Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 7, No. 4, August 2017, pp. 2261 2277 ISSN: 2088-8708 2261       I ns t it u t e  o f  A d v a nce d  Eng ine e r i ng  a nd  S cie nce   w     w     w       i                       l       c       m     A Pr oposal f or End-to-End QoS Pr o visioning in Softwar e-Defined Netw orks Francesco Lucr ezia 1 , Guido Mar chetto 2 , Fulvio Risso 3 , Michele Santuari 4 , and Matteo Ger ola 5 1,2,3 Polytechnic of T urin, T orino, Italy 4,5 CREA TE-NET , T rento, Italy Article Inf o Article history: Recei v ed: Feb 17, 2017 Re vised: Jun 2, 2017 Accepted: Jun 17, 2017 K eyw ord: QoS Pro visioning Netw orking SDN ONOS ABSTRA CT This paper describes a frame w ork application for the control plane of a netw ork infras- tructure; the objecti v e is to feature end-user applications with the capability of requesting at an y time a customised end-to-end Quality-of-Service profile in the conte xt of dynamic Service-Le v el-Agreements. Our solution tar gets current and future real-time applications that require tight QoS parameters, such as a guaranteed end-to-end delay bound. These ap- plications include, b ut are not limited to, health-care, mobility , education, manuf acturing, smart grids, g aming and much more. W e discuss the is sues related to the pre vious Inte grated Service and the reason wh y the RSVP protocol for guaranteed QoS did not tak e of f. Then we present a ne w signaling and resource reserv ation frame w ork based on t he cutting-edge netw ork controller ONOS. Moreo v er , the presented system foresees the need of considering the edges of the netw ork, where terminal applications are connected to, to be piloted by dis- tinct logically centralised controllers. W e discuss a possible inter -domain communication mechanism to achie v e the end-to-end QoS guarantee. Copyright c 201x Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Name: Francesco Lucrezia Af filiation: Polytechnic of T urin, Italy Address. Corso Duca de gli Abruzzi, 24, 10129 T orino, Italy Email: francesco.lucrezia@polito.it 1. INTR ODUCTION Internet service pro viders (ISPs) are stri ving to inno v ate their netw ork infrastructures at the pace content pro viders do with their services. Digital contents are consumed by smart phones and sophisticated terminal stations that continuously e v olv e together with the applications the y host. Interestingly enough, the e v olution of the Ov er - The-T op (O TT) services is mainly happening without the aid of netw ork service pro viders, within the best-ef fort data traf fic channel in the access netw orks. Recently , a ne w plethora of applications requiring a R TT delay of around 1ms ha v e been grouped under the hat of tactile-internet applications: a tactile s ensor reads information and a connected system reacts with actuators seen by a human within 1 ms [1]. Although we are still f ar from achie ving end-to-end R TT of around 1ms with wireless communications, ISPs need to be ready to re-a rchitect their softw are control-plane in order to fully e xploit the enormous potentials of fered by their infrastructures. The goal of this paper is to present the design and a prototype implementation of a control-plane netw ork application for pro visioning dynamic end-to-end QoS profiles to end-user applications. The current adoption of distrib uted control algorithms forces the use of the same signaling protocol (e.g. RSVP , BGP-LS) in all the data-path nodes, not taking into account the resistances ine vitably present between de vice v endors and between administrati v e domains. F or this reason the Service-Le v el-Agreements (SLAs) between a service pro vider and its customers or between pro viders are still mainly sta tic. Moreo v er , the e xperience has sho wn that the scalability issue of the core netw ork in maintain- ing per -flo w s tate for resource reserv ation in each node along a path pre v ented the dif fusion of RSVP and inte grated services in general. As discussed in [2], per -flo w service treatment does not scale in the Internet core; backbone routers must be f ast and only an a g gr e gate behaviour is feasible. Instead, it is important to enable such treatment J ournal Homepage: http://iaesjournal.com/online/inde x.php/IJECE       I ns t it u t e  o f  A d v a nce d  Eng ine e r i ng  a nd  S cie nce   w     w     w       i                       l       c       m     DOI:  10.11591/ijece.v7i4.pp2261-2277 Evaluation Warning : The document was created with Spire.PDF for Python.
2262 ISSN: 2088-8708 at the edge of the netw ork where mass of users enj o ying a mixture of heterogeneous applications share indistinctly the same portion of the netw ork as in the case of cellular access netw orks; performance de gradation is lik ely to hap- pen in the access links where an increasing number of traf fic sources and sinks can introduce a significant amount of queueing delay . Gi v en these considerations, we belie v e that an h ybrid combination of flo w-based and class-based traf fic treatment, respecti v ely at the edge and in the core of the netw ork, could enable guaranteed services for current and future real-time applications. Since these applications could ha v e terminals deplo yed all around the globe, the end-to-end pro visioning will ha v e to span a chain of administrati v e domains, requiring an east-west communication interf ace to con v e y data that v ary from classic inter -domain routing information e xchanged via BGP . If we mak e the assumption that QoS requirements requested by the customer applications