Projeto Final SOP Artigo

Abstract—We argue that a multimedia session can be the resultof carefully orchestrated actions as opposed to simply the outputfrom a single end-to-end or a multicast transfer mode. This is atthe heart of the difference between the proposed SessionOrchestration Protocol (SOP) to SIP´s companion the SessionDescription Protocol (SDP). As such, a device display space andtime are seen as a virtual multi-input scene where differentviewing areas may be composed of simultaneous streams usingspace and time diversity. To maintain scalability, we do not adoptWeb-based orchestration; instead we suggest a simple SIP/SDPextension, we call SOP.

Our results show that SOP is efficient when under thecurrent multipath scheduling.

Index Terms— multimedia, video streaming, sessionmanagement, orchestration.

I. INTRODUCTION

rchestration has emerged as an important underlyingmechanism for the Web Service Object Architecture[1]. It is hereby conceptually applied in the context of

advanced session management to enhance the end mediapresentation by embedding actions; we refer to as directives,in a well sequenced and prearranged manner. Using suchplanned session detours, a great deal stands to be gained interms of enhanced services and support for new businessmodels as shown later. The costs of which, are small changesto SDP.

O

Current work on multimedia standards allows for themultiplexing of different layers and their independenttransport by using special multilayered video coding [53]. Thepresent research looks at yet a new form of “multiplexing”, ormore specifically media composition. We see it as the outcomeof well established content orchestration and interaction. Wedraw a parallel with yet another current web softwareengineering tendency, namely, mashup. A mashup is a webpage or application that uses and combines data, presentationor functionality from two or more sources to create newservices [35]. There are two important aspects in whichemerging service mashup differs from traditional web servicedesign. First, traditional content is aggregated only at theserver whereas under the mashup philosophy, it can also takeplace at the client. Further, mashup software may presentcontent side by side within a single melting pot offering a newaggregation style. Both ideas are embraced by our SOPproposal. The difference is that SOP´s changes are made at theend-to-end streaming transport protocols level and SOP is

Manuscript received xxx, 2012. D. F. Hadj-Sadok is with the FederalUniversity of Pernambuco, Recife, Brazil (phone: +5581 21268954; fax:+558121268955; e-mail: jamel@ gprt.ufpe.br).

directly concerned with user perception quality issues. We

show through conducted evaluations that in this case,judicious packet level scheduling, loss and bandwidthestimation are primordial for guarantying adequate mediaorchestration.

A. Motivation

The recently created IETF Peer-to-Peer Streaming Protocol(PPSP) working group [5] opens the way to innovativethinking and the advancement of an important area. The WG´scharter establishes the development of two signaling andcontrol protocols in a peer-to-peer (P2P) streaming system fortransmitting live and time-shifted media content with nearreal-time delivery requirements. This is the right window ofopportunity to embrace innovative ideas leading to theadvancement of streaming business and technological models.

When considering solutions for multimedia streaming andsession control, SIP comes to mind as a strong catalytic forceshaping today's Telecom industry. While specifically designedwith reusability, backward compatibility and simplicity inmind, it also works well with Internet protocols such as SOAP,HTTP, XML, VXML, WSDL, UDDI and SDP. For simplicity,SIP was modeled after HTTP, using URLs for addressing andSDP to convey session information. In addition to sessioncontrol, it enables name translation and user location, featurenegotiation, call participant management and call featurechanges. Unlike the ITU-T H.323 suite for multimedia sessionmanagement and interchange, fewer and simpler methods aresupported, namely, INVITE for requesting a session, ACK,BYE, CANCEL, OPTIONS, REGISTER for registeringcurrent user location and INFO used for mid-session signaling[36].

Despite SIP´s acceptance, building upon its success is anongoing exercise. The following are some of the IETFcomplementary initiatives seeking to ensure portability ofservices across these differing realms. They are: PINTs andSPIRITS for interworking, Number resolution (TRIP andENUM), Seamless signaling - SIGTRAN and SIP-T, InstantMessaging and Presence (IMPP), SIP for Instant Messagingand Presence Leveraging Extensions (SIMPLE). In turn, thisdocument puts forward another small but effective extensionframework for the flexible support of new services as part ofthe emerging P2P streaming and multimedia sessionmanagement protocols.

This work is further motivated by the widespread use ofmultimode terminals with 3G, 4G, WLAN, Bluetooth andother interfaces, coupled with the current operator yet perhaps“shy” initial deployments of new multi-gigabit 4G Long-TermEvolution (LTE) services due to backbone limitations. As LTEcarriers need to evolve their metro network infrastructuretoward an Evolved Packet Core (EPC) for wireless broadband,worldwide deployment remains limited. Operators such asSweden´s TeliaSonera, Europe´s Vodafone, Japan´s NTTDoCoMo, US Verizon Wireless, MetroPCS and AT&T have

Orchestration and Path Diversity for Real-timeMultimedia Streaming

Djamel F. Hadj-Sadok

1

launched 4G LTE services in 2010. Reported downstream 4Gaccess rates range between 12-45 Mbps in Sweden and 5-12Mbps in the United States whereas uplink speeds are moreconservative. Despite the considerable user rates, delay nearsthe 40ms. Though seen as a DSL replacement by many, usercosts are also likely to be an additional discouraging factor.Most 4G operators have opted for a tiered (or metered) pricingpolicy for their LTE data plans, unlike 3G popular flatmonthly unlimited data plans. Reported costs include a basemonthly fee and a monthly data limit for example between 10and 30 GBytes.

4G´s choice of metering could potentially discourage manycustomers from using huge volumes of data. Clearly, there isroom for more immediate solutions until such limitations areovercome. As a subscriber to two or more access networks,one is interested in investigating ways of making use of thesein a concerted or orchestrated manner to access servicesdemanding higher transfer rates such as real-time streamingapplications. In this work, we seek to reap the benefits ofsimultaneous multimode access support and IP multi-homingsupport, including scenarios with wireless technologies.

But the current best effort Internet lacks support for efficientstreaming applications due to the presence of delay variations,limited throughput and experienced packet loss. Particularattention is given by this work to the problems of multimediastreaming over possibly harmful Internet paths. For instance, itis known that video quality for popular MPEG-4 degradesconsiderably above a 1% packet loss rate.

More specifically, we examine how continuous mediadelivery may be offered by exploiting intelligentlyorchestrated multipath diversity and parallel packetforwarding. Multipath diversity is of interest as it dispersespacket loss bursts. A good understanding of packet loss in amultipath setup remains non trivial and new insights are stillrequired. This problem is known as real-time packetscheduling while taking advantage of path diversity.

In the following text, video streaming is used as the mainscenario for explaining the new session orchestration ideas.

II. SESSION ORCHESTRATION

Media streaming has room, in our view, to support additionaloperations, or methods according to SIP´s terminology, thatmay act along any of the following stages: transport,interpretation, presentation, interaction, storage, billing, etc.Following the Internet tradition, this problem is pushedtowards the edge where we look at the design of new receiverand / or sender protocols that ensure adequate service levels.In this work, we use SOP, a receiver driven mediaorchestration protocol for the management of the differentmultimedia flows. The present SOP proposal introducesdirectives or activities with functionalities that at some stagemanipulate or interact in some form with the multimediastreams. Similarly to SDP, SOP is not a transport protocol butrather a data format and some of its standard notations wouldrequire IANA1 registration when adopted at a wide level.

