Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Recursive InterNetwork Architecture (RINA) wikipedia , lookup
Zero-configuration networking wikipedia , lookup
Deep packet inspection wikipedia , lookup
Asynchronous Transfer Mode wikipedia , lookup
Distributed firewall wikipedia , lookup
Piggybacking (Internet access) wikipedia , lookup
Computer network wikipedia , lookup
Wake-on-LAN wikipedia , lookup
Network tap wikipedia , lookup
Cracking of wireless networks wikipedia , lookup
appeared in Proc. of 2nd European Conf. on Universal Multiservice Networks ECUMN'2002, April 2002 Using only Proportional Jitter Scheduling at the boundary of a Differentiated Service Network: simple and efficient Thu Ngo Quynh (*) Holger Karl (**) (*) Interdepartmental Research Center for Networking and Multimedia Technology PRZ / FSP-PV / TUBKOM Adam Wolisz(**) (**) Klaus Rebensburg (*) Telecommunication Networks Group Department of Electrical Engineering and Computer Science Technical University of Berlin Strasse des 17. Juni 136 10623, Berlin, Germany Email: {thu, klaus}@prz.tu-berlin.de, {wolisz,karl}@ee.tu-berlin.de Abstract: There exist some studies in Proportional Delay Scheduling Algorithms for Differentiated Service (DiffServ) Network which schedule the packets between different classes proportionally based on queuing delay. Traditionally, it is necessary to implement such proportional delay scheduling algorithms at all the routers of the network for receiving proportional delay between different classes. In this paper, we propose a new and simple model, which does not provide proportional delay but delay jitter between different classes. In addition, we point out that for receiving proportional jitter in such a DiffServ Network, it is only necessary to implement a Proportional Jitter Scheduling algorithm at the boundary of the network. Furthermore, we propose some networks using RJPS (Relative Jitter Packet Scheduling) and WTP (Waiting Time Priority) at different positions (core or egress) and a play out buffer delay adjustment algorithm (Concord algorithm) at the receiver. Finally we compare its performance in terms of normalized end-to-end delay with different loss probabilities at the play out buffer. Our first result showed that a proportional jitter DiffServ Network achieves better performance while keeping the cost of implementation lower in some cases. 1. Introduction Currently, the Internet can only provide a single service class – best effort – that requires no pre-specified quality of service (QoS) contracts and provides no minimum QoS guarantees for packet flows. To provide adequate service, some level of quantitative or qualitative determinism – IP services must be supplemented and that is what QoS protocols are designed to do. A number of QoS protocols have evolved to satisfy a variety of application needs: Integrated Services (IntServ) and Differentiated Services (DiffServ). The DiffServ approach [1] is newer than the IntServ approach and proposes a coarser notion of quality of service, focusing primarily on aggregated flows in the core routers, and intends to differentiate between service classes rather than provide absolute per-flow QoS guarantees. In particular, access routers process packets on the basis of finer traffic granularity such as per-flow or perorganization and core routers do not maintain fine grained state, but process traffic based on a small number of Per Hop Behaviours (PHBs) [4] encoded in the packet header. Since late 1997, the DiffServ working group has discussed several proposals for per-class relative QoS guarantees [2, 3]. With the exception of the Expedited Forwarding service [4], proposals for relative per class QoS discussed within the DiffServ context define the service differentiation qualitatively, in the sense that some classes receive lower delays and a lower loss rate than others, but without quantifying the differentiation. Recently, research studies have tried to strengthen the guarantees of relative per-class QoS, and have proposed new buffer management and scheduling algorithms, which can support a stronger notion of relative QoS [5, 6, 7, 8]. Probably the best know such effort is the proportional service differentiation model [5, 6], which attempts to enforce that the ratios of delays or loss rates of successive priority classes be roughly constant. For two priority classes such a service could specify that the delays of packets from the higher priority class be half of the delays of the lower priority class, but without specifying an upper bound on the delays. The Waiting Time Priority scheduler [6] is such a scheduling algorithm, which implements a well known scheduling algorithm with time dependent priorities. Traditionally, for achieving proportional delay between different classes in a DiffServ network, it is necessary to implement these proportional delay scheduling algorithms as WTP at all the routers. Our earlier work [9] showed that it is possible to meet the requirement of Relative Proportional Differentiated