are satisfied in the core netw ork, at least for what concerns a delay bound, and up to a maximum bandwidth allocation, then we can think to o v erlay an inte grated service scheme on top of the current deplo yments where class-based treatment is applied, as long as the resource admission control system is able to map the dynamic service requests to the statical ly allocated resources in the core. Our proposal is a signaling scheme for pa th reserv ation and configuration whose implementation does not require the in v olv ed data-path de vices to be bound to a single control protocol. The aim is to solv e the interoperability problem in pro visioning end-to-end guaranteed services, in a multi-v endor , multi-technology and mult i-domain en vironment by e xploiting current softw are technologies adv ances; in particular , the decoupling between functional intents and the w ay the y are accomplished is crucial. The paper is or g anised as follo ws: Section 2. gi v es a high le v el description of the complete system w orkflo w , the first part of Section 3. is dedicated to a general introduction to the QoS and to the Internet technologies adopted to achie v e it. Here is where the most of t he related w orks are considered. Then it focuses on our specific use-case scenario and on the RSVP critical issues. In Section 4. we discuss the communication interf ace between clusters or domains of netw orks required to achie v e the end-to-end pro visioning, while in Section 5. a brief conte xtualisation of our w ork into a polic y management system is presented. Section 6. and a ll its subsections contain the specific details of our prototype system implementation and Section 7. presents some benchmark results on the service request computation time. Finally , Section 8. concludes the paper . 2. SYSTEM W ORKFLO W The o v erall process is acti v ated upon the occurrence of some e xternal e v ent, either sensorial or caused by the human will, that triggers the dispatch of a request sent by the user application. The request contains a user au- thentication tok en, an application identifier , information about an endpoint to contact together with additional flo w specifications, and a QoS profile containing delay and bandwidth requirements, plus an amount of time (or an estimate of it) for which the profile is required. A pre vious agreement between the user and the netw ork operator is made in order to con v e y to a traf fic plan based on its dynamism, amount of data, QoS parameters, number of request s and possibly other par ameters. The request is sent to a manager application running on top of the local domai n controller that is listening for incoming connections from re gistered users. The authentication tok en, pre viously generated in a hand-shak e phase is v erified and the content of the payload is parsed and elaborated as follo ws. The endpoint information, either a destination address (L2 or L3, depending on the use-case) or the hostname of the machine to be contacted, is used as look-up k e y to get the collection of candidate paths e xisting between the end terminals of the user application. Then an admission contr ol routine runs to check the a v ailability of a suitable path where resources are to be reserv ed for the subject QoS profile. Once a path is selected, the netw ork application has to instruct t h e core controller to setup a priority flo w between the endpoints of the user application; for the sak e of simplicity , we will refer to point-to-point paths, although the solution is equally applicable to paths with multiple destinations ( > 2). It may be necessary to establish connecti vity between the endpoints, other than traf fic control’ s rules. F or e xample, in a pure Openflo w netw ork where none of the routing protocol suite runs in the data-path de vices, the controller w ould prescribe a set of flo w rules containing forw arding instructions, together with QoS constrai nts. If forw arding rules ha v e been pre viously installed on the data-path de vices, then only traf fic classification and shaping is to be done through de vice-specific configurations. T o setup a priority flo w , the manager application issues the setup of custom queues in the de vices along the selected path and sends an intent request to the controller’ s core with the specification of the tar get flo w . An intent is an abstraction used by the applications to specify their high-le v el desires in form of policies. The ONOS netw ork controller [3], used in our prototype implementation, is the first open-source controller that pro vide such feature to IJECE V ol. 7, No. 4, August 2017: 2261 2277 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2263 ( 2)  R e que s t ( 1)   T opol ogy di s c ove r y ( 4)  P a t h s e t up ( 3)  R e s our c e  r e s e r va t i on DOM AI CO N TRO LLER  A P e e r i ng P r ovi de r A dm i s s i on  C ont r ol ( 4)  P a t h s e t up DOM AI CO N TRO LLER  B P e e r i ng P r ovi de r A dm i s s i on  C ont r ol Figure 1. Reference scenario 10 En d - Po i n t in t e r lin k CL U S T E A CL U S T E B N e t w or k vi e w  of   c l us t e r   A Figure 2. Domain topology abstraction its appli cations. The intent is then split by the core into de vice-specific flo w rule requests dispatched to the proper softw are dri v ers of the underlying de vices. Queues are also configured by means of de vice-specific dri v ers. As mentioned before, the endpoints can normally span multiple domains and, clearly , each domain has to tak e care of its portion of the netw ork. A k e y point