1 The Internet Assigned Numbers Authority (IANA) acts as a global registry service.

A. From SIP/SDP to SOP

The Session Description Protocol (SDP) uses in fact a textdata format that describes multimedia sessions. Its role islimited to session announcement, invitation, and other formsof multimedia session initiation. SDP is also designed with theperspective of a receiver´s view in order to give this enoughinformation to join a session. This includes session name, forhow long it has been active, the media that compose it, howeach medium may be received (address, port, format, etc.).Additional information includes, bandwidth resources andnames of administrative contact persons. Overall, a sessiondescription consists of information that applies to the wholesession and one or more descriptions, each applying only to amedium specific stream. SDP information is merely conveyedto the recipients [39]. It is up to the recipients to act upon it.Table 1 bellow shows the session description syntax (* standsfor an optional item).

Table 1 Session Description Format [39]

Though a data notation, SDP has built in dynamicity. Forexample, it may be used to make periodic announcementsusing the session announcement protocol (SAP) towards aknown multicast address. SDP may convey an arbitrary list ofstart and stop times bounding the session. For each bound,SDP may stipulate repeat times such as "every Friday at 10:00am for two hours". Furthermore, SDP may include pointers inthe form of Universal Resources Identifiers (URIs) for moreinformation about a session using the “URI” field. One “i=”information field may be used by a medium. The connectiondata is given by the "c=” field and contains the network type(typically “IN” for Internet, address type (typically has value“IPv4”) and connection address (typically a Class D multicastaddress). For a full description of SDP features please refer tothe RFC pages [39].

Session description v= (protocol version) o= (originator and session identifier) s= (session name) i=* (session information) u=* (URI of description) e=* (email address) p=* (phone number) c=* (connection information -- not required if included in all media) b=* (zero or more bandwidth information lines) One or more time descriptions ("t=" and "r=" lines; see below) z=* (time zone adjustments) k=* (encryption key) a=* (zero or more session attribute lines) Zero or more media descriptions

Time description t= (time the session is active) r=* (zero or more repeat times)

Media description, if present m= (media name and transport address) i=* (media title) c=* (connection information -- optional if included at session level) b=* (zero or more bandwidth information lines) k=* (encryption key) a=* (zero or more media attribute lines)

2

A SOP Design Principles

The following two main aspects characterize SOP:augmentation and cross concern. By augmentation, oneintends to bind SOP to existing session and streamingcapabilities, giving these additional semantics, options andvalues, instead of defining new ones. Hence it is not SOP´sintent to redefine existing session capabilities, mechanisms orservices naturally spread across many protocols and services,but it is rather incremental to these, assisting in their tuningand enhancing some of these at the session control level. Next,there is a cross layer characteristic to this proposal and isinherent to the first aspect. SOP deals with cross layer servicesthat by nature touch many aspects of a streaming process,presentation, transport, framing etc., during an entire session,in an end-to-end scope.

SDP/SIP is recipient oriented and so are the SOP extensions,see Figure 1. To orchestrate the receiver side actions during asession, one may use an imperative language. The SOA effortadopted the Business Process Execution Language (BPEL) [7]with support for variables, conditional structures, loops,invocations, parallelism and many other structures easilyfound in programming languages. We adopt nonetheless asimpler readable notation to describe a SOP session and themedia it transfers. There is a tradeoff here between simplicityand orchestration support. The present proposal seeks to strikea balance between the two. A full BPEL-like notation is, fornow, not an option as it introduces XML parsing overhead andmay scale poorly in a real-time Telecom scenario.Nonetheless, the conceptual ideas behind orchestration as seenby the BPEL community are inspiring for the design of ourextensions to SDP.

Figure 1 Receiver Driven Orchestration

The current SDP RFC offers a description in three parts:session, time and media. The session part describes the generalinformation including the source of the media. The time partestablishes the duration of the presentation whereas the mediapart presents the different media that make up the sessionincluding audio, video, data, etc. To achieve our goals, someadditions are made. Firstly, session information such as thesource address must also be passed to the media level as weassume that different content may originate from differentsites. Secondly, in addition to the timing information related tothe overall session, media level timing is introduced. As aresult, a simple orchestration is supported where media maybe sequenced, run in parallel or have one wait for the otherone to run simultaneously, etc. Thirdly, media within a sessionare tagged for ease of handling and manipulation. This couldbe the case, for example, when achieving orchestration or

when asking another peer to relay such media to others inorder to improve user quality of perception and cooperate withother peers. Fourthly, more than a source address may begiven within a session description, it is expected that thiswould ensure for example load balancing as different receiversmay choose different sources on average. Fifthly, repeatactions are not only associated to an entire session but also togiven media hence increasing the granularity level for mediamanipulation. This could for example be used to include acarousel type of advertizing material which keeps rewindingand playing automatically when it reaches its end. Finally, anode may be told that other peers (partners) would need it tostart relaying the given media it is receiving.

There are many reasons for wanting to enhance the sessiondescription scope by allowing many entities (peers or partnersin SOP´s terminology) to coordinate their efforts in creatingand controlling the way media is exchanged in both space andtime. By doing so, they should be capable of offering newvalue added functionalities in the form of a composite multi-dimensional session.

In this work, we remove the media continuity concept wheresome went to the extent of saying that there is no media unit.Our proposal, introduces media units in terms of two (time)and three (space) dimensions to increase control over thestreamed media. In [4], some semi-structured features areadded to the content but their operations were limited tosearch and query.

Though attributes have been included into SDP as a meanfor its extension, we used these in few places only. This is dueto their limited attribute-value syntax. For more complexdescriptions, complete new tags are introduced by SOP. Next,only the proposed extensions to SDP are detailed and theirroles in the orchestration of session information are explained.

A Remote Invocation

SOP envisages some advanced interaction between thestreamed media and the external Web information objects. Forexample, along the content one may dynamically point to anURL that fits into the current context. Such action could resultin some content right management control activation,consultation of purchasing information, authentication orbilling data, advertizing of similar videos, addition of useruseful information, conducting an opinion poll, etc. UnlikeSDP´s URI, when a URL is processed, new content may begenerated, an action may be taken, usage statistics may beupdated, etc. The following SOP/SDP extension,“a=URL:run”, is used to invoke an external URL. Of course,security is a concern and should be dealt with adequately forexample through prior authentication which is outside thescope of this proposal.

B. Resource Information

A source may indicate to the receivers the time validity ofsome content using a removal action as in “a=del:Date”. Thedate information must be given in the NTP format. This end-to-end dialogue enables some resource configuration whereboth sender and receiver have a say, though ultimately thereceiver is responsible for any enforcement.

3

C. Regulatory Differences

Depending on the location, different regulations and rulesmay apply for content distribution. For instance, some contentmay not be distributed at certain locations or can only bedistributed at some others [3]. A distribution map mayaccompany given content and act as a time and space filter toit, but this is not easy to implement nor to enforce as it wouldrequire tampering with existing media formats. Ideally, asender peer should be unaware of such restrictions, hencefiltering per receiver location dependent additional contentmay take place along the path. In practical terms, the senderneeds to include information within the session descriptionthat could be used to restrict its distribution.

