Autonomous QoS Management and Policing in Unmanaged Local Area Networks

The high increase of bandwidth-intensive applications like high definition video streaming in home and small office environments leads to QoS challenges in hybrid wired/wireless local area networks. These networks are often not QoS aware and may contain bottlenecks in their topology. In addition, they often have a hybrid nature due to the used access technology consisting of, for example, Ethernet, wireless, and PowerLAN links. In this paper, we present the research work on a novel autonomous system for hybridQoS in local area networks, calledQoSiLAN, which does not rely on network infrastructure support but on host cooperation and works independently of the access technology. We present a new QoS Signalling Protocol, policing and admission control algorithms, and a new lightweight statistical bandwidth prediction algorithm for autonomous resource management in LANs. This new QoS framework enables link based, access-medium independent bandwidth management without network support. We provide evaluation results for the novel bandwidth prediction algorithm as well as for the QoSiLAN framework and its protocol, which highlight the features, robustness, and the effectiveness of the proposed system.


Introduction
Quality of Service (QoS) becomes more and more relevant to consumer networks, due to the increased usage of High Deinition (HD) video streaming, IPTV applications, and Voice over Inter net Proto col (VoIP) communicat ion.Nowadays, video streaming is already topping the list of consumer traic [1] with more than 35% of mobile devices using WiFi connections and even Ultra HD Internet video streaming is evolving.Regarding the consumer Internet video traic, video streaming is expected to grow further massively from 19% as of 201 to 53% in 2017, as forecasted by Cisco's Visual Network Index [2].Currently, most consumer routers and switches support QoS management in local area networks (LANs) on the basis of the Diferentiated Services (DifServ).Additionally, there are already consumer systems in the market, which perform self-organised traic low identiication and classiication for prioritised packet scheduling [3].his means that, at best QoS is only available for traic passing the gateway and no QoS guarantees are possible as compared to the QoS model the Resource Reservation Protocol (RSVP) is based on [ ]. herefore, LAN internal traic remains completely unmanaged, which afects mainly intra-LAN communication like video and audio streaming from Media Network Attached Storage (Media-NAS) devices, which is a common operation in networked consumer households.Additionally, IntServ protocols, like the RSVP, usually cannot no be found in consumer network hardware.Even Microsot has dropped support for RSVP in Windows operating systems with the release of WindowsXP [5] in 2001.State-of-the-art LAN management protocol frameworks like H.622 [6] or UPnP-QoS [7] and QoS NSLP [8] emphasise the need for QoS management in LANs.hey propose strategies, which rely on network support for session-based QoS to enable bandwidth reservation.
Since the broadband connection and wireless links in a LAN are still bottlenecks for services with high bandwidth requirements, a QoS scheme is required to protect real-time service lows from congestion.Internet providers solve the problemonthelastmilebyusingseparatemulticastandQoS enabled Internet links for their IPTV delivery.his provider control ends at the router.Problems in a LAN arise mainly when multiple user traic lows utilize a bottleneck link in t h en e t w o r ka n do n eo ft h el o w sh a sr e a l -t i m em e d i aQ o S requirements.his problem in unmanaged network scenarios is still a research challenge in state-of-the-art research and t e c h n o l o g y .hec u r r e n tt r e n do na p p l i c a t i o ns i d et ou s e rate-adaptive streaming does not solve the problem either.Rate-adaptive streaming aims at reacting on congestion and bandwidth degradation by stepping down the resolution and bandwidth [9], instead of preventing the congestion from not real-time dependent sources.
To overcome these issues we propose a QoS in local area networks (QoSiLAN) framework, which enables QoS service guarantees in the forms of end-to-end bandwidth reservation between the hosts within the LAN, as well as for traic lows, passing the Internet gateway. he proposed QoSiLAN framework addresses the problem of QoS management in unmanaged LANs, without the essential dependence on network infrastructure support.Its goal is to manage the real-time traic autonomously with a host-based, cooperative approach.he QoSiLAN framework is designed to work autonomously, taking into account the assumption not to rely on network cooperation, but to support it.Since the QoSiLAN framework completely relies on host cooperation, multiple key technology modules are needed on the hosts in the LAN to support this architecture as depicted in Figure 1.For the irst time, this paper describes all modules and their interaction algorithms in detail.

Contribution
In this section the main contributions of our research are presented and discussed.hese are the statistical classbased bandwidth prediction (SCBP) algorithm and the QoSiLAN policing and admission modules.hese two key solutions provide the core functionality to the QoSiLAN framework, since the bandwidth prediction is essential to support autonomous traic management, which works independently of support from media applications.Additionally, the policing and admission control module processes all data from the other key technology modules to enable smart policing decisions.A complete overview is shown in Figure 1, where all modules are presented.Especially for selforganisation, the physical topology map of the LAN must be discovered, including a bottleneck bandwidth detection for the path, which is assisted by the iperf tool [10] (topology discovery).Also, the traic must be monitored, analysed, identiied, classiied (Flow Identiication and Classiication), and predicted in an autonomous manner (Flow Bandwidth Prediction).For signalling, a new protocol is needed to enable QoS management communication between the hosts (QoS Signalling Protocol).For this purpose, a signalling solution built on top of the latest IETF recommendation is presented, including a critical examination on the overhead and scalability issues this approach causes.he policing and admission control enables the smart provisioning of resources, which employs cooperative traic shaping (traic shaping) on all hosts in the LAN to enforce bandwidth preservation (policing/admission control).In the following, these modules are presented in detail.
2.1.Topology Discovery.First, to manage the traic on a p e rl i n kb a s i s ,ac o m p l e t em a po ft h el o c a ln e t w o r km u s t be known.he map reveals the link layer interconnection between hosts, switches, and access points, which is needed for link based resource reservation.he QoSiLAN framework does not rely on network core element support.his means that only end-hosts are involved in the signalling and enforcement processes.Although the network is not involved in the QoS management, the network is actively monitored.To be able to control the traic in the LAN, its physical topology must be known.his enables the QoSiLAN framework to manage resources link-based and to react on topology changes and congestion directly.
his is accomplished using the Link Layer Topology Disc o v e ry( L L T D )p r o t o c o l .heL L T Dp r o t o c o lw a sp r e s e n t e d i r s t l yb yB l a c ke ta l .[ 1 1 ] .I ti sal a y e r2t o p o l o g yd i s c o v e r y protocol, which was developed within the Microsot Rally program [12].Later the LLTD protocol was introduced and distributed with Microsot Windows since 2006. he protocol speciication was published by the Microsot Corporation [13].It works on wired IEEE 802.3 and wireless IEEE 802.11 networks.It is based on probing packets sent by a central entity,called"mapper,"toeachLLTDsupportinghost,called "Responder," in the network.All end devices implementing theprotocolappearinthenetworktopologymapthemapper builds.he protocol aims at investigating the address tables of Ethernet switches.Since one cannot read the Address Information Table (AIT) in most consumer switches, the mapper sends Emit packets to the hosts to make them send Probe packets to the switches using spoofed MAC addresses, according to the connections reasoning technique presented by Sun et al. [1 ].Using this algorithm the mapper trains and probes each switch's address tables to ind out if the hosts are attached to the same or diferent switches.According to the IEEE 802.3 speciication a switch forwards packets with an unknown MAC address to all ports.Once it has received a packet with the MAC address as source one time it subsequently forwards incoming packets with this MAC as the destination address only to this single port. he other hostsinthenetwork,whichhavealsolistenedtothetraining packet also answer to the mapper if they also have received the packet or not.For this purpose, the LLTD protocol uses MAC address spooing with addresses starting with 00-0D-3A.his is a dedicated MAC address range, reserved by Microsot.In order to make use of the LLTD capabilities, we completely reimplemented the protocol and the mapper logic and integrated it into the QoSiLAN framework.