of the system is the topology abstraction (Figure 2): all forw ard- ing de vices of the local domain are e xposed to the controller with the same abstraction model, as is an entire remote domain topology , vie w as a single de vice (a big switch) whose ports are connected to terminal endpoints or to other domain topologies. When the endpoint of the application requesting a priority path resides on a dif ferent domain, the admission control routine recognises the presence of a virtual de vice associated to the remote domain and as a con- sequence queries the netw ork manager application of the remote domain in order to establish an end-to-end resource reserv ation along the path between the endpoints (Figure 1). Once the process con v er ges, the domain controllers can simultaneously setup the path in their o wn portion of the netw ork, the manager application that recei v ed the original request sends a response back to the user application that can start sending priority data o v er the netw ork. 3. QOS PR O VISIONING 3.1. General Discussion QoS is introduced in pack et-switched netw orks in order to apply a special treatment to specific data pack et flo ws. At the de vice le v el, QoS is achie v ed through traf fic classification, shaping, and scheduling at the e gress ports; at the ingress ports pack ets are filtered for policing. Classification determines which treatment each data pack et has to under go, shaping is used to control customer input data rate to conform to the SLAs, while scheduling af fects delay and throughput of the data pack et flo ws. Gi v en conformance to a SLA, scheduling at the netw ork interf aces is the lo west le v el operation gearing QoS pro visioning, also kno wn as service discipline . At the path le v el, QoS is achie v ed through resource reserv at ion o v er the whole set of nodes along the path. The reserv ati on must be secured end-to-end and the resource allocation in one node must be consistent with the others along the path. A path selection with the end-to-end delay constraint is subject to inaccurate information due to the dynamic nature of the delay and hence subject to theoretical intractability ([4] [5]). Ho we v er , end-to-end delay and delay jitter bounds ha v e been computed by means of netw ork calculus applied to queueing systems modelling the A Pr oposal for End-to-End QoS Pr o visioning ...(F r ancesco Lucr ezia) Evaluation Warning : The document was created with Spire.PDF for Python.
2264 ISSN: 2088-8708 netw orks ([6, 7, 8]). Academic and industrial communities ha v e been v ery acti v e in the last decades in in v estig ating netw ork models and algorithms to solv e the QoS r outing problem, also kno wn as the multi-constr ained path compu- tation problem ([9, 10, 11, 12, 13, 14, 15]). In [2] the y co v er all the important component s of the QoS pro visioning in Internet as it has been concei v ed in the last decades: inte grated services, RSVP ([16, 17]), Dif fServ [18], Multi Protocol Label Switching (MPLS [19]) and constraint-based routing. T oday netw ork operators emplo y MPLS mainly for layer 2 and Layer 3 V irtual-Pri v ate-Netw ork (VPN) ser - vices ([20]), while constrain-based routing for T raf fic Engineering (TE) operations is comple x to achie v e in a complete distrib uted control-plane. Moreo v er , TE is not useful in presence of congestion. This happens at the bottleneck links that typically reside in the last mile to w ards the cust o m ers. Dif fServ model within a single AS is emplo yed by ISPs for class-based tr eatment of the data pack ets; b ut the v alidity of the T ype of Service (T oS) field in the pack et IP header may lose completely meaning when tra v ersing multiple administrati v e domains. In other w ords, within a single ad- ministrati v e domain the class-based QoS pro visioning is theoretically easy t o achie v e and technology is not the hurdle, while polic y and economic f actors ha v e the major impact on the f ate of the QoS pro visioning in the multi-domain scenario. IntService and the RSVP signaling protocol instead did not tak e of f e v en within the single administrati v e domain. RSVP is used for labels distrib ution in G/MPLS b ut it f ailed in its primordial intent. In order to accomplish the process described in Section 2. an end-to-end guaranteed service must be pro vided. In the ne xt section we discuss the critical issues of RSVP and in what our proposal dif fers from it. 3.2. Comparison with RSVP P ath Computation and Routing . RSVP uses a combination of Constrained Shrotest P ath First (CSPF) algo- rithm and Explicit Route Objects (ER Os) to determine ho w reserv ed traf fic and signal ing messages are routed o v er the netw ork; RSVP is a distrib uted signaling protocol and it needs routing to w ork. ER Os are a mean to e xplicit indicate some nodes that must belong to the reserv ed path; in order to force a specific path through a set of nodes you should enter and configure each node with specific ER Os instructions. If the total bandwidth reserv ation e xceeds the a v ailable bandwidth specified across the link for a particular path se gment, the path must be recomputed through another route. If no se gments can support the bandwidth reserv ation, path setup f ails and the RSVP session is not established. In our solution the path computation is independent from the routing protocol. Dif ferent routing and for - w arding scheme can be used to b uild the path in the underlying netw ork: flo w-rules, labels, tunnels. But no routing of the signaling scheme is needed; centralised path