AS number lists “a=allow:values” and;AS number lists “a=forbid:values”

A simple solution, as shown by the commands above, hasbeen adopted to try and filter out unwanted receivers. It isbased on the (loose) simple concept of Autonomous System(AS) numbers. Future advanced filters may be based on timezones, IPv4 or IPv6 address ranges, etc.

D. Multipath Streaming

Multipath streaming is gaining support with the increasingpresence of multimode terminals. As such, it is taken at theheart of the SOP proposal. Many video splitting schemes haveemerged [8]. The advantages stem not only from the addedthroughput but also in the reduction of correlation betweenpacket losses in a wireless environment. Further, asmultimedia sessions are expected to relatively last longdurations, network conditions are likely to change. Multipathstreaming is better suited for adapting to such changes.

E. Stream Multiplexing

Similarly to the above scenario, one may think of situationswhere “intelligent” or enriched stream processing isperformed. Streams may be merged from different sources tobe presented in a synchronized way, each of these may be toldto share with others the presentation media, or periodicallytake over some part of a display, etc.

1) Space Multiplexing Here, space directives are mixed with content at different

locations. This could be useful if an ISP may want to add someadvertizing material. Such additional data may be subject topre-established agreements among providers. For example, atransport provider may see a business opportunity in relayingP2P traffic, instead of shaping or removing this as is the casesometimes today, as long as it could insert some advertizingmaterial.

There is an added benefit from the above mechanism, that ofcopyright protection. By merging new content onto a givenvideo, this facilitates tracing it back through the transportnetwork. To support CRM, this mechanism may be used toinclude watermarks and log content streaming along a networkpath.

2) Time MultiplexingSimilarly to space multiplexing, time-based stream content

multiplexing uses time diversity to include at different timeswithin a session, new material. This will then be merged withthe original content using directives similar to BPEL´sactivities such as wait, sequence, if , while, etc. The directives

adopted in this context allow the definition of periodical,repetitive, or time spaced media insertions.

Table 2 Example showing SOP extensions

In order to manipulate individually streams and media, thesehave been tagged using a simple “per-session” distinguishedname. Hence, following each line with an SDP mediadeclaration there is a tag-name declaration. The two followinglines (10 and 11) associate the name audio-from-Globo to thecurrent audio signal.

Line 1 in the above description from Table 2 shows theversion number and line 2 the origin of the session. Thesubject is given in line 3 with additional information in lines 4to 6. Line 7 gives the session address and line 8 shows itstiming information. Next a set of media are described. Theabove example shows three media, namely, audio-from-Globo,scenes, and advertizing. Their associated coding formats arealso given. Note that all of these media may draw by defaultfrom the origin specified in the session description part, exceptaudio-from-Globo which may be received from differentaddresses as specified in the “p” command (short for paths)given in line 11. The tagname attribute extension associatesnames with each of the media defined in this example asshown in lines 10, 13 and 15. Line 16 states that theadvertizing material must be repeated every 15 minutes during5 minutes. The “f=” extension requests from the receiver to actas a possible server or relaying point. According to the abovedescription, it is possible for a relay node to convert to a newformat before relaying the media identified by tagname.

Last but not least, line 18 does the actual orchestration. Inour simple example, both audio-from-Globo (radiobroadcaster) and the scenes video are played at the same time(in parallel). Next, in sequence, the advertizing material isplayed to end the session.

3) Network Session CoordinationA major shortcoming of current multimedia streaming and

control protocols is their lack of or limited intelligence inadapting to network conditions. Sessions are negotiated andestablished with limited information and are mostlyconfigured in a static manner. In this work, we suggest thatboth traffic and network should, as separate entities, be able to

1. v=12. o=djamel 2890844526 2890842807 IN IP4 126.16.64.43. s=SOP Research4. i= Thoughts on the Session Orchestration Protocol 5. u=http://www.gprt.ufpe.br/DjamelSadok/sop.pdf 6. [email protected] (Djamel Sadok) 7. c=IN IP4 224.2.17.12/127 8. t=2873397496 2873404696 9. m=audio 49170 RTP/AVP 010. @=tagname:Audio-from-Globo11. p=djamel 2890844526 2890842807 IN IP4 126.16.64.4

IP4 151.16.10.1 IP4 128.37.17.6712. m=video 51372 RTP/AVP 31 13. @=tagname:scenes top14. m=video 51372 RTP/AVP 3115. @=tagname:advertizing bottom16. r= 900 30 017. f=recvrelay scenes RTP/AVP 3118. q= audio-from-Globo | scenes / advertizing

4

adapt in a very controlled way nonetheless, in order not toupset the network engineering equilibrium.

For instance, to deal with network congestion, the P2POctoscope software withstands heavy congestion using asimple session management trick; a live stream session isfirstly opened to a few peers in the network and is then spreadto the rest [1]. This mechanism is taken at the heart of the SOPdesign for orchestration support. Such smart session controlmay be made on-demand and extended to any applicationsimply by allowing SOP support at a source. Similarly, SOPwould be used to redirect existing and mostly new sessions toalternative sources in order to balance traffic, reduce load on asource as could be the case in a TV channel broadcast, etc.

By offering multiple possible sources for some session, see“p=” SOP command example above, SOP allows thebalancing of the network load and distributing the load amongdifferent media servers.

F. Mobility Challenge

The availability of multi-interface wireless devices allowsmobile users to select and switch between different radioaccess technologies such as 3G or Wi-Fi occasionally orswitch its attachment within the same system. This operationmay cause a P2P Streaming session quality to degrade due tothe introduced latency or even its termination [3]. A Contentassisted handoff between technologies may be achieved bySOP and coordinated by different peers to mitigate such effect.Let us consider a case where other nearby peers may beinstructed to help smooth content streaming during handoff.One way to achieve this is through changing a session´slayout. This is a clear example, where an end service interactswith the access networks to deal with technology inherentlimitations such as handoff latency and temporarydisconnection, both vertical and horizontal ones. Making asteaming service aware of such a challenge, may prompt thereceiving peer service to slow down the presentation, and thesource to increase its transfer rate in order to ensure thebuffering of enough video frames to mask the effect ofmobility.

The question is how could we make SOP help survive ahandoff or similar temporary disconnections. One way is usingthe multipath “p=” command information presented earlier tocreate new connections through the different interfaces of agiven terminal or peer. For example, a WLAN terminal mayinitiate new sessions over GSM or 3G (UMTS) to lowerpacket loss and delay such as when facing a handoff.

An increase in the number of peers for example relaying agiven session using the SOP command “f=recvrelay” followedby the advertizing of their addresses using the “p=”potentially increases the number of source peers anddiminishes disconnection problems. As a result, a coordinatedorchestration of a single flow over different paths, within asingle or composite session may ensure data continuity as longas these paths do use different interfaces in multimode peersand do not share any common bottlenecks.

G. New Business Model for Unwanted Peer-to-PeerSessions

Peer to peer traffic is seen by some ISPs as a nuisance. It istraffic that often clogs their networks while offering these littleor no return. Though many operators have announced anincrease in their wireless 3G and 4G access networks to fewhundred Mbps, they remain wary of the effects it will have ontheir core networks. Most see that they are not prepared forsuch traffic, let alone if they will be getting little gain from it.Other content players, see this as an opportunity to deploytheir own broadband multi-gigabit fiber access networks.Major cellular systems manufacturers are pushing towards anearly deployment of the Evolved High Speed Packet Access(HSPA+) release 72 with 4G downlink speeds in the range ofup to 28 Mbps.