Flow Identiication and Classiication.
To enable the QoSiLAN to work independently of applications, the traic lows must be identiied and classiied in an autonomous manner.his is accomplished using our solution, the enhanced Statistical Protocol Identiication (eSPID) algorithm [15].We further developed the Statistical Protocol Identiication (SPID) algorithm [16] to identify audio/video streams reliably.he SPID algorithms are based on the Kulback-Leibler divergence (KLD): using twelve diferent statistical measures of traic behaviour, like bytes-per-direction, number-of-direction-changes, bytefrequency, and nine more. he eSPID algorithm needs to learn a protocol from 30 preclassiied sample lows, as evaluated in our previous work [15].he twelve statistical measures are applied to the sample lows and stored as probability distribution arrays in a database.his makes it robust against UDP/TCP port changes and up to a certain degree against encryption. he KLD in (1) is a logarithmic measure of the relation between the relative frequency of the observed () to the trained lows (),summedforeach attribute measure. he ( ‖ )is matched to a database of learned ( ‖ ) from other protocols.he protocol with the smallest divergence ( ‖ ) is then identiied.he distance represents the probability of matching.he eSPID algorithm allows for robust identiication of lows ater 2 0p a c k e t s ,a ss h o w ni no u rp r e v i o u sw o r k[ 1 5 ] .he r e f o r e , it is a fast and reliable method for near real-time media identiication. he outgoing traic is identiied and classiied using the individual protocol name and the media type.hese media types are, namely, audio, video, or unknown if no matching was found.New real-time protocols can be learned by the system in a semisupervised way.Needed are at least 30 sample lows, which are identiied by the user.F o rt h e s et h eK L Di sc a l c u l a t e da n dt h er e s u l ti ss t o r e d in the database.he identiication probability of encrypted protocols is not substantially less.As the content based measures have no statistical information due to the entropy of the encrypted content, only the behaviour focused statistical measures apply.his does not need to essentially lower the recognition probability, especially not for diferent protocols, but for the content diferentiation within a protocols.

Statistical
Class-Based Bandwidth Prediction.Once a low is identiied and classiied, the needed resources must be estimated in order to reserve them.his is enabled by a novel statistical class-based bandwidth prediction (SCBP) algorithm, which we developed for the special purposes of theQoSiLANframework.AsshowninFigure1,itisapplied ater the eSPID process and the results are fed directly into the policing and admission control module.he statistical class-based bandwidth prediction (SCBP) algorithm is our novel approach for Internet video traic bandwidth prediction for single streams.It addresses the problem of diferent transmission characteristics of state-ofthe-art streaming technology and the resulting low bandwidth forecasting accuracy.Nowadays, the traditional continuous streaming characteristic, for example, known from the Real-Time Transport Protocol (RTP), is highly underrepresented.Instead, HTTP based streaming technology like HTTP Dynamic Streaming (HDS) [17], HTTP Live Streaming (HLS), and the Real-Time Media Protocol (RTMP) is very common.hose streaming protocols support the feature of chunkeddatatransfer .hisallowstheCDNstosigniicantlysave their traic costs, since only those parts of a video are transferred, which are actually consumed by the user.If a user stops a video the transfer also stops.If a user moves to a position at the end of a video, there is no need to transfer the parts before.So, the data is not transferred continuously according to the encoded bit rate, but in chunks, with a burst characteristic, oten with several seconds of traic silence in between the bursts.he goal for our SCBP algorithm is to predict the average bandwidth requirement of a low, for the irst minute of transfer, ten seconds ater start.As shown in Figure 2, ater 60 s the average of the last 60 s is chosen to update the reservation state.hese values were chosen to fulil system requirements for the QoSiLAN framework.So, ater 60 s the reservation update is not using the SCBP values, but the measured average throughput of the low from the last 60 s.
Since the QoSiLAN framework needs to communicate with all hosts in the network to coordinate the resource reservations, it causes high signalling efort.herefore, the reservation state live time was evaluated to be 60 s, to avoid too frequent state updates and therefore network congestion from signalling.hat is why the SCBP algorithm is designed to predict the average bandwidth consumption for single multimedia streams for the irst 60 s ater 10 s recording time.According to our evaluations a minimum of 10 s recording time from the start of the low was evaluated to be required to catch the lows characteristic and to enable proper predictions.As depicted in Figure 2, ater 60 s of measurement, t h er e s e r v a t i o ns t a t ei su p d a t e dt ot h em e a s u r e da v e r a g e bandwidth consumption ( 60 ) of the last 60 s.Within this paper we present our simple but highly precise prediction algorithm and validate its applicability using evaluations, as presented in Section 5.1.
he SCBP algorithm classiies the traic into six diferent traic classes (A-F), according to their behaviour within the irst ten seconds of transmission, as shown in Figure 3.To achieve the highest accuracy, the measurement must start at the very beginning of a connection.During the observation time throughput measurement samples are collected, where is an even number.From this we get the set , containing bandwidth measurement samples.For better low characterisation we divide the set of into two subsets: ⊂fl { 0 ,..., /2 } and ⊂fl { /2+1 ,..., }, where we calculate the maximum ( ,max and ,max )a s well as the average ( and ) values.In addition also max is determined as the maximum value in the set of . he prediction is calculated at time and forecasts the average bandwidth consumption for the whole irst 60 s of transmission.he sampling interval is deined as = / .he evaluations are based on =1 0s/10 samples =1s. he ratio , deined in   aburstatthebeginningandnodatawithinthelasthalfofthe observation interval of 10 s. hrough evaluations, we validated that the -values assigned to the classes A-F were applicable to the ranges as best deined in Table 1.Hence, the classes relect the bandwidth characteristic, including the start-up behaviour of the low in a simple way.For the class-range assignment evaluations,themeasurederrorvalueswereputintoadensity histogram to identify the ranges of similar deviations.
We designed the prediction formula, shown in for the balance of simplicity and accuracy.It takes into account the application in embedded systems with less computation power, which also saves battery life on mobile systems. he prediction formula in (3) implements the correction factor , which relects the class characteristics.he value for was found using evaluations, as carried out in Section 5.3. he QoSiLAN protocol uses, like other common QoS protocols, a sot state mechanism.his means, the hosts keep a reservation state as long as a time-out has not been reached.To keep states alive, a reservation refresh message (QoSiLAN RESERVE), with the same parameters as the irst one, must be sent before the time-out is reached.
QoSiLAN hosts (QH) inform the QM about their traic situation and request for resources.he QM asks all QH for cooperation for currently active QoS states to establish the bandwidth reservations commonly.
he NSIS General Internet Signalling Transport (GIST) protocol [20] also brings along a host discovery feature, which is used to detect and identify the QoSiLAN enabled hosts i nt h en e t w o r k .heQ o S i L A NM a n a g e r( Q M )s e l e c t i o ni s performed by the smallest switch/hub-hop distance to the router, evaluated using the LLTD protocol.If the router itself is QoSiLAN enabled, it announces a distance of zero.If the router is not QoSiLAN enabled and two QH have the same distance, the QH with the longest uptime period is selected.
(i) Reservation (QoSiLAN RESERVE).her e s e rv a t i o nm e ssages use the NSLP common header, as all GIST NSLP objects [20] do. he message contains a sequence number (RSN), a REFRESH PERIOD, and a BOUND SESSION ID. he QSPEC object deines the QoS information parameters.It lists the requested resources using the TMOD-1 parameter, which contains the peak data rate for traic shaping, and the Excess Treatment parameter, which deines the shaping policy, as deined in [21].
In addition, we introduced the new Reservation Path Parameter for IPv4, to communicate the ive-tuple, which deines what low the reservation is for, as shown in Figure .he irst 32 bits are deined according to the QSPEC parameter header [21]. he source and destination IPv addresses aswellasthesourceanddestinationportsandthetransport protocol ID are included.For IPv6 a similar parameter header format with 6 -bit IP address ields needs to be deined accordingly.
Reservations are deleted either on sot state time-out or by sending a QoSiLAN REQUEST message using the BOUND SESSION ID without a QSPEC object deined.
(ii) Response (QoSiLAN RESPONSE).he response message is intended to report success or error codes to the requesting  node. he success case is indicated using the error=0x00 value.If a reservation state is rejected, a negative acknowledgement is signalled through a error=0x47 value set in the INFO SPEC header.he RSN object is taken from the corresponding QoSiLAN REQUEST message to match the response to a request.
(iii) Notify (QoSiLAN NOTIFY).hen o t i f ym e s s a g ei s intended to report signiicant best efort traic low statistics to the QoSiLAN Manager (QM).his is needed to keep the QM informed about the current state of the network, especially when there is a lot of traic occupying which is not classiied for QoS. he message structure is similar to the corresponding QoSiLAN REQUEST and QoSiLAN RESPONSE messages as described before.he QoSiLAN NOTIFY message also uses the Reservation Path Parameter for IPv within the QSPEV header to announce detected lows.
2.4.2.QSLP-LAN Signalling Procedure.In Figure 5(b) the s i g n a l l i n gi ss h o w nt ob ee x e m p l a r ya n du s e dt oe s t a b l i s h bandwidth reservation in a simple LAN scenario.In this LAN one QoSiLAN Manager (QM) host and three client QHs are connected using two diferent switches.All of them, except the switches, are QoSiLAN aware. he initiator (Host 1) sends a reservation request message (QoSiLAN RESERVE) to the QM. he reservation request shall contain at least the physical addresses and the IP addresses of the two communicating parties and the requested resource, the predicted bandwidth to reserve. he QM analyses the location of the QHs and sends sophisticated reservation requests (QoSiLAN RESERVE) for all other nodes (Host 1 ,H o s t2 ,a n dH o s t3 )i nt h en e t w o r k ,b a s e du p o nt h e L A Nt o p o l o g ya n dt h eL A N ' st r a i cs t a t u sk n o w l e d g e ,t o encourage all QHs to obey the limitations for the afected physical links.In return, the nodes acknowledge the request to the QM (QoSiLAN RESPONSE), and the QM reports (QoSiLAN RESPONSE) the result/success to the initiator (Host 1) at the end.To cover the case where both the sending and receiving host do not support QoSiLAN, but the QM is a gateway and detects the low, another example is given.As shown in Figure 5(a), on resource reservation initiation by a gateway, enabled as QM, no request/response communication to another initiation node is required in contrast to the host initiation case, depicted in Figure 5(b).In both cases the reservation requests/responses are sent to/from each QH in the network, initiated from the QM.To avoid unnecessary signalling efort, the hosts do not report their traic statistics regularly, but only triggered by events.In case of signiicant best-efort traic detection, a QH reports the monitored bandwidth of the stream to the QM using a QoSiLAN RESPONSE message.Details on the policing and admission control algorithm are described in Section 2.5.3.