Service Model in terms of delay jitter by using a simple new scheduling algorithm called RJPS (Relative Jitter Packet Scheduling). This paper also examined the behaviors of RJPS in different contexts as variable window size, variable packet size, variable load condition, etc…within only one hop. In [11] the performance of two networks which use only RJPS or only WTP at its routers showed that our scheduler RJPS could produce smaller delay than WTP in some cases, but the delay in a RJPS network fluctuates much more than a WTP network while WTP is a very stable scheduler under different load conditions. Hence we proposed some other topologies, which use RJPS as Ingress or Egress router while others are WTP for gaining proportional jitter between different classes and minimizing the delay at the receiver after the play out buffer. It is very important to know that we should not only examine the behaviors of delay and jitter proportional schedulers in the same network to verify which can produce better network delay, but the influence of these network delays at play out buffer algorithm because play out buffer delay is adjusted with the variation of network delay (that means delay jitter) and loss rate. In this paper, we propose another Model of DiffServ Network which does not provide proportional delay between different classes, but proportional jitter. We prove that such a proportional jitter network requires only proportional jitter scheduling algorithms at the boundary of the network, and hence decreases extremely the cost of implementation. Whether to implement a model of Relative Proportional Differentiated Service in terms of delay or delay jitter in a network, it is very noteworthy to know that the objective of our works is to improve end-to-end quality of service, and in this case, a better end-to-end delay (the sum of network delay and play out buffer delay) We continue to focus our research on some networks, which could provide proportional delay and/or jitter between different classes and develop a new performance criteria for comparing its qualities based on end-to-end delay, which is called normalized end-to-end delay. This paper is organized as follows. In Section 2 we give an overview of the current works on relative per class QoS guarantees and some play out buffer delay adjustment algorithms. In Section 3 we propose some different network topologies. Section 4 models our network and the performance criteria for comparing the quality of different topology. The detailed result is shown in Section 5 and in Section 6 we present brief conclusions. 2. Related Work Proportional Jitter Scheduler: RJPS is a workconserving packet scheduler that serves N queues, one for each class. The goal of this algorithm is providing the proportional jitter between different classes. Suppose that each router has a pre-specified number of jitter classes N. Each jitter class is served by a single first-in-first-out (FIFO) packet queue. Packets of a flow belonging to a jitter class i are queued in the corresponding queue in each router that the flow passes through. All flows with the same jitter class specification share the same FIFO queue at the router. The goal of RJPS scheduling algorithm is to serve the packets such that the short term average jitter (calculated over a moving window) and long term average jitter (calculated from the beginning time of simulation) experienced by packets in a jitter class will stay proportional for all pairs of classes. In [9] the behaviors of RJPS in different contexts as variable window size, variable packet size, variable load condition, etc. are examined. Proportional Delay Scheduler: Early work by Dovrolis outlined definition and main issues of relative service differentiation. It also introduced two scheduling algorithms for delay differentiation, namely the Backlog Proportional Rate Scheduler (BPR) and the Waiting Time Priority scheduler (WTP) [6]. The BPR scheduler is a version of Weighted Fair Queuing, whose rates can be dynamically adapted to the load situation of the system in order to maintain delay differentiation between classes of service. Delay Differentiation under BPR is found to be inaccurate in the case of moderate traffic load or asymmetric load distribution between the classes of service. In contrast, WTP can obtain more accurate delay differentiation by decoupling delays from service rates. Another proportional delay scheduler is proposed by Timely group, which is called MDP [8]. This scheduler is more convergenz than WTP because it maintains a window of packet over time for calculating the average delay and priority of each class of the expense of the complexity. We conclude that among delay proportional schedulers WTP, on the one hand, can estimate delay differentiation very well. The BPR scheduler, on the other hand, can support throughput differentiation and link sharing due to its load-controlled nature, but its performance in terms of delay differentiation is unpredictable. MDP is more complex and depends strongly