In such context, we will show how the use of SOP couldhelp unlock the current transport versus content providersdebate under bilaterally beneficial new business scenarios.SOP is receiver oriented and as such allows for the possibleinclusion of entities representing a network operator forexample or a content delivery network (CDN), to manipulateand play a part in the orchestration of a session. This way, thetransport network may have a saying and benefits also frommultimedia streaming including that of unwanted P2P traffic.To do this, its own streaming entity may be included as one ofthe media sources using the “m=” SOP command, see Figure2 illustrating how this is done.

Figure 2 New Business Model for TransportPartner

Using this mechanism, one is able to make better use ofstreaming associated services such as electronic programguide servers (EPG). New business models and added servicesmay explore to the full such capability. SOP´s directives maybe used to add value to the content opening the way for newbusiness models where both the end peers as well as those thatparticipate in the overall realization of the session, in this casean operator, benefit.

H. Session Cooperation

When there are insufficient numbers or peers in a certainregion, as is the case of some Chinese P2P streaming systems,the instability of dynamic peers turns it more difficult toensure a good service as there are no stable end-to-endbandwidth assurances from the source to those peers outside

2 HSPA+ Release 8 deployment offers 42 Mbps downlink speed and is expected as early as 2010.

p=djamel 2890844526 2890842807 IN IP4 126.16.64.4 IP4 151.16.10.1IP4 128.37.17.67.

5

Session Orchestration

Packet Error ModelBandwidth Estimation

Traffic SplittingPacket Scheduling

Receiver Control

Sender Path Diversity

Error Concealment

China. As a result, SOP may adopt a hierarchical or bilateralP2P Streaming Distribution among P2P service providers.

SOP may be used to enable cache cooperation and evenforced caching to alleviate the problem and remove the needfor unified client software to watch P2P Streaming Programson global scale and across P2P content providers. Here, SOPredirects sessions to caches (surrogate servers) in order toimplement cache cooperation. Hence, a foreign cache may beseen as a remote peer that cooperates within a SOP session.For each media declaration, a cache specification may begiven.

h= IP4 126.16.64.4 IP4 151.16.10.1 IP4 128.37.17.67The above example shows the use of the SOP directive “h=”

to indicate alternate cache servers for a given session. Usingthis mechanism, operators, peer-to-peer and content deliverynetworks may share cache servers leading to the optimizationof their resources and better user quality of experience.

III. SOP PATH DIVERSITY MODEL

SOP orchestrates multiple streams to create a more advancedpresentation. In this section of the paper we analyze a peculiarscenario, that of path diversity. Though there is a large numberof streaming scenarios that could be analyzed, we look at theuse of SOP in the simple scenario of a single transmitter overmultiple paths to a receiver as shown in Figure 3.

Figure 3 Simple SOP Multipath StreamingScenario

Parallel paths are better explored when they have disjointbottlenecks as this reduces correlation among consecutivepacket losses and allows continuous transmission policyadjusting to the varying path states. In other words, thelikelihood of having a “good” path is persistent. Previouswork such as [12] may be used to determine if commonbottlenecks exist among paths. Note that the paths may sharecommon links as long as these are not bottlenecks [13].

We narrow our study to that of sending pre-storedcontinuous media. Though erasure codes have also been usedin a number of studies [10][21][17][19] we want first tounderstand the full benefits of multipath streaming without thecomplexities of such coding.

The problem of packet scheduling over competing parallelpaths may be seen as a resource allocation problem, one thatcan be dealt with using analytical models and a range ofstrategies that perform:

Splitting of traffic: with approaches that include loadbalancing using load sharing or traffic splitting, greedyallocation heuristics, depth and breadth search, binpacking algorithms, or a Generalized Multi-objectiveMulti-tree models (GMM), etc. When used the latterlead to a Pareto equilibrium which can then be solvedusing evolution theory for example, etc;

Queue-based packet scheduling: includes round robin,random, weighted round robin and weighted fairqueuing. Round robin traffic splitting is simple toimplement and may be used as a baseline forcomparison;

Packet filtering: performs packet selection using tablebased or direct hashing;

Congestion Control: acting on feedback obtained fromclosed loop control theory models such as the MultipleInput Multiple Output (MIMO) models.

Packet Loading: Similarly to bit strategies used todecide how many bits are loaded into a frame structure(a.k.a channel coding) with strategies such as water-filling, etc;

In many cases, the resulting allocation problem iscomplex and a strategy such as the Lagrange optimization maybe used to relax the constraints on the targeted optimizationfunction.

IV. PROPOSED ARCHITECTURE

Multipath streaming may negatively impact the buffer spacerequired for reordering packets, the network load withduplicate packets and the delay in reconstructing thetransmitted video. Accurate network topology, receiver andsenders´ state information must ideally be shared. This can bea complex undertaking and such measurement overheadshould be avoided in practice for the benefits of scalability andreal-time operation. As a result, our multi-path streamingmodel should perform well with inaccurate information assuggested in [9] while minimizing the network reorderingoverhead. We also stipulate that sometimes a controllableminimal amount of additional traffic overhead could be worthadding to ensure a suitable user perception quality.

In line with our SOP proposal, we suggest a receiver drivenscheduling policy as in the approach from [14]. In a way, thereceiver is calling the shots and orchestrates the streams frommultiple sources. We assume that a receiver collects a highlevel definition of the streams to be received using a protocolsuch as SOP.

Figure 4 An Architecture for SessionOrchestration

We consider that packets are lost over links or when missingtheir reconstruction deadline due to delay. Unlike existingworks, we also examine extreme situations such as when all

6

channels exhibit high losses as in wireless networks. Packetsare also dropped when one or more of their ancestors have notbeen delivered as in [11].

The proposed architecture consists of the sender and receivercooperating modules, namely, packet error modeling,bandwidth estimation, error concealment, a packet schedulerand finally traffic splitting (or a link scheduler) as shown inFigure 4. These are described in the next sections.

A. The Packet Error Model

We consider building a theoretical and practicalunderstanding of this problem. In the theoretical context, mostexisting works have adopted a two states Markov Chain,Gilbert model or Gilbert Elliot models to represent thebehavior of a lossy path [10][13] [17] [18][19]. In [21] and[22] a continuous-time version of the Gilbert Model is used.Nonetheless, the accuracy of the Gilbert models has beenquestioned in [24][25]. Though they may describe packet lossbursts, they fail to convey any feedback or long termdependencies. Despite their low complexity, they have anumber of deficiencies with serious consequences on ForwardError Correction (FEC) and network overflow analysis. Otherdrawbacks of Gilbert and Gilbert-Elliot is a lack ofcorrespondence between the bad period length and that of apacket burst loss. Hence, the model may enter a bad periodwhile no loss events are experienced. This way, the length of abad period is greater than that of packet loss bursts.

For the present study, three packet error models areexamined and explained next: the 4-state Markov chain model,the 2-state Markov chain model and the channel noise model.