computation is clearly much f aster and protocol-independent. The controller only needs the vie w of the topology as a connected graph. T o force the reserv ation in a specific path, you can directly reserv e resources and install forw arding rules into the proper de vices at once. The candidate paths are collected from the store and a suitable one is found before injecting resource reserv ation rules into the netw ork. This occurs in the centralised controller within a single softw are process. RSVP is simplex . In RSVP the reserv ation process is applied in a single direction of the path. T o ha v e full duple x reserv ation, the number of operations and the messages e xchanged are doubled. In a centralised netw ork controller , you can e q ua lly pro vide simple x or duple x reserv ation via a single soft- w are entity . Admission and P olicy Control . RSVP Admission and polic y control is applied to ea ch node. So RSVP implementation must be inte grated with each node’ s traf fic and polic y control module, thus increasing the chance of interoperability . RSVP must pro vide QoS service characterisation within opaque objects parsed by each netw ork node. In our proposal the QoS service characterisation is completel y decoupled by the signaling protocol/scheme and must be embedded only into the front-end APIs consumed by the customer applications. The admission and polic y control is applied only once, in each domain, in the central controller . An RPC-based mechanism is used for configur - ing the traf fic control module gi v en the possibility to fulfil the request. Our prototype relies on ONOS. ONOS pro vides common abstracted beha viours for traf fic selection and treatments that are translated into de vice-specific rules. F irst speak er . RSVP session initiator is the inbound router running RSVP in conjuction with other protocols (e.g. MPLS or GMPLS in case of label distrib ution). If RSVP is embedded in a host application, then the first net- w ork node should speak RSVP , otherwise a tunnel between the applicati on and the first RSVP-capabl e de vice shall be IJECE V ol. 7, No. 4, August 2017: 2261 2277 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2265 created thus increasing the number of operation required for RSVP signaling to w ork. In the presented solution the session initiator is a user application featured with proper APIs for contacting the controller . The idea is to k eep the API consumer implementation as simple as a REST client that has the capability to create, remo v e, update and de lete a priority path. Such client w ould be pro vided for dif ferent softw are en vironments. Scalability . As per [21], the scaling problems of RSVP are link ed to the resource requirements (in terms of processing and memory) of running RSVP . The resource requirements increase proportionally with the number of sessions. Each session requires the generation, transmission, reception and processing of RSVP P ath and Resv mes- sages per refresh period. Supporting a lar ge number of sessions, and the corresponding v olume of refresh messages, presents a scaling problem. A centralised control plane presents the same scalability issues concerning the state maintenance of an in- creasing number of sessions in the data-path nodes. But it only matters the traf fic control and flo w rule s, while processing and singnaling o v erhead are significantly lo wered. Complexity . Finally , and maybe the most rele v ant obstacle to success, RSVP is comple x because it w as designed with IP multicast in mind, intermediate nodes ha v e to mer ge resource reserv ation requests coming from the recei v er nodes. Moreo v er , the basic RSVP reserv ation model is "one pass": a recei v er sends a reserv ation request up- stream, and each node in the path either accepts or rejects the request. This scheme pro vides no easy w ay for a recei v er to find out the resulting end-to-end service. T o solv e this issue an enhancement w as proposed [22], introducing further comple xity in the concretization of RSVP and the inte grated services in general. Orchestrating the data-path nodes from within a central controller a v oids the issues related to the e xchange of asynchronous signaling messages. The decision process not being distrib uted decreases the comple xity in maintaining a single state of the system. 3.3. End-to-End Beha viour The end-to-end QoS profile model shall follo w the one described in the Specification of Guaranteed Quality of Service [23]. As per [23], " the end-to-end behaviour pr o vided by a series of network elements is an assur ed le vel of bandwidth that, when used by a policed flow , pr oduces a delay-bounded service with no queueing loss for all conform- ing data gr ams ". W e in vite to refer to the specifications for further clarification about the QoS model tak en into account. Each netw ork node must pro vide a service that matches, with some error bounds, the fluid model ([24, 25]) through the tok en b uck et scheme with paramters ( b; r ; p ) , respecti v ely the b uck et size, the tok en rate and the peak rate. The QoS request includes a maximum end-to-end delay bound, d r eq , that shall be guaranteed between the application terminals. In the centralised controller , the link pro viders are responsible for notifying the presence of the information on the links the y are pro vider for (propag ation delay , transmission capacity and the maximum trans mission unit); these information are stored in the controll er database upon disco v ery of the link itself. Lik e wise, the de vice pro viders in the southbound must e xport other rele v ant information, such as the delay error terms representing ho w the de vice’ s implementation of the guaranteed service de viates from the fluid model in each netw ork interf ace (the C tot and D tot in the formula 2 belo w). On a link l with capacit y c l , we define a minimum bandwidth reserv ed to the best ef fort traf fic, R be l . Let R i be the allocated bandwidth for a flo w i . On each link, the total number of accepted profiles N is subject to: N : c l R be l + N X i R i (1) The end-to-end delay bound as defined in [23] is: [( b M ) =R ( p R ) = ( p r )] + ( M + C tot ) =R + D tot + X l d p l (2) W ith r < = p < = R , M being the path Maximum-T ransmission-Unit and P l d p l the propag ation delay sum. A Pr oposal for End-to-End QoS Pr o visioning ...