one or more similarly truncated activities sent within time Y sec, where 0 (cid:20) Y (cid:20) (cid:28). The expected value of Y is
Y(cid:22) =(cid:28)(cid:0)(cid:21)1(1(cid:0)e(cid:0)(cid:21)(cid:28)); therefore, theaverageduration ofafailedcontentionperiodis(cid:13)+2(cid:28)(cid:0)(cid:21)1(1(cid:0)e(cid:21)(cid:28)). Thelength
of the averagebusy period is then B(cid:22) =(cid:13)+2(cid:28) (cid:0) (cid:21)1 +e(cid:0)(cid:21)(cid:28)((cid:14)+(cid:28) + (cid:21)1). The average idle interval is again I(cid:22)=(cid:19)+ (cid:21)1.
2
Substitution into Eq. 1 yields Eq. 7.
3.2. Scheduled Group Coordination Schemes
In contrast to the contention time in RGC, (cid:13) in SGC denotes the time to transmit a (cid:13)oor token to the next host.
Two SGC approaches, polling, and reservation, have previously not been used in telecollaboration. Polling involves
long wait times, and reservation schedules can quickly become obsolete. We focus on token passing, where the (cid:13)oor
isbeing askedfor,oro(cid:11)eredto hostsin aprede(cid:12)ned serviceorder. We discuss threemaincasesofcontroltopologies
for hosts.
In Direct Coordination (STD), each host in a group is fully connected to every other host. STD improves the
n(n(cid:0)1)
responsetime, however,the number of links is 2 and growsas the square of the number of hostsin the session.
Many small-scale commercial video conferencing systems follow this unscalable model.
Theorem 5. The e(cid:14)cacy of STD without multicast support is
(cid:14)
(cid:17)STD = 1 (8)
(cid:14)+n((cid:13)+(cid:28) +(cid:15))+(cid:19)+ (cid:21)
T
Xr Xg A I
t
g t g t d +ne ti+1/ l
Figure 7. Typical STD timeline.
Proof: The timeline for STD is shown in Figure 7. The average utilization period is U(cid:22) = n(cid:14)Ps, because (cid:13)oor
capture is perfect and any one of the n active hosts can acquire a (cid:13)oor of holding time (cid:14) with success probability
1
Ps = n. A (cid:13)oor may not be released at FH, until successor FH’ received it. The control packet overhead is (cid:13) and
the propagation delay is (cid:28). Unicasting a (cid:13)oor request to the n(cid:0)1 active hosts amounts to (n(cid:0)1)((cid:13)+(cid:28) +(cid:15)), plus
((cid:13)+(cid:28) +(cid:15)) for a reply. The averageactivity duration is A(cid:22)=(cid:14) and may be trailed by an idle interval consisting of a
period (cid:19) and an averageinterarrival time for all hosts, I(cid:22)=(cid:19)+ (cid:21)1. Substitution into Eq. 1 results in Eq. 8. 2
Corollary 3. The e(cid:14)cacy of STD with multicast support is
MC (cid:14)
(cid:17)STD = 1 (9)
(cid:14)+3((cid:13)+(cid:28))+(n(cid:0)1)(cid:15)+(cid:19)+ (cid:21)
With multicast support, the request-reply-releaseexchangeof control packets takes a time of 3((cid:13)+(cid:28))+(n(cid:0)1)(cid:15), if
we assume that every host incurs host processing overhead from its n(cid:0)1 neighbors.
19,20
InRing-basedCoordination(STR),a(cid:13)oortokencyclesthroughalogicalringarranginghosts. Varioussystems
based on this idea havebeen discussed. A host that is ready to start an activity, captures the passing token, inserts
a command sequence with address and control information, sends the activity packets within this turn period and
transfersthetokenaftercompletiontothesuccessorhost. Ahostwithoutpending(cid:13)oorrequestspassesontheo(cid:11)ered
token. Tokens held in excess time may expire and incite automatic transfer. The prede(cid:12)ned token passing schedule
may not re(cid:13)ect spontaneous interactivity. The (cid:13)oor can be granted ahead of its token position to a successor host,
oritmayonlybeacquiredbyahostwhenthetokenpassesthroughthathost. Likewise,atokencanbeimmediately
released after transmission (RAT), or released after one more reception (RAR) at the sending host. For e(cid:14)ciency
reasons we focus on RAT. In ourmodel, the pre-arrivalthink time (cid:12)1 plus the token-presencetime (cid:12)2 <(cid:12)1 must be
smaller than the ring cycle time. If the (cid:13)oor is taken, then (cid:12)2 (cid:25)((cid:14)+(cid:28) +(cid:13)).