Figure 5 4th Order Markov Model

1) 4-State Markov Channel ModelThis model was first adopted in VoIP related research [26]

where a series of Internet traces were analyzed using the 4-State Markov model and later implemented in the networkemulator (NETEM) project [49]. It consists of two 2-stateMarkov sub-models representing burst and gap periodsrespectively, see Figure 5. A gap is a region between two

consecutive errors whereas a burst is one where the packeterror rate is higher than a certain threshold. Note thattheoretically, any k-order Markov model may be reduced torepresent Gilbert and Bernoulli simpler models. Hence bothare seen as special cases.

The four states S= {S1 , S2 , S3 , S4 } represent successful

receipt, receipt within a burst, loss within a burst and isolatedloss within a gap respectively. A burst (in state S3 ) is seenas the longest sequence beginning and ending with a lossduring which the number of consecutive correctly receivedpackets is less than some threshold value (for video streamingapplications in IP networks a suitable limit would be 64 or128). A gap state on the other hand (as in S4 ¿ may bedefined when there is a loss rate lower than some limit orsome consecutive number of received packets.

As S4 reflects isolated packet loss, its transition

probability to S1 is p4,1=1 . If

D=p13 . p23+ p23 . p31+ p14 . p23 . p31+ p13 . p32 , weget the following state probabilities solution vectorΠ= {π1 , π 2, π 3, π4 } :

{π1=p23 . p31 /Dπ2=p13 . p32 /Dπ3=p13 . p23 /D

π 4=p14 . p23 . p31 /D

To implement this loss model into our orchestrationprotocol, we need a simpler and more intuitive way for theSOP protocol to set these probabilities. The approach taken forthe Linux implementation of the network emulator in [27]gives a user understandable meaning equivalent to the statevector Π . Five new parameters, namely, probability of

loss Pl , mean burst size B , density of packet loss

within a burst Pb , isolated loss within the good state

Pg and the mean good burst length Bg are introduced.These four first measures are straight forward to establish andare given by the expressions in (1):

{P l=π1+π2

Pb=π 1

π2+π3

Pg=π 4

π1+π 4

Bg=1p32

(1)

The average burst size B , is obtained based on the

system´s visiting the states S2 (good sub-burst within a

burst) and S3 (bad loss sub-burst) with durations

1p23

∧1

p31+ p32

respectively. Work in [28] provides a simple

7

0

11

0

P(1/0)=p

P(0/0)=1-p

P(1/1)=1-p

RV X RV Y

and intuitive way for calculating the average burst length. IfT ok→err and T err→ok are transitions from the correct

(good) to error states and vice-versa, then their absolute

difference |T ok→err−terr →ok| is bound by 1. As these

transitions have equal probabilitiesπ 2 p23+π1 ( p13+ p14 )=π 4 p41+π3 ( p31+ p32 ) , this

difference tends to 0 as n→∞ . The average burst B¿¿

and gap ( G ) lengths are the quotient of the probability ofloss or receipt to the corresponding transition probabilitiesrespectively:

pπ1(¿¿14+ p13)+π2 p23

p¿

B=π 4+π3

¿G=

π1+π 2

π1(¿¿14+ p13)+π 2 p23

¿

(2)

These new five measures are more intuitive to select andsufficient to calibrate the 4-state Markov channel model in ourevaluations than the initial model parameters we started with.

A different approach is used in [28] where burst and gapprobability distributions are defined and a numericalapproximation of the Markov matrix of probabilities is madeto obtain the steady state solution. Considering that thetransition matrix P for a m-State Markov chain has transitionspij . The solution is given when

πP=π (⇔Pnπ n=πn ) with ∑

i=1

m

π i=1 . With

n→∞ , P∞ converges towards a steady solution

limn→∞

Pn=1Π . P is then decomposed into the product of

the matrices of eigenvalues and diagonal D such that

P=U DU−1 . We have P∞=U ΑU−1 with A

´s only non-zero element being A (1,1 )=1 . The transition

probabilities pi , j: i , j∈ {1. .4 } for the 4-state machineare then estimated as shown in the case of the 2-state modelintroduced next.

2) State Markov Channel ModelFor completeness, we also consider the 2-State Markov

Channel Model in our evaluations as shown in Figure 6.Obtaining the two transition probabilities p12 and

p21 is straight forward. The value for p12 may beapproximated by the ratio of the number of transitionsfrom state S1 to S2 ( T S1→S2

¿ over the number

of received packets. Similarly, p21 is given by the ratio

of the number of transitions from state S2 to S1 (

T S2→S1¿ divided by the lost packets.

Figure 6 Two States Markov Error Model

The steady state probabilities P1 and P2 corresponding to states S1 and S2 are given as:

P1=π1=p21

p21+ p12

and P2¿ π2=p12

p21+ p12

(3)

3) Channel Noise ModelThe root cause of packet loss is congestion which is different

to that of bit errors, typically being circuit noise or jitter.Despite of this, we decided to also model packet loss as achannel noise model borrowed from the field of informationtheory.

Figure 7 Channel Packet Error Model

We regard end-to-end packet loss over a given path as anoise channel and use channel capacity theory to provide anequivalent model. A congested link traversed by packets isseen as a jammed channel where the video connection isconsidered to be the communication channel whereas the othercompeting packets are considered white Gaussian noise. Apacket is seen as a symbol in this model. Code ´1´ representswhen a symbol (video packet) is sent or received while code´0´is used when no packet is sent/received or that a packetfrom a competing traffic is sent. The resulting link model isshown in Figure 7 with the probabilities for correct receptionP(0/0) and P(1/1) whereas P(1/0) represents the probability ofpacket error “p”.

4) MPEG Mean Frame Drop MPEG-2 and MPEG-4 videos consist of independent intra-

frames (I), Predictive (P) and bidirectional (B) frames. Agroup of pictures (GOP) starts with an I-frame and may only

8

contain one of these. We denote T as the length of a GOP, seeFigure 8. I-frames are encoded independently. P-frames areencoded relatively to the previous I or P frame whereas B-frames use the latest information from a previous orsucceeding I or P frame. Consequently, the loss of a referenceI-frame results in discarding the whole GOP. The loss of a P-frame may corrupt the remaining frames in the GOP in what isreferred to as the drifting error.

Figure 8 Example of a GOP Video Structure

The event of a frame loss after k frames have been receivedhappens at a rate of (T-k)/T. The mean frame dropout (MFD)rate is the sum of two terms: the probability of losing the firstframe or I-frame and the probability of receiving framesbefore a P-frame is lost.