(F r ancesco Lucr ezia) Evaluation Warning : The document was created with Spire.PDF for Python.
2266 ISSN: 2088-8708 Statement 1 imposes that the sum of the allocated bandwidths for N flo ws must not e xceed the capacity of the link. Flo ws requesting a maximum delay bound are assigned to higher priority queues w .r .t. to the best ef fort traf fic. It is possible to assign the same high priority queue to more distinct flo ws, as long as statement 1 holds. R shall be chosen such that d r eq is greater or equal to the v alue computed in equation 2, pro vided that d r eq is greater than the fix ed delay terms D tot and P l d p l . These constraints must apply on all links of a candidate path between the end terminals of the customer application; a ne w queue is created for a ne w flo w if the y are satisfied. As mentioned in the pre vious section, the admission contr ol routine occurs only once per domain (more details in Sec. 4.), in the central controller . Netw ork element s must e xport the proper information, while the dri v ers ha v e to translate a service request into de vice-specific traf fic control rules. The pro visioning of a guaranteed service along a path of se v eral de vices and links is possible only through a cross-v endor and cross-technology solution. This leads to the adoption of softw are dri v er modules installed into the centralised controller . These dri v ers con v erts the protocol-agnostic rules into de vice-specific instructions and are essential to s olv e the interoperability problem deri v ed by the presence of de vices from multiple v endors and technolo- gies. F or e xample, in a L TE cellular netw ork, the high-le v el profile is mapped to a standardised QoS Class Identifier (QCI) by the proper softw are dri v er; the mapping w ould be follo wed by the setup of the pack et data netw ork g ate w ay and the mobile station with some scheduling rules applied to the tar get data flo w [26]. The same high-le v el profile has to be translated into a specific setup on the backhaul that pro vides the connecti vity to w ards the core netw ork consisting of all the required switches to aggre g ate the traf fic from the access cellular netw ork [27]. The se switches could be, for instance, pure Linux de vices in which case the dri v er w ould e x ecute a remote procedure call configuring the in v olv ed interf aces with the well-kno wn commands s uite tc qdisc , tc filter and tc class for setting up the queues. If instead a switch is an Openflo w-enabled de vice, classification and priority come within the forw arding rules, while the queue configuration for the service discipline must be supplied on a separate communication channel, for e xample, through O VSDB protocol in Open-vSwitch [28] (Fig 3). T ogether with the local domain (or local cluster in the single administrati v e domain) our frame w ork adds the reflection of such operations into the remote domain (cluster) where the endpoint of the customer appl ication requesting the service resides. Note that we control the edge de vices on each side of the communication while lea ving aside the backbone routers where per -flo w service treatment does not scale. While rfc-2212 states that all the nodes of a path should tak e part of the resource reserv ation process for equation 2 to hold, we ar gue that the dynamic resource reserv ation at the edge of the netw ork can occur transparently w .r .t. the statically allocated resources of the core where the QoS e xists only for classes of traf fic, rather than flo ws, in form of virtual circuits created with protocols such as MPLS or GMPLS (Fig. 4). If the backbone is treated as a composition of these cir cuits rather than a composition of nodes and links, then a resource mapping between the dynamic and the static portions of the netw ork resolv es in representing these circuits as aggre g ate elements into the topology vie w of the controller . P ath-Computation-El ement (PCE) describes a model to address the problem of constrain-based path computation in conjunction with a label switched protocol ([29, 30]). An all in one orchestration frame w ork for the complete set of the netw ork elements is left as a future w ork. I ngr e s s   P ol i c i ng F or w a r di ng P R I O  0 P R I O  1 Q ue ui ng I ngr e s s   P ol i c i ng F or w a r di ng Q ue ui ng C ont r ol l e r L i nux  tc dr i ve r O pe nf l ow dr i ve r OVS DB dr i ve r L i nux de vi c e O V S  br i dge RP C tc qdi s c tc fi l t e r tc cl as s Q ue ue   c onf i g que ue  i d que ue  m i bw que ue  m a bw F l ow   c onf i g s e t  que ue s e t  pr i or i t y Figure 3. Dri v er modules in the controller are the means for de vice-le v el configuration of the queues IJECE V ol. 7, No. 4, August 2017: 