Policing/Admission Control.
As shown in Figure 1, the policing and admission control algorithms are the key components to enable autonomous QoS management.Within this paper we propose appropriate policing and admission control schemes to enable the QoSiLAN framework.he resource management within a LAN requires detailed knowledge about the available resources on all links between the LAN entities.Only if the link layer topology and the capacities of its links are known, the QoSiLAN framework can autonomously manage the resources.In the following, we describe the management procedures to acquire the required information.
2.5.1.Resource Discovery.One host in the network, which p r e f e r r e dt h eg a t ew a y ,f u l i l st h er o l eo ft h eQ M .A ss u c hi t mapsthenetworktopologyandactsasresourcecoordinator. he mapping process is executed each time a new host is discovered in the network.he host discovery is based on broadcast and ARP packet monitoring. he map, generated by the LLTD Mapper module, described in Section 2.1, relects the link layer topology between the hosts, switches, hubs, and access points in the LAN.Additionally to the topology itself, the bottleneck capacity of each link needs to beevaluated.hisisdoneduringnetworkorlinkidletimes. he mapper advises the hosts to measure the bottleneck bandwidth of their links by active probing.he QM gathers the results from the QHs and adds this information to the connection information table within the topology map.
2.5.2.Policing Procedure.Each QH in the network continuously monitors its outgoing traic.To make autonomous QoS policing possible, the QM needs to gather and maintain all information about the LAN and its links, by using the LLTD topology mapper module.his includes the physical paths and measures their bottleneck bandwidth .As depicted in Figure 6, the eSPID module analyses on each QH the outgoing lows to identify data with QoS requirements, like video and audio transport.Once a host discovers, for example, a VoIP communication, it estimates the bandwidth requirements for this particular low using the SCBP algorithm.he QoSiLAN MBAC algorithm decides upon the admission of the reservation, taking into account the diferent measures.If the admission was granted, a QoSiLAN REQUEST message is sent to the QM to request the bandwidth reservation for the detected resources.A QoSiLAN REQUEST message contains a ive-tuple: the sender and receiver IP and port addresses as well as the estimated bandwidth.
he QM receives the QoSiLAN REQUEST message and checks the map and the available resources for the low's route within the LAN.If the requested resources are available, the QM sends individual QoSiLAN REQUEST messages to each QH in the network.hese messages contain a list of QSPEC parameters containing the low's ive-tuples and their bandwidthlimitstobeobeyedbythereceiver.hesemessagesare generated individually for each QH, since the afected links on a route to another QH in the LAN difers, depending on the location in the LAN's topology.Each QH checks the request for validity and if the requested resources are available locally.In any instance of an error or conlict, the QH sends a negative acknowledgement QoSiLAN RESPONSE message back.If the request is accepted, the QH sends a positive acknowledgement QoSiLAN RESPONSE message back to the QM.
Once all QHs in the network responded with a positive acknowledgement QoSiLAN RESPONSE message, the QM conirms the resource request with a positive acknowledgement QoSiLAN RESPONSE message to the requesting QH.From this moment, the low is protected by traic shaping r u l e so na l lQ H si nt h en e t w o r k ,w h i c ha p p l yt oa l ll o w s , except the protected one.he reservation state is deleted either on sot state time-out or on end-of-stream discovery, like TCP FIN lag detection.