MFD=P2+P1∑k=1

T−1 (T−k ) p11k−1 p12

T

MFD=1−P1

1−(1−p12)T

T . p12

MFD=1−p21

p21+ p12

1−(1−p12 )T

T . p12

(4)

A suitable GOP or more precisely the parameter T in ourmodel, could be selected to ensure that the mean framedropout is under control. Recall that video quality for MPEG-4 degrades considerably above a 1% frame loss rate. Thisinformation could be proactively used by SOP to change videocoding by increasing the GOP (larger values of T) to improveuser perceived quality or PSNR. Looking at the MFD

expression, the function f (x)=1−ξ1−(1−ζ ) x

ζx where

0≤ζ ≤ 1 and 0≤ξ≤ 1 is increasing for all x>0 .It therefore reaches a minimum for x=1 and asx→∞ , giving a value of (1-ζ or 1-P1). Since the GOP

size is by definition larger than 1, we see that as the GOP sizetends to large values, the average frame dropout tends to 1-P1

or P2.Considering SOP with M different paths, these can be

modeled by M independent 2-state Markov error models. Wenext establish the corresponding MFDM . We denote j the pathindex and Pj as corresponding 2-state probabilities todistinguish between the path loss characteristics. Theprobability of losing all streams at t1 is:

(P j ,2−P j ,1 (1−p j ,11t1 ))−¿∏

j=1

M

(P j ,2−P j ,1 (1−p j ,11t1−1 ))

∏j=1

M

¿

The mean frame dropout over M paths (MFDM)is:

[(P j ,2−P j1 (1−p j ,11n ))

(T j−n)T j

]−¿∏j=1

M

[ (P j ,2−P j ,1 (1−pj,11n−1) )

(T j−(n−1))T j

]∏j=1

M

¿

¿¿

P j ,2+ ∑n=1

TM−1

¿

MFDM=∏j=1

M

¿

(5)When the same video is split across different paths such as

within our evaluation scenarios, we haveT j=T ∀ j∈[1..M ] .There are a number of problems with the above approach

however in (4). First the use of large GOP structures meansthat there are less I frames in video as there is only one perGOP. Secondly, a large GOP also suffers from error drift orpropagation. Hence if there is an I or P frame loss, then aconsiderable number of dependent packets are not decodedand consequently lost. Though the usage of a large value GOPmay be useful for reducing bandwidth, its loss effectoutweighs such benefits. The limitations of the above lossmodel in (4) and (5) are inherent to the fact that it does notconsider the peculiarities of the MPEG structure andrelationships among the frames as shown in Figure 8.Unfortunately most existing MPEG frame loss models sufferfrom oversimplification as they ignore the architecture of theGOP structure [18][51]. A distortion analysis due to packetloss is made in [52] but it only considers videos with twotypes of frames I and P. Next, we develop a two states Markoverror model that overcomes these limitations.

Given that P1 and P2 are the probabilities of losses in thegood and bad states respectively, the average packet error

P is given as:

P=p=P1π1+P2π2 (6)Having (3) and let U(Ii) and U(Pi) denote the probability that

the ith I-frame and P-frame are useful respectively. An I-frameis not useful if it is lost or corrupted. A P-frame, on the otherhand, is not useful at the receiver if it is lost or its predecessoris lost. Let also M denote smallest standard distance between Ior P frames and T denote the GOP size. N I, NP, NB representthe average numbers of packets used to transport I, P and Bframes respectively. We obtain the following expressions forU(Ii) and U(Pi):

U (I )=(1−p)NI (7)The kth P-frame can be decoded if all of its P-frame type

predecessors are received correctly including the I-frame andof course itself. We therefore have:

U (Pk )=(1− p)N I+k NP (8)

From (7) and (8), and if Ng is the number of GOPs in a givenvideo, then the expected number of correctly received anduseful I-frames and P-frames are given by:

9

{E [U ( I )]=(1−p)N I∗Ng

E [U (P ) ]= (1−p )N I∗∑j=1

TM

−1

(1−p)j NP∗N g

(9)

B-frames loss does not affect other frames but its value(usefulness) depends on other I and P frames in the GOP. Notethat the last B-frame in a GOP also depends on the next I-frame from the next GOP. We therefore have:

{U (Bk )=(1−p)N I∗(1−p)k NP(1−p)N B1≤k ≤

TM

−1

U (Bk= TM )=(1−p)2NI∗(1−p)

(TM

−1 )N P

(1−p)N B(10)

Similarly, the expected number of correctly received anduseful B-frames within a GOP is given as:

E [U (B ) ]=(M−1 )∗∑k=1

T /M

U (B k )∗N g=(M−1 )∗(1−p)NI+NB∗[(1−p)

TM

∗NP

+∑k=1

T /M

(1−p)k∗N P ]∗N g

(11)Hence the effective frame loss rate is given as:

Nj∗E ¿¿

Eloss=1− ∑j∈{ I ,P ,B}

❑

¿ (12)

B. Bandwidth Estimation Models

Available bandwidth estimation remains one of the mostrelevant research areas. Its importance to real time streamingapplications in best effort networks is beyond any doubt.Existing approaches fall into either the self-congestion classsuch as Pathrate and Pathload [30] and Initial Increasing Gap(IGI)/Packet Transmission Rate (PTR) [31] or model-based.While the first are intrusive and may themselves affect theobserved parameters, the second class contains both modelsbased on loss-related metrics [29] and delay-related metrics[32]. Though the results from [33] show that delay basedmodels tend to be more accurate, we develop preferably lossbased ones due to the loss data availability at the SOP receiver.Delay information is difficult to obtain in a real network dueto the need for clock synchronization using GPS for example.As such, delay models will not be discussed further.

1) Bandwidth for a Channel Noise ModelConsidering a channel noise model, we estimate the

available bandwidth Bw for a single multimedia stream.

According to Shannon´s channel capacity model for a linkwith rate r, information theory teaches us that bandwidth isgiven in terms of average information or entropy of randomvariables (RVs) X and Y as in:

Bw=H t ( x )−H t ( x / y )=r [H ( x )−H ( x / y )]Where H (x / y ) is the conditional entropy of RVs X and

Y. We then have:

{ H ( x )=−∑i=0

1

[P(xi)lo g2P(x i)]

H ( x / y )=−∑j=0

1

[P( y j)∑i=0

1

[P(x i / y i)lo g2P(x i / y i)]]

(13)

r−1¿+ log2(r−1)+r log2r

¿Bw=−r log2¿

For large rate values (r ≫1¿ , we have:

Bw=log2r+rH ( p )=r (H (p )+log2r

r) (14)

2) Simple Loss based Bandwidth EstimatorWe suggest a simple additional model that can be calibrated for use as bandwidth estimator as given next.

Bw=r (1−p)αwith α>0 (15)

C. Packet Scheduling

This module decides which packet goes next on one of thechannel transmission links. Note that it does not select theactual channel to use as this is the responsibility of the nextmodule in the proposed SOP architecture. The followingstrategies are selected for evaluation in this work.

a) No Scheduling: when no scheduling is applied, packetsare sent according to their natural order in the stream.This strategy is used as a baseline for comparing theothers;

b) Deadline-based: packets are given deadlines and then sorted according to this information in an ascending order.

c) Priority based: packets are ordered according to a pre-established cost function. Though many cost functionsmay be tested, we consider a simple expression thatcombines the damage caused if such packet is nottransmitted, the deadline and the importance of theframe according to its type in reconstructing the stream.Examples of simple cost functions are:

{Δ1 (n )=

importance(n)damage(n)

Δ2 (n )=importance (n )∗damage (n )

deadline (n )

Δk (n )=[ I k .importance (n )+Dk∗damage (n ) ]

[G k .damage (n ) ](16)

with I k ,Gk ,D k≥0 and importance, damage and

deadline respectively, being normalized weights ∈ [0,1 ] ;

10

d) Distortion splitting: here we allow some video distortionto take place and replicate I-frames and P-frames whichare sent instead of normal B type frames.

More specifically, let´s denote δ n the scheduling action

carried on packet n, ri(t) the media rate over path i, t na

packet n arrival´s time, t nl its deadline for reconstruction,

ε>0 and π∈Π the scheduling policy.