2261 2277 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2267 Dy n a m i c r es o u r c e a l l o c a t i o n St a t i c r es o u r c e a l l o c a t i o n (e. g . M P LS c h a n n el s ) Figure 4. T w o allocation schemes for tw o layers of the netw ork 4. EAST -WEST COMMUNICA TION INTERF A CE When the terminals of the application requiring a priority path are located in tw o portion of the netw ork pi- loted by distinct controllers, a communication mechanism between these controllers is necessary in order to e xchange the proper information during the reserv ation process. From hereafter , we will use the term domain and cluster inter - changeably to indicate distinct portions of the netw ork. At the origin of the communication, there is the route disco v ery phase to share the endpoints of each domain. A design choice is to be made on ho w and when to e xpose the netw ork element parameters related to the topology itself. There are tw o options: a) sharing these i nformation during the route disco v ery phase, or b) a v oid to pre-share the parameters and collect them during the resource reserv ation process on-demand, that is, e v ery time a ne w request arri v es. Suppose we ha v e a portion of the netw ork under domain A , and another portion under domain B , then suppose that at some point in time an application connected to A asks for a priority path that includes an endpoint in domain B . The tw o dif ferent approaches are described in the follo wing sections. 4.1. Pr e-Shar ed Netw ork P arameters and Band width Resour ce W ith option (a) , the controller A has already all the information required for the end-to-end admission control routine when the request arri v es. This means that the topology e xposed by B to A shall include all the necessary netw ork infrastructure parameters, ( c l ; M T U ; C tot ; D tot ; d pr op tot ) p for all p in the set of paths that B is willing to e xpose to A , which in turns implies that the e xposed topology shall be detailed enough for A to select a suitable path. This requires a more comple x topology abstraction than the single big switch as depicted in Fig. 2, because clearly with the si ng l e node abstract ion you cannot ha v e as man y path properties as you w ould ha v e with a netw ork of nodes. Y ou can al w ays achie v e a certain le v el of aggre g ation by hiding element s of the B topology , by computing the aggre g ate parameters for some paths to w ards the destinations and then e xposing a virtual topology composed by only these paths. In this case each domain controller should maintain a mapping between the local ph ysical de vices and links and the virtual ones e xposed to the remote domains. Other than the topology abstraction, there is one major issue with this approach: the computation of the aggre g ate, rate-related delay error term, C tot , which is theoretically not feasible when the computation occurs, for instance, in the domain controller A for some de vices of domain B , because C depends on the parameter r , the rate requested by a user application. So either C , e xpressed as a function of r for each de vice, is shared during the topology disco v ery phase, which means e xposing the entire topology , or it must be computed on-demand, by the domain controller B as described in the ne xt section. 4.2. On-demand Netw ork P arameters and Band width Resour ce In this case, the path parameters are collected and e xchanged during the admission control routine. The do- main B is requested to run the resource reserv ation process in its o wn domain. Controller A forw ards the application request parameters (i.e. d r eq , the tuple ( b; r ; p ) , the domain ingress point and the tar get destination) to B , which replies with a tuple ( M ; C tot ; D tot ; d p ) B and an upper bound on R , chosen based on a proper selected path, if a v ailable, so that A can compute the end-to-end delay bound before proceeding with the actual reserv ation and path setup. This w ay the topology abstraction can be k ept as simple as a single big switch whose function is to merely of fer a point of connection t o the remote destinations t o an y domain controllers with which there is a peering. The netw ork parameters and the resource select ion comes directly from an up-to-date decisi on process made within the concerned domain. The computation of C tot can be actually computed while masking the details of the local topology . This approach is the choice of our implementation prototype described in the rest of the article. A Pr oposal for End-to-End QoS Pr o visioning ...(F r ancesco Lucr ezia) Evaluation Warning : The document was created with Spire.PDF for Python.
2268 ISSN: 2088-8708 4.3. Resour ce Management The netw ork parameters used for the delay bound computation in equation 2 are mainly static, e xcept for the bandwidth R , the dynamic netw ork resourc e under consideration. W e assume unlimited b uf fer space for the queues, or at least enough to fill the entire bandwidth resource in an y link of the netw ork. W ithin each domain a resource management system should track the allocated and the a v ailable bandwi dth in each link of the underlying netw ork. From the inter -domain communication perspecti v e, the bandwidth resource management unfolds three cases depending on the use-case scenario: Full shar e . The bandwidth is completely shared between applications, re g ardless of the domain the y reside. This can be the case where the control plane of a single administrat i v e domain is split into multiple re gions for performance