Admission Control. he proposed admission control
algorithm works according to the principles of coordinated resource admission control.his shall prevent overbooking of resources.he QM provides the function of a inal decision point.It carries out priority considerations in terms of network resource availability, based on client requests.It also takes care that resource reservation request does not block best efort signalling traic and that individual links are not overbooked.According to comparative simulation results from Jamin et al. [22] a network utilisation rate of 77% is achievable in a multihop scenario, with a utilisation target of 80% =1− and no packet loss due to congestion.For this, we deined a threshold of =2 0 % residual capacity, which shall not be blocked by reservations and is instead reserved for best-efort traic. he following algorithm deines that a reservation will be denied by the QM: if the sum of current reservation states ] for one of the afected connections/links including the predicted bandwidth and residual capacity for a low exceeds the link capacity .his algorithm enforces bandwidth allocation by collaborative shaping.By default, the QHs in the network shape their traic to the residual bandwidth, to isolate the reservation's traic from congestion.his works ine for a two-host network. he probability of multiple hosts using the full capacity of residual bandwidth and therefore exceeding it in sum grows with every host joining the network.herefore, an additional control function is needed to manage the residual bandwidthandtosharethisresourceforthebesteforttraic among the hosts.A reactive approach, as proposed by Hock et al. [23] cannot be applied to this scheme, since the traic characteristic is not predictable enough.As carried out in Section 2.3 the traic characteristic is not continuous, but bursty with pause periods.herefore, a QoS degradation by congestion is hard to detect by the receiver.Hence, our approachworksasfollows.
SinceeveryQHinthenetworkismonitoringitsoutgoing traic continuously, it detects streams with signiicant best efort traic rates.he rate is regarded as signiicant, if the output rate exceeds the residual capacity shared by the number of QHs ≥( 1 / ) .I ft h i si st h ec a s e ,t h eQ H informs the QM about its current average best-efort traic output rate and the destination of the stream.he QM collects this information from all QHs and maintains it per link in the network topology map. he QM takes care, that the sum of best-efort traic from all QHs will not exceed for each single link in the LAN.If a reservation violation is detected by the QM, the afected QHs are advised to shape their traic accordingly for traic crossing the afected links.

Related Work
In this section, related work for the new technology proposed in this paper is discussed.hese are the QoS management in LANs, the statistical class-based bandwidth prediction for lows, and the policing and admission control in LANs.

Approaches for
Figure 6: QoSiLAN policing information low diagram.
An approach, most similar to the QoSiLAN framework, has been proposed by Louvel et al. [2 -26], who propose a network resource management framework for multimedia applications distributed in heterogeneous home networks.In this solution to QoS for multimedia applications in LANs, also a central management entity, called the Global Resource Manager (GRM), is used as resource coordinator.On the local devices, the needed components are bundled in a Local Resource Manager (LRM). he LRM provides a resource estimation method, implemented using the Bienaymé-Chebyshev inequality algorithm [27] and a scheduling tool for traic prioritisation, the Linux iproute2 tool's tc command [28]. he GRM measures the available bandwidth on the links using the iperf tool [10] and coordinates the resources.As main diferentiation criteria, Louvel's proposal does not take into account the network topology and limits the approach to a one-hop star topology with heterogeneous network interfaces and diferent devices attached.Since the GRM as the central entity is able to manage all resources, a dedicated QoS protocol is not needed and obviously out of scope of the approach.his limits its practical applicability in home networks dramatically.In the deined topology  it beneits from the nonintrusive and adaptable resource management approach, since the end-devices do not need to be modiied essentially to achieve the desired QoS level.
A state-of-the-art approach for a common home network protocol comes from the Universal Plug and Play (UPnP) forum in forms of the UPnP-QoS Architecture [7] to enable QoS services in LANs, consisting of a single IP subnetwork.his architecture deines policing and admission control for prioritised, parametrised, and hybrid QoS control for individual links, as well as path property discovery for them.hree services are required to implement this functionality, as presented in Figure 7. he QosPolicyHolder Service gathers path information and provides appropriate policies for the traic, described by a TraicDescriptor structure.he QosManager Service, invoked by an application and implemented within a UPnP Control Point, requests the required resources from the QoSPolicyHolder Service. he QosDevice Service is responsible for establishing the QoS for a new traic stream.To support end-to-end prioritised QoS, the QosDevice Service needs to be implemented on all network devices along the data path.For links not supporting the parametrised QoS on the path, prioritised QoS is selected ho p ingfo rDifServsu p po rtb ythenetwo rk,r esul tinginan hybrid QoS operation.Network segments not supporting the UPnP-QoS architecture result in a QoS establishment failure for the whole end-to-end path.his is an aspect the proposed QoSiLAN framework overcomes, since it does not rely on network support and it can operate end-to-end, even if not all devices support the system.In addition, the UPnP-QoS framework requires not only implementation of its services on all devices, but also support by the applications causing the traic.his problem was addressed by Laulajainen and Hirvonen [29], who propose a background service running on traic causing devices, which performs fast application detection on basis of statistical analysis of the irst four packet sizes of a stream, which provides a basic identiication functionality.his measure is also a small part of the eSPID algorithm, described for the QoSiLAN framework in Section 2.2.Suraci et al. state that the UPnP-Q o Sa r c h i t e c t u r ew o r k sw e l li nt h ec a s eo fm o d e r a t et r a i c loads but may fail whenever the network becomes overloaded [30].hey demonstrate their admission control and drop solutions using test-bed evaluations.For the admission control algorithm they rely on the bandwidth information provided by the UPnP framework.If a network segment on the data path has no suicient residual bandwidth, the admission is rejected; otherwise it is admitted.he drop strategy decision algorithm is realized using a binary tree to estimate lowest cost for packet dropping.hey determine the cost by the importance/priority of the low. he QoSiLAN framework also deines an admission control algorithm, which also supports not QoS aware network segments, in contrast to the presented solution.A dedicated drop strategy is implicitly given by the traic shaping appliance the QoSiLAN requests from the operating system.Castrucci et al. go one step deeper in the scheduling of packets with their proposal for an application QoS management and session control in a heterogeneous home network using inter-MAClayersupport [31].heyproposeanarchitecturaland procedural deinition of the home context using the UPnP-QoS and SIP frameworks.Within this OMEGA called architecture they introduce a convergence layer between IP and MAC layer to manage all traic using the information provided by the UPnP-QoS framework.his architecture requires implementation on all network devices and also uses a centralised coordinator within the network's gateway to manage the resources.Chen e ta l .[ 3 2 ]p r o p o s eaD i f S e r vf o c u s e ds c h e m ef o rQ o S management in heterogeneous home networks.it also adopts to the QoS-UPnP speciication by adding monitoring and resource management functionality to the framework.hey monitor real-time network traic to adaptively control t h eb a n d w i d t ha n dm a n a g et or e d u c ej i t t e rl a t e n c ya n d packet loss signiicantly.he focus of this particular work is set on the last mile from the service provider to the home network, which is out of scope of the QoSiLAN framework.Since the QoSiLAN framework also may control the Internet gateway, it also manages these resources and controls the Internet line.Furthermore, Westberg et al. use the OSGi's resource management in DifServ (RMD) [33] architecture to interface to the Per Hop Reservation (PHR) and Per Domain Reservation (PDR) protocols to manage the network traic not on per low basis, but on link basis.In this way they also cover wired and wireless QoS concerns into admission control using the proposed adaptive QoS mechanism, proved by evaluations in an heterogeneous test-bed.Lee et al. [3 ] propose an enhanced UPnP QoS architecture to support network-adaptive media streaming in home networks.hey state that the initial UPnP QoS architecture does not provide methods for dynamic network monitoring.hus, they propose to enhance it by adding a dynamic network monitoring and adaptation scheme.Although the UPnP QoS 2.0 speciication introduced the GetRotameterInformation method to retrieve network status information of other important features, it still lacks the QoS-based adaptation method and the capability of guaranteeing streaming quality over time-varying networks.hey propose to enhance the UPnP QoS Device with an dedicated Status Monitor component and the UPnP QoS Manager with an QoS Adapter.h e i r purpose is to acquire continuous network status information for a dynamic QoS management.hey use this enhanced functionality to adapt the video streaming quality dynamically.his approach is diferent from the QoSiLAN framework, since it does not aim to prevent congestion, but only to react on network performance degradation, which leads to lower video quality and therefore probably l o w e rQ o E .B r e w k ae ta l .[ 3 5 ]p r o p o s ea ne n h a n c e m e n t to UPnP QoS for automatic QoS provisioning, which is a missing feature to have a better comparability to the QoSiLAN framework.hey describe the problem of autoclassiication of the traic from non-UPnP-QoS devices present in UPnPQoS enabled networks.A limitation of their proposal is the assumption that the home gateway and all network devices are UPnP QoS aware and support their enhancements.Only end-devices are allowed to be not UPnP QoS compliant.An advantage is the integration of the provider network using GMPLS transmission, which provides better QoS enforcement possibilities.heir simulationshowssimilarresultsasourswiththesamesetup time issues, which are inherited from the reactive traic identiication and classiication approach.
Also the ITU-T proposes an architectural framework of a home network that supports multimedia services within the recommendation H.622 [6]. he ITU-T identiies two diferent roles that home networks fulil and name them as primary and secondary domain.For the primary domain the home network is considered as an extension of the access network from the provider point of view.For the secondary domain, they consider the home network as an intra-LAN transmission medium for data distribution among home devices from the user point of view.As an extension of the access network, they state that providers expect it to behave similar to their access network with the same functional QoS services with security and management entities that can be found typically in provider networks.In the role of interconnecting home devices these features may not be needed.For QoS they also deine two diferent QoS models: class-based QoS and session-based QoS. he session-based QoS is recommended to be realized as UPnP [36] and classbased QoS using [37,38].hey emphasise the features of these models, like for class-based QoS the less complexity, scalability, and priority-based mechanism.For session-based mechanisms they criticise that some network devices may be unaware of signalling protocol, because network devices need a complicated mechanism and that additional session setup time is introduced by the resource reservation process.Interestingly, they also consider NSIS QoS NSLP [19] and UPnP QoS [7] as emerging new QoS technology which need further consideration.his is exactly what also the QoSiLAN framework does by further developing the ideas from NSIS and UPnP to enable autonomous session-based QoS for unmanaged networks.In that way the QoSiLAN framework complies with the H.622 recommendation for the primary as well as the secondary domain and ills the gaps of the identiied drawbacks of existing and referenced solutions.h e r ea r eal soQo Sa p p r oa c h e sf o rh i gh e rl a y e ra p p l i c ation, for example, the routing layer.Haikal [39,0].his is an Open Shortest Path First (OSPF) link-state routing protocol extension, which works independently of the used QoS architecture.his kind of routing-level QoS architecture works well for large scale hierarchical, routed networks but does not provide a solution to unmanaged local networks using a single subnet, which is thetargetenvironmentfortoQoSiLANframework.