δ n={ c j if packet nissentpath

j

0 if packet n is droppedThe successful decoding of a packet n, χn(π )=1 when

packet n has been sent ( δ n≠0 ), it does not miss its

deadline for decoding ( t na≤ t n

l+ε ) and all frames needed

to reconstruct n have been correctly received within theirdeadlines. Otherwise χn is null. The overall benefit is

then given as the weighted sum Δ (π )= ∑n: χn( π)=1

ωn .

Hence, an optimal scheduling policy π*is defined as thesolution for the optimization problem [11]:

Δ (π¿ )= max∀ n: χn (π )=1

ωn (17)

D. Traffic Splitting

This module selects for each packet the correspondingoutgoing link. For this, we compute channel proportionality,represented by the positive α i values, as factors of datasuch as loss, bandwidth information, delay, etc., orcombinations of these. We preferably will use simplemeasures that can easily be calculated at the receiver. Thefollowing strategies are considered:

a) Round robin packet splitting (RR): this is simple toimplement and is used as a mere baseline for comparingnew approaches.

b) Shorter delay first: we exploit delay differences amongpaths and assign more packets on those with shorterdelays. This approach has also been taken in previousworks such as [21]. We take the RTT information as anapproximation of one way delay (OWD). This is in linewith the proposal from [23] where MPLOT intelligentlymaps packets that are not required immediately to pathswith longer delays and those that are needed soon toshort delay paths.

c) Less for lossy paths: here the α i values are selectedto be lower in the case of paths suffering from high lossvalues. A simple strategy would be to schedule packetsin an inversely proportional level to each path loss rate.

d) Redundant Splitting: unlike previous load splittingworks, we suggest the use of a load (or what we call aBandwidth Margin) to overcome losses even if it meantoccupying more aggregate bandwidth. Hence,considering the traffic splitting vector[∝1 ,∝2 ,.. ,∝k ] , instead of having a traffic

partition where ∑i

∝i=1 we consider a partition

sum larger than the video rate, where ∑i

∝i≥1 .

Note, that this strategy is only used when there isbandwidth available to accommodate such increase. Wewant to see if this may contribute with any benefit toSOP. The idea is that packets are duplicated (on aseparate path d) with a probability that is a θk pd

where pd is a parameter representing the basic

duplication priority level and θk being the prioritylevel of such packet. We pay particular attention to theimpact of duplicating I-frames and P-frames.

E. Error Concealment

Advanced concealment of transmission errors may beachieved by estimating using interpolation (or extrapolation)the lost information or missing pixels. The key observationhere is the existence of a significant amount of correlationalong the spatial and temporal video dimensions.

To build the receiver´s perceptual quality a.k.a. quality ofexperience (QoE), a simple error concealment technique isused, whereby a k-length burst of lost frames i+1, i+2,…,i+kis replaced by the last frame correctly received i. This way, amore representative PSNR metric is computed. This is adefault action during the experiments made in this work.

V. MODELING SOP

REFERENCES

[1] Thomas Erl, Service-Oriented Architecture (SOA): Concepts,

Technology, and Design, Prentice Hall, ISBN-10: 0131858580, 2005.

[2] www.Octoscope.com

[3] P2P Streaming for Mobile Nodes: Scenarios and Related Issues, Internet

Draft, http://tools.ietf.org/html/draft-lu-ppsp-mobile-03

[4] Shojiro NISHIO , Opening up New Vistas on Advanced Multimedia

Content Processing, New Generation Computing, Ohmsha, Ltd. andSpringer-Verlag 18(2000)295-303, 2000.

[5] https://datatracker.ietf.org/wg/ppsp/charter/

[6] Service Oriented Architecture: A Field Guide to Integrating XML and

Web Services, Thomas Erl, Prentice Hall PTR Upper Saddle River, NJ,USA, ISBN:0131428985, 2004

[7] S. Weerawarana, F. Curbera , F. Leymann , T. Storey and D. F. Ferguson,

Web Services Platform Architecture: SOAP, WSDL, WS-Policy, WS-Addressing, WS-BPEL, WS-Reliable Messaging, and More, ISBN-10:0131488740 , 456 pages, Prentice Hall, 2005

[8] Xunqi Yu, James W. Modestino and Ivan V. Bajic, Modeling and

Analysis of Multipath Video Transport over Lossy Networks UsingPacket-Level FEC, Proceedings of DMS'05, pages 265-270, Banff, AB,Canada, September 2005.

[9] J. Apostolopoulos, T, Wong, S. Wee and D. Tan, On Multiple

Description Streaming with Content Delivery Networks, In the IEEEINFOCOM, New York, June 2002.

[10] L. Golubchick and J. C. S. Lui, Multi-path Streaming: Is it worth the

Trouble?

[11] D. Jurca and Pascal Frossard. Video Packet Selection and

Scheduling for Multipath Streaming, IEEE Transactions on Multimedia,Vol. 9, No. 3, pages 629-641, 2007.

[12] R. Rubinstein, J. Kurose and D. Towsley, Detecting Shared

Congestion of Flows via end-to-end measurements. In the ACMSigmetrics Conference, Santa Clara, California, June, 2002.

11

[13] B. Abdouni, W. C. Cheng, A. L. Chow, L. Golubchik, W. J. Lee and

J. C. S. Lui, Muti-path Streaming: Optimization and performanceevaluation, Proceedings of the SPIE Conference on MultimediaComputing and Networking, vol 5680, San Jose, California, pages 216-2276, 2005.

[14] T. Nguyen and A. Zakhor, Multiple Sender Distributed Video

Streaming, IEEE Transactions on Multimedia, Vol. 6, No.2, 2004.

[15] P. A. Chou and Z. Miao, Rate-distortion Optimized Streaming of

packetized media, IEEE Transactions on Multimedia, Vol. 8, No. 2,pages 390-404, 2006.

[16] K. Chebrolu and R. Rao, Bandwidth aggregation for real-time

applications in heterogeneous wireless networks, IEEE Transactions onMobile Computing, vol. 5, no. 4, pp. 388–403, Apr. 2006.

[17] A. L. Chow, L. Golubchik, J. C. S. Lui and W. J. Lee, Multi-path

streaming: Optimization of load Distribution, Performance Evaluation,pages 417-438, 62 (2005).

[18] I. Lee and L. Guan, Reliable Video Communication with Multi-path

Streaming using MDC, 2005.

[19] A. L. Chow, H. Yang, C. H. Xia, M. Kim, Z. Liu and H. Lei, EMS:Encoded Multipath Streaming for Real-time Live StreamingApplications, The 17th IEEE International Conference on NetworkProtocols (ICNP'09), Princeton, New Jersey, October 2009.

[20] Bing Wang, Wei Wei, Zheng Guo and Don Towsley, Multiple LiveStreaming via TCP: Scheme, Performance and Benefits, Wang, B., Wei,W., Guo, Z., and Towsley, ACM Transactions on Multimedia ComputerCommunications Appl. 5, 3, Article 25, August 2009.

[21] M. Kurant, Exploiting the path propagation time differences inmultipath transmission with FEC,” in Proceedings of IEEE INFOCOM,April, 2009.

[22] H. Y. Hsieh and R. Sivakumar, A transport layer approach forachieving aggregate bandwidths on multi-homed mobile hosts. InProceedings of the ACM Annual International Conference on MobileComputing and Networking (MobiCom), 2002.