reason or because of dif ferent underlying ph ysical netw orks. In this scenario, it is necessary to update the e xposed resource each time an application obtains or releases a portion of the bandwidth. Upon a ne w allocation or release in one domain, a mes sage is sent in broadcast to all the other domains with which there is a peering. The message is read by the remote controllers and their local resource stores are updated accordingly . In the on-demand mode of communication, Section 4.2., each domain manages its bandwidth resource independently and only during a reserv ation process the resource a v ailability in a remot e domain is determined. P artial shar e . This case is equal to the pre vious one b ut only a portion of the bandwidth is shared with the remote domains. P artial static shar e . A domain controller adv ertises to each remot e controllers a virtual v alue of the bandwidth resource during the route disco v ery phase and no further update messages are e xchanged between controllers. This v alue is the static portion of the bandwidth allocated to each remote domain. In this case, also with the on- demand mode of communication a controller can determine the a v ailability of bandwidth e v en before contacting a remote domain. This is the case of the multi administrati v e domains scenario. 5. POLICY ENFORCEMENT Polic y-based QoS management is of primary importance for netw ork operators. If the service discipline at the netw ork interf ace is the lo west le v el operation gearing QoS, at the top le v el we ha v e the SLAs e xpressed in terms of pol icies. The SLAs consist of a set of specifications that are translated by the netw ork manager into de vice le v el primiti v es (e.g., forw ardi ng rules, queue configurations, traf fic shaping policies, etc.). In [31] [32] the authors pro- pose automatic polic y based management system in the Internet Dif fServ architectures. The y present a frame w ork for polic y management that reacts to netw ork state changes or customer users requests to dynamically re-adapt the polic y enforcement. In [33] a management frame w ork for automati c polic y enforcement is introduced in a netw ork controller based on Openflo w; the y describe all the necessary functional components of the system without entering in the implementation details of an y of them thus a v oiding to discuss ho w do the y actually interact between each other . The focus of the present article is on the QoS pro visioning in the economic conte xt of dyn a mic SLAs; this frame w ork foresees the possibility to be inte grated with an e xisting polic y-based management system. The concerned SLAs are between a service pro vider and its customers and between service pro viders who cooperate to pro vide an o v erall service that can span multiple administrati v e domains. Between the customer and the pro vider , a set of APIs can be embedded directly into the customer applications and layered on top of an e xisting polic y management system. The netw ork operator could also pro vide ready-to-use applicati ons for specific services (e.g. a remote health control system). Here we limit the discussion by listing the additional information needed by the polic y manager in order to conform the ingress traf fic to the dynamic SLAs. P olicy to regulate the interaction with customer applications: List of user and application IDs that are allo wed to request a service Upper bound on the amount of bandwidth each user could request Maximum amount of time each user is allo wed to retain a priority path Amount of bandwidth reserv ed to the best ef fort traf fic Set of destinations for which a user could request a priority path Set of netw ork elements that cannot be part of the reserv ation process Pre-configured queues for selected customers or applications IJECE V ol. 7, No. 4, August 2017: 2261 2277 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 2269 P olicy to regulate the interaction betw een pro viders: List of peer domains that are allo wed to interact with the local domain List of destinations to e xpose to the remote domains T opology abstraction to e xpose to the remote domains V irtual bandwidth resource associated to the e xposed destinations Aggre g ate parameters selection, M T U ; C tot ; D tot ; d p Number of total service requests that a remote cluster is able to perform Pre-configured queues for selected peers 6. ARCHITECTURE 6.1. High-le v el System Components The main functi onal modules in the control plane are pr otocol a gnostic thanks to the separation of concerns gi v en by the netw ork controller architecture of ONOS (see Section 6.2.) that is partitioned into: Protocol-a w are netw ork-f acing modules that interact with the netw ork Protocol-agnostic system core that tracks and serv es information about netw ork state Applications that consume and act upon the information pro vided by the core At the application layer resides the admission contr ol and r esour ce allocation routine. It guarantees the correctness of the QoS pro visioning to the end-user applications; it has to dynamically setup a n d teardo wn multiple and concurrent QoS profile sessions and v erify that e v erything in the underlying netw ork is up-to-date and in a consistent state . In the core controller there are se v eral components required to accomplish the compl ete reserv ation process, see Figure 5, while in the netw ork-f acing layer we ha v e as man y dri v ers as the number of dif ferent d e vices in the underlying net- w ork and a comm unication interf ace used to e xchange