et al. propose a d i s t r i b u t e dQ o Sa d a p t i v er o u t i n ge n g i n ea r c h i t e c t u r eb a s e d on OSPFxQoS
he Data Distribution Service (DDS) for real-time systems [ 1] is a middleware architecture for device, service, and QoS management for data centric communication in highly dynamic distributed networks.It follows the publishsubscribe communication model and is able to provide QoS in any environment, where users, devices, and services are potentially mobile.Al-Roubaiey and Alkhiaty provide an architecture for a QoS aware DDS middleware in an ubiquitous environment [ 2]. heir proposed solution as well as the DDS speciication does not provide technical solutions, but only high level descriptions of solution principles and are therefore not directly comparable to to QoSiLAN framework.

Bandwidth Prediction for Flows.
In the past, the scientiic community addressed the problem of traic prediction mainly to continuous traic lows, Internet backbone traic or even more speciic on video codec level.he algorithms, designed for encoding bandwidth prediction, use algorithms to exploit the nature of MPEG video to allocate the bandwidth on a scene basis [ 3]. he algorithms used for Internet backbone prediction address scenarios, where multiple streams run through one link and predictions aim on providing forecasts for the multiplex, the sum of streams [ ].In contrast, our solution addresses only single stream predictions, in a local network scenario.In publications addressing streaming media, mainly traditional streaming protocols were investigated.To the best of our knowledge there is no comparable scientiic publication looking at the nature of real Internet traic from modern cloud services and Content Delivery Networks (CDNs) on a per low basis.Even in the area of traic modelling no comparable work could be found.An oten addressed problem is the traic prediction on a large scale basis for Internet backbones [ 5], which aims on statistical predictions like number of streams, amount of bandwidth, and occurrence probability.Some general applicable algorithms, like the Recursive Last Square (RLS)inanapplicationoftraicprediction [ 6]andmachine learning algorithms like SVM [ 7], were investigated.It was found that those algorithms are on one hand designed to predict the traic on a short term but do not perform very well in an inert framework like the QoSiLAN addresses, with forecasting intervals of 60 s.In addition, those algorithms cause high computation complexity, but the QoSiLAN framework's design requires for a lightweight approach with a minimum of computation costs, since it is designed to runo nthinandalsomob ilesystemswi thlimi tedCPUand power sources.hat is why we aim on a simple approach to predict the needed bandwidth as accurate as possible.hese requirements are mostly hit by linear prediction algorithms.He et al. [ 8] distinguish between Formula Based (FB) and History Based (HB) algorithms.For the FB algorithms they propose linear prediction algorithms.Among others a Moving Average (MA) predictor was presented.We also employed and conigured it for our application and referenced it as Mean Estimation (ME), for our evaluation of results in Section 5.3. he HB algorithms do not it for our application, since they require a large set of throughput measurements from previous transfers in the same path, which behaved similar.In particular, the similar behaviour of protocols is an assumptionwecouldnotreproduceinourevaluations,where similar video streaming lows from diferent large CDNs behave very individually.