Theorem 6. The e(cid:14)cacy of STR without multicast support is
(cid:0)(cid:21)((cid:12)1+(cid:12)2)
(cid:14)(1(cid:0)e )
(cid:17)STR = n (cid:0)(cid:21)((cid:12)1+(cid:12)2) 1 (10)
2((cid:28) +(cid:13)+(cid:15)+(cid:12)2)+(cid:14)(1(cid:0)e )+(cid:19)+ (cid:21)
Tt-1 Tt
At-1 Xt At
t
d t b 1 b 2+g d +ne t b 1
Figure 8. Typical STR timeline.
Proof: The timeline forSTR is depicted in Figure8. We assumeperfect (cid:13)oorcaptureandonly onehostis active
at any time. The average utilization is U(cid:22) = (cid:14)Ps, with Ps as the success probability that the token is available in
the period X(cid:22) = (cid:12)1+(cid:12)2. The probability that a (cid:13)oor can be claimed by a user is Ps = 1(cid:0)e(cid:0)(cid:21)((cid:12)1+(cid:12)2). The token
cycle time is a function of the active host set n, and with growing session size it is less likely that all hosts will
n
engage in (cid:13)oor contention. The cost to transfer a (cid:13)oor token in a cycle involves on the average 2 hosts, amounting
to n2((cid:13)+(cid:28) +(cid:15)+(cid:12)2) including processingoverhead n2(cid:15). The averageturn lasts hence T(cid:22)= n2((cid:13)+(cid:28) +(cid:15)+(cid:12)2)+A(cid:22)+I(cid:22).
Theidletime isagainI(cid:22)=(cid:19)+(cid:21)1. SubstitutingU(cid:22) andT(cid:22) intoEq.1yieldsEq.10. Theoverheadtomaintainthetoken
2
is not included in this result.
Corollary 4. The e(cid:14)cacy of STR with multicast support is
(cid:0)(cid:21)((cid:12)1+(cid:12)2)
MC (cid:14)(1(cid:0)e )
(cid:17)STR = n (cid:0)(cid:21)((cid:12)1+(cid:12)2) 1 (11)
2((cid:28) +(cid:13)+(cid:12)2)+(cid:14)(1(cid:0)e )+(cid:19)+ (cid:21)
n
With multicasting, a token is sent only once to the network interface, but it takes on the average 2 hops to cycle
back for another turn option, however,the processing overhead (cid:15) vanishes.
Tree-based Coordination (STT) allows for more e(cid:14)cient inter-group collaboration and hierarchical mixing of
21
media sources. Its control infrastructure can work in symbiosis with reliable multicasting over a single shared
22
acknowledgment tree, allowing more scalable and economic transmission of session data. STT control messages
traversebranches of the tree in a parent-child relation re(cid:13)ecting multicast group membership. Each host must only
deal with messages from immediate neighbor hosts. Floor directives can be coalesced into single messages, and can
beforwardedthroughthe treein anaggregatedfashion. Singlehosts willhencenotsu(cid:11)erfrom messageimplosion,a
problem knownfrom reliablemulticast with regardto acknowledgingreceivedorlost messages. The communication
delay depends on the height of the controltree, rather than the session size. The star topologyis a special case of a
tree, which works for small sessions.