data between the domain netw ork controllers. Such interf ace has to tak e into account se v eral aspects of the system: routes to destinations disco v ery , netw ork topology elements e xposition, resource reserv ation parameters and a polic y-dri v en mechanism to abide t o the SLA made between the in v olv ed administrati v e domains. W e le v erage on a project called ICON A to fulfil the remote topology and destina- tions disco v ery function, see Section 6.4.. The resource reserv ation parameters are currently e xchanged during the admission contr ol routine at the application layer , using a prototype REST channel interf ace. The inte gration with a polic y manager is left as a future w ork. P a t h M a na ge r A dm i s s i on C ont r ol F r ont - E nd  A P I s P r of i l e  S t or e Q ue ue  C onf i gur a t i on I nt e nt  I ns t a l l e r T opol ogy M a na ge r N e t . E l e m . S t or e R e s our c e  M a na ge r D e vi c e  M a na ge r D om a i n M a na ge r I nt e nt  M a na ge r F l ow R ul e M a na ge r D r i ve r  M a na ge r D e vi c e  D r i ve r s R e m ot e  D om a i T opol ogy P r ovi de r A ppl i c a t i on C or e P r ot oc ol s P r ot oc ol s Figure 5. Controller’ s Main Components A Pr oposal for End-to-End QoS Pr o visioning ...(F r ancesco Lucr ezia) Evaluation Warning : The document was created with Spire.PDF for Python.
2270 ISSN: 2088-8708 A ppl i c a t i on N or t hbound  I nt e nt  F r a m e w or k ( pol i c y e nf or c e m e nt , c onf l i c t  r e s ol ut i on) D i s t r i but e d C or e ( s c a l a bi l i t y , pe r s i s t e nc e , a va i l a bi l i t y) S out hbound ( di s c ove r , obs e r ve , pr ogr a m c onf i gue ) O pe nf l ow N e t C onf OVS DB Figure 6. ONOS distrib uted architecture 6.2. An o v er view of ONOS The Open Netw ork Operating System (ONOS) is a softw are defined netw orking (SDN) OS for Service Pro viders, that is tar geting scalability , high a v ailability , high performance and abstractions to mak e it easy to cre- ate apps and services [34]. ONOS implements a distrib uted architecture in which multiple controller instances share multiple distrib uted data stores with dif ferent le v el of consistenc y . The entire data plane is managed simultaneously by the whole cluster . Ho we v er , for each de vice a single controller acts as a master , while the others are ready to step in if a f ailure occurs. W ith these mechanisms in place, ONOS achie v es scalability and res ilienc y . Figure 6 sho ws the ONOS internal archi- tecture within a cluster of four instances. ONOS is based on softw are modules managed by the Apache Karaf suite [35], a set of ja v a OSGi based runtime and applications. It pro vides a container into which v arious component can be deplo yed, installed, upgraded, started and stopped at runtime, without interfering other components. The southbound modules manage the ph ysical topology , react to netw ork e v ents and program/configure the de vices le v eraging on dif- ferent protocols. The distrib uted core is responsible to maintain coherent infor mation, to elect the master controller for each netw ork portion and to share information with the adjacent layers. In case of a f ailure in the data path (switch, link or port do wn), an ONOS instance becomes a w are of the e v ent through the s outhbound modules, computes alter - nati v e paths for all the traf fic crossing the f ailed element, and notifies them to the distrib uted core; then, each master controller configures accordingly its portion of the netw ork. The northbound subsystem of fers an abstraction of the netw ork and the interf ace for applications to interact and program the NOS. Finally , the Application layer of fers a container in which third-party applications can be deplo yed. Applications on top of ONOS can benefit of the Intent Frame w ork. The ONOS core accepts the intent specifications and translates them into actionable operations on the netw ork en vironment. These actions are carried out by the intent installation process, such flo w rules being installed on a switch, or optical lambdas (w a v elengths) being reserv ed. 6.3. The Manager A pplication The manager application is a standard, on-platform ONOS application. Its main function is the admission contr ol routine. W e separate the routing and resource assignment process into tw o steps: the collection of candidate paths between the application terminals through the ONOS P athMana g er and the selection of the one that satisfies the constraints imposed by the basic admission contr ol scheme des cribed in Section 3.. The state of the underlying netw ork resources is track ed by the Resour ceMana g er , back ed by a distrib uted store, whic h is queried during this process to reserv e the bandwidth resource. The local path parameters are collected by the internal store populated with the information of t h e underlying netw ork elements, while i n presence of virtual domain de vices, the parameters are queried to the remote controllers and then mer ged with the local ones. If a suitable path is found using algorithm 1, the bandwidth in each link is temporary reserv ed (lock ed), priorities queues are setup in each de vice through the QueueMana g er module and an intent is submitted by the IntentInstaller (Fig. 7). If the installation is successful, the Resour ceMana g er is requested to allocate the pre viously lock ed resource, otherwise a rollback is performed on all the IJECE V ol. 7, No. 4, August 2017: 2261 2277 Evaluation Warning : The document was created with Spire.PDF for Python.