Policing and Admission Control in LANs.
Af r a m e w o r k for providing end-to-end QoS for individual lows was proposed by Yang et al. [ 9] with the goal of keeping the scalability of the DifServ model.hey propose the On-Demand QoS Path (ODP) framework, which supports per low admission control and end-to-end bandwidth reservation.In contrast to our QoSiLAN framework, the ODP framework targets interdomain/Internet QoS by involving edge-and core-router.he ODP framework enables scalability by using class-based service diferentiation in the network c o r e .heO D Pf r a m e w o r kr e d u c e st h es i g n a l l i n ge f o r tb y using a hierarchical bandwidth management scheme.From evaluations they conclude that the ODP Central Control and Router-Aided approaches provide end-to-end guarantees to individual lows with signiicantly less overhead than IntServ QoS like RSVP.
In terms of admission control, one can distinguish between parameter-based admission control (PBAC) and measurement-based admission control (MBAC) algorithms.Whereas the PBAC algorithms rely on a priori knowledge and accurate network traic models to allot the resources, the MBAC algorithms rely on actual measurements and accurate estimation of QoS parameters.Brewer and Ayyagari [50] compare and analyse MPAC and PBAC in testbed evaluations.hey conclude for bursty traic patterns that the MBAC approach provides better network utilisation and a higher admission rate than the PBAC approach.Similar results from Mancuso and Neglia [51] prove true the superiority of MBAC algorithms about PBAC algorithms, in special for scenarios with bursty nature of self-similar lows.hey discovered that MBAC algorithms make the system robust to statistical traic properties.hat is why we also decided to investigate more on MBAC algorithms and designed our approach according to this scheme.Moore [52] identiied ive characteristics for an appropriate MBAC algorithm.First a MBAC must provide a relationship between the traic characteristic and the calibration control.Second, the estimatormustincorporatethestatisticalnatureoftraic.hird, the estimator and the MBAC must be matched to the task required.Fourth, the algorithms must be implementable with realistic resource requirements.Fith, the policy performed by the MBAC inluences the overall performance critically.Overall, he concludes that the correct maintenance of the current provision values is more important than the accuracy of short term traic characterisation.Independently, we also designed our MBAC algorithm for QoSiLAN according to these principles and share the experiences.Jamin et al. [22] evaluated three MBAC and one PBAC algorithm in terms of performance for controlled load service.hey conigured the PBAC algorithm for capacity bounding.he three MBAC algorithms are based on equivalent bandwidth, acceptance region, and measured bandwidth.Although they do not aim on giving inal conclusions on their simulation results, their evaluations reveal that a higher utilisation target than 80% causes packet loss in the network.In another survey Jamin and Shenker [53] observed that all known MBAC algorithms can be reduced to one formula, as shown in and be tuned with parameters to give the same result curves.In (5) ] is the measured load, is the link bandwidth, and (⋅) and (⋅) a r ef u n c t i o n so ft h es o u r c e ' sr e s e r v e dr a t ea n dt h e number of admitted sources.herefore, they conclude and propose to focus future research on the tuning parameters, i n s t e a do ft h ea l g o r i t h m si t s e l f .A n o t h e ro b s e r v a t i o ni st h e structural limitations of MBAC algorithms.First, long lasting connections will statistically dominate the reservations over short lasting connections.Second, lows that traverse multihop paths have a higher risk of a rejected admission, if the switches perform the admission independently.algorithm Section 2.5.2. he application of MBAC algorithms in the context of QoS for Quality of Experience (QoE) was shown by Latré and De Turck [5 ]. he authors propose a MBAC algorithm for provider based video rate controlling.hey deine policies of how providers can use MBAC algorithms and video rate control policing for the goal of revenue maximization or QoE.his is a passive approach to react on QoS degradation.Instead, the QoSiLAN framework aims on preventing congestion and interference traic actively.

Materials and Methods
his section discusses the approaches, environments, and scientiic methods used to achieve the presented results and e v a l u a t i o n sa sw e l la st h er e s e a r c hm e t h o d o l o gyt od e v e l o p the proposed research solutions.Firstly, it presents the evaluation test-bed most evaluations were based on.Aterwards, for each proposed QoSiLAN framework key module, the special evaluation or simulation setup is described.
4.1.Home Scenario Evaluation Testbed. he evaluation testbed was selected to represent a typical home environment with six active hosts, as presented in Figure 8. hree of the hosts are connected using ixed line 100Base-T Ethernet links and three hosts are connected using WiFi IEEE 802.11g links.
In special, the simplex/duplex nature of the diferent link typesandtheh ybridQoSbehaviourarerepresentedbythis setup.Each one wireless and one wired host are serving as media server and therefore as data source.he others are conigured as media clients to consume media and demand for resources interactively. he Host was conigured as QoSiLAN Manager (QM), to manage the resources, since t h ei n t r a -L A Nt r a i ci si nf o c u sf o rt h i ss c e n a r i o .F o r s e t u p s ,w h e r et h eI n t e r n e tt oL A Nt r a i ci si nf o c u s ,t h e router is conigured as QM. he hosts are Netbook devices with Windows 8 operating system. he router is a Linksys WRT-5 GL device [55] running with the Linux based DD-WRT operating system.All devices, including the router, are equipped with the portable QoSiLAN framework stack. he test-bed network is isolated from the laboratory's traic using therouter' sNATandirewallfunctionality .heInternetlink is routed through the laboratory LAN, sharing a 100 Mbit Internet link, provided by the facilities of the University of Applied Sciences Mittelhessen, Germany, which is connected to the German Scientiic Network (DFN) backbone [56]. he DFN Internet backbone X-WIN is a science network, connecting more than 60 universities, science institutes, and science related companies within Germany, Europe, and abroad using one of the most powerful iber-based communication networks in the world.
It was found that the LLTD algorithm is the most appropriate s t a t eo ft h ea r ta p p r o a c ho fp h y s i c a lt o p o l o g yd i s c o v e r yi n LANs. he protocol was implemented by Microsot for their Windows operating system products since the release of Windows Vista, including a closed source LLTD Mapper service and LLTD Responder service for Microsot Windows and an open source LLTD Responder for Linux based operating systems.he most important part, an API to the LLTD Mapper service or an open source implementation of it, is not available publicly.Although technical protocol descriptions exist and the algorithm was presented within a conference paper [57], a lot of implementation and algorithmic details of the LLTD mapping process are not published.So, to be able to use the technology within the QoSiLAN framework and for its evaluations the LLTD mapper had to be reverse engineered andreimplemented.hiswasamajortask,sincethemappingprocess is very complex and especially discovering deep segments and hosts is not trivial.To reverse engineer the LLTD algorithm and to test the LLTD mapper during the implementation and reverse engineering phase, an Ethernet network simulator was developed, which simulates the behaviour of switches, hubs, and LLTD responder nodes. he network simulator, as shown in Figure 9 provides basic Ethernet functionality to emulate the Ethernet communication behaviour, addressing the switch's AIT building behaviour.In addition, the LLTD responder nodes also implement the LLTD responder behaviour for LLTD message sending and responding and the "sees" list.All addresses, tables, and lists are inspectable through the user interface.Additionally, the network simulator provides functionality to pause and continue the communication to provide rich debugging possibilities.It provides a live watch feature, to follow the Ethernet packet traversal through the LAN using animations. he LLTD mapper application, shown in Figure 10, is a separate component, which was developed to support both communication with the simulator through named pipes and portable Ethernet operation using the libpcap/winpcap API interfaces for Linux and Windows based systems [61,62]. he LLTD mapper as well as the Ethernet Simulator was developed using the Microsot .NET framework [63] and the Mono project framework [6 ] to support platform independence.For the other evaluation scenarios, which make use of the LLTD features and as inal regression test, the LLTD mapper was tested and productively used within the evaluation test-bed, as presented in Figure 8.Further details on the work of the LLTD are not included in this paper, since we did not enhance this technology but only used it for the QoSiLAN framework and its test-beds.