Theorem 7. The e(cid:14)cacy of STT without multicast support is
(cid:14)
(cid:17)STT = (K+1)P(cid:22)((cid:13)+(cid:28) +(cid:15))+((cid:14)+P(cid:22)(cid:28))+(cid:19)+ (cid:21)1 (12)
Proof: The timeline for STT is depicted in Figure 9. We have again perfect (cid:13)oor capture due to explicit token
exchange, with U(cid:22) = n(cid:14)Ps, and Ps = n1. With the normalized average path length P(cid:22), the average duration of the
(cid:13)oor capture period amounts to 2P(cid:22)((cid:13) +(cid:28) +(cid:15)). Assuming that a host does not know the location of FH or FC,
T
Xr Xg A I
t
P(g+t+e) P(g+t+e) d Pti +1/l
Figure 9. Typical STT timeline.
exploratory unicast of a control message from one host to its parent and to each of its children would amount to
X(cid:22) (cid:20) (K +1)P(cid:22)((cid:13)+(cid:28) +(cid:15)) plus a targeted reply costing P(cid:22)((cid:13)+(cid:28) +(cid:15)). However, as outlined in Section 4, hosts can
be tagged with unique labels, which allow for e(cid:14)cient absolute or relative routing of control directives toward the
FC or FH. Consequently, a control directive must be sent only to one neighbor node as the gateway on the path
to FH, hence K = 1. Signaling the conclusion of the activity period adds another propagation delay, A(cid:22) = (cid:14)+P(cid:22)(cid:28).
The activity period may be trailed by another idle period of averagelength I(cid:22)=(cid:19)+ (cid:21)1. Substituting into Eq. 1 gives
2
Eq. 12.
Corollary 5. The e(cid:14)cacy of STT with multicast support is
MC (cid:14)
(cid:17)STT = 1 (13)
(cid:14)+2(cid:13)+3(cid:28) +(cid:19)+ (cid:21)
With multicasting, a request-reply pair takes two (cid:13) and (cid:28), plus another (cid:28) to signal completion of the turn. A close
correlationbetweenthe controltreeofthe STTprotocolandtheend-to-end multicast treeis assumed. A hostsends
only one message to the network interface, i.e., P(cid:22) =1, K =1, and (cid:15) becomes negligible.
3.3. Results
MC
We compare (cid:17) of the discussed paradigms in four cases: (1) a small group in a network with low link latency;
(2) a small group in a high-latency network; (3) a large group in a low-latency network; and (4) a large group in a
high-latencynetwork. Groupsizesaren=5(small)andn=300(large),correspondingtoPC-conferencingsystems,
or average MBone session size. Latency is indicated by (cid:28) =0:005 s (low), and (cid:28) = 0:4 s (high). The time to sense
0
(cid:13)oorinformationis(cid:13) =0:25s. Acontrolpacketof25byteslengthismeasuredwith(cid:13) =(cid:15)=0:02s. Thenormalized
(cid:14) (cid:14)
activitytime is(cid:14) =1andtoken-ring\thinktimes"aresetto(cid:12)1 = 2 sand(cid:12)2 = 10 s. Thetypicalidletimeischosen
(cid:14)
as (cid:19)= 5 s. Figures 11 and 10 plots the resulting e(cid:14)cacy.
CCoommppaarriissoonn ffoorr ssmmaallll sseessssiioonnss // llooww llaatteennccyy CCoommppaarriissoonn ffoorr llaarrggee sseessssiioonnss // llooww llaatteennccyy
11..0000 11..0000
RRSSII RRSSII
RRSSFF RRSSFF
RRAASS RRAASS
SSTTDD SSTTDD
SSTTRR SSTTRR
00..8800 SSTTTT 00..8800 SSTTTT
C)C) 00..6600 C)C) 00..6600
MM MM
a (a ( a (a (
EtEt EtEt
acy: acy: acy: acy:
EfficEffic 00..4400 EfficEffic 00..4400
00..2200 00..2200
00..0000 00..0000
00..0011 00..1100 11..0000 1100..0000 110000..0000 11000000..0000 00..0011 00..1100 11..0000 1100..0000 110000..0000 11000000..0000
OOffffeerreedd llooaadd:: GG OOffffeerreedd llooaadd:: GG
Figure 10. E(cid:14)cacy with multicast support for low network latency.
Thee(cid:14)cacyofRSIisbelow20%inallfourscenarios. SystemsemployingRSFbene(cid:12)tfromcoordinationattempts
and improve slightly over RSI, approximating 25% e(cid:14)cacy in networks with faster links. Both schemes are instable