[23] V. Sharma, S. Alyanaraman, K. Kar, K. Ramakrishnan and V.Subramanian, MPLOT: A transport protocol exploiting multipathdiversity using erasure codes. In Proceedings of the IEEE Annual JointConference of the Computer and Communications Societies INFOCOM,2008.

[24] X. Yu, J. W. Modestino and X. Tian, The Accuracy of GilbertModels in Predicting Packet-Loss Statistics for a Single-MultiplexerNetwork Model, In Proceedings of the 24th Annual Joint Conference ofthe IEEE Computer and Communications Societies (INFCOM), pages2602 – 2612, vol. 4, 2005.

[25] X. Yu, J. W. Modestino and X. Tian, The Accuracy of MarkovChain Models in Predicting Packet-Loss Statistics for a SingleMultiplexer, IEEE Transactions on Information Theory, pages: 489 -501 , Volume: 54 Issue: 1, 2008.

[26] Indepth: Packet Loss Burstiness, Available athttp://www.voiptroubleshooter.com/indepth/burstloss.html

[27] S. Salsano, F. Ludovici and A. Ordine, Definition of a general andintuitive loss model for packet networks and its implementation in theNetem module in the Linux kernel, Technical Report, Version 2.2,University of Rome, Available at:http://netgroup.uniroma2.it/NetemCLG, September 2010.

[28] L. Estrada, D. Torres, H. Toral, Characterization and Modeling ofPacket Loss of a VoIP Communication, World Academy of Science,Engineering and Technology 66, pages 926-930, 2010.

[29] K. Salamatian, B. Bruno, T. Bugnazet, Interpretation of lossesobserved on the Internet by inferring traffic characteristics, in:Proceedings of the DIMACS Workshop on Internet and WWWMeasurement, Mapping and Modeling, Rutgers University, Picscataway,NJ, USA, 2002.

[30] Ravi Prasad, Manish Jain and Constantinos Dovrolis, EvaluatingPathrate and Pathload with Realistic Cross-Traffic, Talk at bandwidthestimation workshop 2003.

http://www.caida.org/workshops/isma/0312/slides/rprasad-best.pdf,2003.

[31] N. Hu and P. Steenkiste, Evaluation and characterization of available

bandwidth probing techniques, IEEE JSAC, Volume 21, No. 6, pages879-894, 2003.

[32] X. Hei, D.H.K. Tsang, B. Bensaou, Available bandwidthmeasurement using Poisson probing on the Internet, in: Proceedings ofthe 23rd IEEE International Performance Computing andCommunications Conference, IPCCC 2004, Phoenix, AZ, pages 207–214, 2004.

[33] X. Hei, B. Bensaou and D.H.K. Tsang, Model-based end-to-endavailable bandwidth inference using queuing analysis, ComputerNetworks 50, pages 1916–1937, 2006.

[34] H. Sanneck and G. Carle, A Framework Mo del for Packet LossMetrics Based on Loss Runlengths, Proceedings of the SPIE/ACMSIGMM Multimedia Computing and Networking Conference 2000(MMCN 2000), San Jose, CA, January 2000.

[35] http://en.wikipedia.org/wiki/Mashup_%28web_application_hybrid%29

[36] Alan B. Johnston, SIP: Understanding the Session InitiationProtocol, Second Edition. Artech House. ISBN 1580531687, 2004.

[37] H. Schulzrinne, S. Casner, R. Frederick and V. Jacobson, Request

for Comments 3350, RTP: A Transport Protocol for Real-TimeApplications, July 2003.

[38] T. Friedman, R. Caceres and A. Clark, RTP Control Protocol

Extended Reports (RTCP XR), Request for Comments 3611 November2003.

[39] M. Handley, V. Jacobson and C. Perkins, Request for Comments

4566, SDP: Session Description Protocol, July 2006.

[40] György Dán, Viktória Fodor and Gunnar Karlsson, Analysis of thePacket Loss Process for Multimedia Traffic, in Proceedings of ICTSM12, pages 83-94, July 2004.

[41] I. Cidon, A. Khamisy and M. Sidi, Analysis of packet loss processesin high speed networks, IEEE Transactions on Information Theory,volume IT-39, number 1, pages 98–108, 1993.

[42] Denise M. Bevilacqua Masi, Martin J. Fischer and David A. Garbin,Video Frame Size Distribution Analysis, The TelecommunicationsReview, pages 74-86, 2008.

[43] D. Gross and C. M. Harris, Fundamentals of Queuing Theory. Wiley,New York, (1998).

[44] L. N. Singh and G. R. Dattatreya, A novel Approach to ParameterEstimation in Markov-Modulated Poisson Processes, Department ofComputer Science The University of Texas at Dallas Richardson, Texas75083-0688 http://www.utdallas.edu/~datta/papers/mmpp.pdf.

[45] L. P. Raj Kumar, S. Kumar, D. M. Reddy and M. R. Perati,Analytical Model for Performance Study of the Switch under SelfSimilar Variable length Packet Traffic, Proceedings of the WorldCongress on Engineering and Computer Science 2010, Vol 1, pages 243-247, WCES 2010.

[46] H. Heffes, A Class of Data Traffic Processes: Covariance

Characterization and Related Queuing Results, BSTJ, Volume 59,Number 6, pages pp. 897-929, 1980.

[47] Daniel P. Heyman and David Lucantoni, ?Modeling Multiple IP

Traffic Streams with Rate Limits,? IEEE/ACM Transactions onNetworking, Vol. 11, No. 6, pages 948-958, December 2003.

[48] A. Andersson and T. Ryden, Maximum Likelihood Estimation of a

Structured, MMPP with Applications to Traffic Modeling, 13th ITCSpecialist Seminar, Monterey, CA, 2000.

[49] Soshant Bali and Victor S. Frost, An Algorithm for Fitting MMPP toIP Traffic Traces, IEEE Communications Letters, Vol. 11, No. 2, pages207-209, February 2007.

[50] S. Hemminger, Network Emulation with NetEm, Proceedings of the6th Australia's National Linux Conference LCA´2005, 2005.

[51] I. Lee, Y. He and L. Guan, Centralized P2P Streaming with MDC,IEEE 7th Workshop on Multimedia Signal Processing, pages 1-4, 2005.

12

[52] L. U. Choi, M. T. Ivrlac,E. Steinbach and J.A.Nossek,IEEE International Conference on Image Processing, ICIP´2005. pages 189-92, 2005.

[53] R.L. Lagendijk, Jacco Taal and Alan C. Bovik(Editor), Layered, Scalable, and MDC Coding(Synthesis Lectures on Image, Video, & MultimediaProcessing), Published by Morgan & ClaypoolPublishers, ISBN: 1598292366, 2008.

Djamel F. Hadj Sadok (M’95) is an Associate Professor at Centro deInformatica, The Federal University of Pernambuco. He has been working inthe area of Computer Networks, traffic classification, routing and contentdelivery networks. The author received a Ph.D. in Computer Sciences fromthe University of Kent at Canterbury in 1990.He has been with UFPE since 1993 where he currently leads a group ofresearchers working in close collaboration in a number of projects with theindustry including Ericsson Research in Sweden.

Dr. Djamel Hadj Sadok holds a number of patents, acts as a reviewer for anumber of high level journals in the area and takes part in the technicalcommittees of many national and international conferences.

13