Enhanced Statistical Protocol
Identiication.Preceded b yi n t e n s i v er e l a t e dw o r ks t u d y ,t h eS P I Da l g o r i t h mw a s identiied as best itting for the QoSiLAN framework since it fulils the requirements of the targeted environment.It is light weight of high precision and enables protocol identiication and application payload identiication at the same time, even for encrypted or compressed traic.Ater related work study and irst tests the SPID algorithm was selected as appropriate base. he published implementation was evaluated to be not well implemented in terms of performance and memory usage and needed to be reimplemented.In addition, laboratory test showed that the SPID performance was not as good as expected.herefore, all measures and parameters were evaluated and optimised.In addition, the measures were reconigured and additional once developed and tested, to ind a better measure coniguration for the QoSiLAN framework evaluation test-bed.he completely reengineered algorithm was called eSPID. he eSPID implementation is written in C++ using the libpcap/winpcap API interfaces [61,62].
he eSPID evaluations to ind the best measure and algorithm coniguration were carried out using a set of 3135 lows from 17 diferent protocols.he set of evaluation lows was recorded from real web-browsing and application usage under various usage scenarios, to cover diferent protocol behaviours for the same protocol.During the development process and for the ine-tuning of the algorithm, the diferent measures were tested individually to verify their performance andusefulnessforthewholesetofmeasures.
4.4.Statistical Class-Based Bandwidth Prediction. he SCBP algorithm was developed from the motivation to have a simple and light-weight algorithm with low computing complexity.Ater literature research and investigation of realworld traic from major video and audio streaming portals it was found that the state-of-the-art literature solutions do not handle the characteristics of nowadays Internet media streaming traic explicitly.For that reason, the SCBP algorithm was designed from the practical observation that streams need a prior classiication and case by case handling before predications should be applied.In particular, the irst ten seconds of transmission were found as signiicant for the overall transfer behaviour.Dependent on the characteristic of the irst ten second transfer behaviour, a systematic deviation from the expected results could be discovered.herefore, intensive evaluations were carried out to optimise the classiied results using individual correction factors for each class.
In addition, diferent optimisation approaches were followed in parallel and compared to ind the best optimisation set for the prediction results.All evaluations for the SCBP were performed with real Internet traic from common WebTV, IPTV, Internet Radio, and on-demand platforms, located in Germany, United Kingdom, France, and the United States of America. he evaluation was carried out according to Figure 11 on Host 1. Host 1 is connected using 100BaseT Ethernet to an Internet gateway, which provides access to the University Of Applied Sciences Mittelhessen's (THM) Internet connection. he SCBP implementation is written in portable C++ using the libpcap/winpcap API interfaces [61,62]. he streams were automatically identiied using the eSPID algorithm [15] and classiied according to the algorithm discussed in Section 2.2.Only streams with a minimum transfer time of 60 s were included in the evaluation, which resulted in a set sized of 63 samples.his allowed us to get the signiicant average bandwidth consumption value for the irst 60 s ( 60 ) for each low. he ( 60 )-value served as reference and was used to validate the prediction accuracy ater ten seconds ( 10 ).

QoS Admission Control and Policing
. To e va lu ate t he concept of the QoSiLAN framework and the efectiveness of its admission and policing algorithms, the scenario shown in Figure 12 was conigured to overload the network for prooing the system in a critical situation.In addition, the features of hybrid inter-access-medium QoS, per link bandwidth reservation, and simplex/duplex handling are shown within this example.To show the efectiveness of the QoS approach, Figure 13 depicts a direct comparison of a bandwidth conlict scenario with and without QoS enabled.
his comparative evaluation was executed within the testbed shown in Figure 12.Host 1, Host 2, and Host 5 request TCP streams with 8 Mbps from Host 3. In Figure 13(a) the results are presented, where there is no QoS conigured, whereas in Figure 13(b) the results were generated with QoSiLAN's QoS functionality enabled and requested for all of the streams.All stream have a bandwidth demand of 3 Mbps.S i n c et h e ya l la r es h a r i n gt h es a m ea c c e s sm e d i u m ,w h i c h was assessed with a capacity of 20.6 Mbps before a bandwidth conlict occurs when all streams are running at the same time.In Figure 13(b), the conlict is resolved by QoSiLAN, which detects the overprovisioning and hence does not admit the r eserva tio nr eq uestf o rS tr ea m3.her ef o r e,i tlimi tsS tr ea m 3t o3M b p s .F i r s t ,H o s t1r e q u e s t sS t r e a m1a n dc a ns t a r t without disturbance.Ater 20 s Host 2 requests Stream 2. Since the access medium capacity is not exceeded in both scenarios, no problem occurs.he main diference until that poin tintimeisthevisibleefectofthetraicsha per ,which harmonises the variance of the bandwidth consumption.Ater 0 s Host 5 requests Stream 3. Figure 13(a) shows that this harms the transmission of Stream 2, which shares the residual bandwidth with Stream 3, now.Also Stream 1 is afected and disturbed in transmission.Figure 13(b) shows the QoSiLAN approach's efect, where the start of Stream 3 does not harm the transmission of the other streams.Only in the beginning of Stream 3 there is a short disturbance of Stream 2 until the traic shaper in Host 5 is applied ater 20 packets of low recognition and identiication.
he whole QoSiLAN framework was also evaluated using the test-bed in Figure 12. here, Host was setup as QoSi-LAN Manager.Table 2 shows a tabular view on the actions performed within the evaluation.
Before the start of the evaluation, Host already performed the LLTD mapping process and assessed the wireless link TCP throughput from the wireless Host 3 to the other wireless Host 1 with = 10.6 Mbps.his is a realistic throughput for IEEE 802.11g connections with both nodes connected to the same wireless access point, due to the simplexnatureoftheaccessmedium.hetestedthroughputfrom the ixed Host 6 to the wireless Host 1 was measured to be 20.6 Mbps.
he maximum TCP throughput between the ixed Host and Host 6 was tested with 87.7 Mbps. he tests were carried out under congestion free network conditions and a close physical link for the wireless hosts, using the iperf TCP and UDP bandwidth performance measurement tool [10].In the beginning, the network is in idle state.For our evaluations, Host 1 starts requesting a video stream (Stream 1) from Host with an average bandwidth of 8 Mbps, as shown in Figure 1 .Ater 20 packets the eSPID module identiied the stream type as video and ater 10 s the average output bandwidth 1 of Stream 1 was predicted with 8.0 Mbps. he admission control algorithm decided to admit the reservation, since it is in the rage of 80% link capacity 8.48 Mbps = ≥ =8 .0Mbps of the wireless link D. he QM initiated the QoS signalling and advised all wireless hosts to shape their outgoing traic to other wireless hosts to = 2.12 Mbps residual bandwidth.Exactly 20 s ater start, Host 2 also requests TCP data (Stream 2) from Host 3 with a data rate of 2 Mbps, which was predicted with2 = 2.1 Mbps.his stream was not identiied as audio/video stream and therefore no reservation was initiated.he predicted bandwidth of Stream 2 ( 2) is smaller than the residual bandwidth and thus causing no QoS problems.Although for this low no reservation is requested, the QM is informed about the current bandwidth occupation.Now, both streams are running in parallel without any disturbance or interference.When Host 5 also requests  3 Mbps is detected.Since no reservations are applied to the links A and B the reservation is admitted and does not afect the other reservations.his new reservation is also not communicated to the wireless hosts, since the bottleneck bandwidth for the path from the wireless hosts to Host 6 is lower than the residual bandwidth on link B. Host 6 has no limits to Host , due to the duplex nature of the Ethernet links and no reservations apply in that direction.In addition, there is no need to communicate the Ethernet reservation to the wireless nodes, since their outgoing traic limit is lower than the one on the Ethernet links.Host limits its outgoing traic for other lows than Stream to Table 3 shows the shaping policies as applied at the end of the evaluation, when all reservation states are active.As one can see, all host receive individual policies, according to their location in the network.In that way perlink reservation states are enforced.(ii) AV-streaming/IPTV (higher priority); (iii) VoIP/IP telephony (highest priority).
In the case of two conlicting reservation requests, the irst come irst served policy is applied.So, the last incoming conlicting request is declined, unless it belongs to a higher prioritised traic class.If a higher prioritised request is recognised, the QH owning the lower one is informed with at e a rd o w nr e q u e s tf o ri t sr e s e r v a t i o ns t a t e ,i nf o r m so fa QoSiLAN RESERVE message with the tear lag set.Also all other QHs in the network are informed to delete the lower prioritised reservation state and to establish the higher one.his enables the QH to inform the user about the loss of QoS for the running service.6 gives an overview about the results achieved for the diferent conigurations.here, also the median ( MPAR) and the mean accuracy (MPAR) values are listed, which enable a irst assessment of the results.While the mean accuracy value emphasises the average network utilisation, the median accuracy value gives an indication about the distribution of values.Figures 16(a)-16(c) give an overview about the three diferent optimisation aspects, whereas Figure 16(d) shows the results ater the Euclidean distance calculation, which relects a combination of the previous ones.In Figure 16(a) the Over-Estimation Rate optimisation, with a target value of 1, indicates that the Mean Estimation (ME) and CLPHR aspects perform best. he CBQoS performs only on the third place together with CLQoS aspect.In Figure 16(b) the Mean Prediction Accuracy Ratio results with a target value of 1 that indicates the optimum for the CLMPAR and CBMPAR aspects, as expected, but the CBQoS aspect performs as next best.he Prediction Hit Rate aspect results in Figure 16(c) show the best performance for the CBPHR and the CBQoS r e s u l t s ,w h i c hs h o u l db en e a rt o1t ob et h eb e s t .hi si s interesting and shows that the optimisation for Prediction Hit Rate without classes cannot provide an average ratio better than 0.77. he class-based approach gives a better  16(d) illustrates the advantage of the CBQoS approach, which does not perform best in all categories, but collectively over all optimisation aspects it shows the best performance.he statistical class-based bandwidth prediction algorithm using the CBQoS optimisation coniguration was proved to perform best for predicting bandwidth requirements for a relative long period of 60 s.On the irst view the classless results look also very good and perform in some conigurations similar to the class-based approach or even better.Finally the Euclidean distance comparison in Figure 16(d) reveals the advantage of the CBQoS optimisation approach, since it brings the deviation of results into account, combining the diferent measures.Also, we showed, as presented in Figure 16, that a 100% prediction accuracy with a minimum error is not desirable, as it causes a higher underestimation probability than, for example, an QoS optimisation case.Generally, it is more critical for the media streams, if their resources are underprovisioned, since they are not properly protected from congestion then.In contrast, blocking of more resources than needed is critical for the overall network performance and utilisation.In the application of QoS, a little overestimation of resources is also to be regarded as positive, since this allows the streams to prebufer faster at the receiver, which results in more robustness against the variability of network performance and provides a better stream isolation against disturbing traic and congestion.

Conclusion
In this paper we presented and evaluated the QoSiLAN framework in its whole for the irst time. he interaction of the diferent key technology, which enable link based resource reservation by autonomous policing and admission control for QoS in LANs, was shown.he QoSiLAN admission control and policing algorithm was designed to take care not to block more resources in the network than needed, by allowing to react dynamically on reservation violations and network congestion in an appropriate manner autonomously.he QoSiLAN framework was presented to be an efective method to reserve bandwidth on individual links in LANs without the network's QoS assistance and therefore to provide self-organised QoS for unmanaged networks.he admission control and policing function in the QM utilizes the QH information from the topology mapping and the traic-analysis and -reporting to take informed decisions for managing the LAN's resources eiciently.he new SCBP algorithm was presented along with detailed evaluations to optimise it for the best resource predication performance within the QoSiLAN framework.he evaluations show that QoS conlicts are detected reliably and a solution is enforced in an autonomous way. he novel QoSiLAN QoS model, designed to enforce cross-access-technology linkbased bandwidth reservation by collaborative traic shaping without network assistance using the dedicated QSLP-LAN protocol, was proved within the evaluations.he QoSiLAN framework is implementable in a lightweight manner.It does n o tr e l yo nn e t w o r ks u p p o r t ,b u to nh o s tc o l l a b o r a t i o n .I n addition, not all hosts in the network need to support the framework essentially.For a proper operation, the support by at least one of the communication parties within the LAN is required, if the router is QoSiLAN aware.Traic sources like Media-NAS devices and the Internet router may implement the framework preferably to allow for an optimised operation.his makes the QoSiLAN framework easy to implement and to be deployable in realistic scenarios.he autonomous coniguration and operation features qualify the framework for nonexpert deployment and application.

Figure 1 :
Figure 1: Overview of the QoSiLAN key components.

Figure :
Figure : Reservation path parameter for IPv .

Figure 5 :
Figure 5: QM and receiver initiated message low.

Figure 9 :
Figure 9: LLTD network simulator with the test topology loaded.

Figure 10 :
Figure 10: LLTD mapper application with the test topology's segment tree.
outside the wireless link D a shaping limit of 1 Mbit was c o m m u n i c a t e d .A ss o o na st h i sn e wp o l i c yi sa p p l i e db y the hosts, the throughput of Stream 1 recovers to its desired state, as depicted in Figure1.In parallel, a video stream from Host to Host 6 (Stream ) with3  = 8.

Figure 16 :
Figure 16: Performance assessment of optimisation results.
[19].QSLP-LAN Message Format. he QoS Signalling Layer Protocol for QoSiLAN is designed based on the NSIS message format, similar to the QoS NSLP header format described in[19], Section 5.1.
the data path.In the following, we present the QSLP-LAN protocol messages and the signalling procedures. he protocol behaviour within the QoSiLAN framework is presented as examples within the evaluations in Section 5.2. 2 QoS in Local Area Networks.Beside the QoS approach presented in this work, other solutions for enabling QoS in LANs or home networks are described in the literature. he approaches vary from the access layer to the application layer.Since the QoSiLAN approach is an approach targeting cross access technology scenarios, other approaches working on layer 1 and layer 2 are not discussed.
5.1.3.Limitations.Since the QoSiLAN approach relies on client support, most of the possible reasons limiting the framework's efectiveness are caused by lacking QoSiLAN support by hosts: (i) One host, with lacking QoSiLAN support, can be managed by the corresponding host.In cases if the sender is not QoSiLAN aware, on behalf, the traicreceiver may tell the QM about the needed resources.
network.In addition, the QM has to calculate the r e s e r v e dc a p a c i t i e si nt h en e t w o r ka n dt ot a k ec a r e that it reserves not more resources than available.It has to mind a minimal bandwidth capacity for all hosts in the network to enable best efort for at least signalling applications.

Table 6 :
Overview of results.