on traffic pattern. We assume that WTP is the best among delay proportional schedulers. Play out Buffer Delay Adjustment Algorithm: Although using such proportional schedulers at the routers in Internet, a characteristic of the network is that the total delay experienced by each packet is a function of variable delays due to physical media access and relay queuing, in addition to fixed propagation delays. The result is that the time difference between transmitting any two packets at the source is unlikely to be the same as that observed between their arrivals at the destination. This is a problem for a stream of multimedia packets, because the presence of delay variation can have an impact propagation delays. The result is that the time Network Sender Buffer Receiver on the audio-visual quality as perceived by a human user. buffer delay of each packet dynamically such that the total end to end delay of all packets are TED, that means one packet that arrives later than TED will be thrown away, and all the packets that suffer network delay Jitter smaller Figure 1. System Elements To solve these synchronization problems, one must introduce additional delay by buffering packets at or near the point of presentation. Clearly, if every packet is delayed in the buffer such that it suffers cumulatively (in the network and buffer) a delay equal to the maximum network delay, the receiver can reproduce a jitter-free play back. Because play out buffer delay is adjusted with the variation of network delay (that means delay jitter) and loss rate, an interesting question is not only to examine the behaviors of delay and jitter proportional schedulers in the same network to verify which can produce better network delay, but the influence of these network delays at play out buffer algorithm. In our network, we use Play out Buffer Adjustment Algorithm described in [11], which is called Concord Algorithm. This algorithm is described as follows. Concord Algorithm: As shown in Figure 1, we have a network N, over which sender S wishes to send stream M to receiver R, with the packets in M being produced every T seconds. All packets have sequence numbers. For the network N the Concord algorithm constructs a packet delay distribution (PDD), an estimate of the probable delays suffered by packets in the network over a time window. This PDD may draw on existing traffic conditions, history information or any negotiated service characteristics to derive estimate for minimum, maximum and/or mean delay distributions. PDD 100% D than Playoutbuffer TED = TED- D will network be D network buffered during the , we will receive the loss rate of L%. We could write: TED TED = f PDD ( w, J network (t ), L) L is the predefined loss rate. W is a window of where packets over it the Probability Delay Distribution is calculated, and at the time t. J network (t ) is the actual variation of delay Hence we could rewrite D Playoutbuffer as follows: D Playoutbuffer = TED - D network = Playoutbuffer f PDD ( w, J network (t ), L) 3. Formal Description of Proportional Differentiated Service Model The model for Relative Proportional Differentiated Service: The proportional model states that certain class performance metrics should be proportional to the differentiated parameters that the network operator chooses. A generic description of the proportional differentiation model follows. Suppose that qi (t , t + τ ) is the performance measure for class i in the time interval (t , t + τ ) where τ > 0 is the monitoring timescale. If differentiation over short timescales is desired, the value of τ should be relatively small. The proportional differentiation model imposes constraints of the following form for all pairs of classes and for all time interval (t , t + τ ) in which both q i (t , t + τ ) and q j (t , t + τ ) are defined: L qi (t , t + τ ) 0% M in q j (t , t + τ ) D elay TED Figure 2. PDD Distribution In the example shown in Figure 2, a PDD distribution is computed. From this distribution, when we have a predefined loss rate ratio L, we can easily find the value of total end-to-end delay TED. If we adjust the play out where = ci cj c1, c 2 ,..., c N are the generic Quality Differentiation Parameters (QDPs). The basic idea is that, even though the actual quality level of each class will vary with the class loads, the quality ratio between classes will remain fixed. The model of Relative Proportional Differentiated Service for Delay: Based on the previous model, in the context queuing delay, we can write as follows: d i (t , t + τ ) d j (t , t + τ ) where parameters Parameters and = ∆j ∆i ∆ i are the Delay Differentiation d i (t , t + τ ) are the average queuing delay of class i packets (class i is better than class j if ∆ i > ∆ j ). The model of Relative Proportional Differentiated Service for Jitter: In the context of queuing delay jitter, we created a new model of Proportional Jitter Differentiated Service Network, which could provide proportional jitter between different classes. This model is illustrated by the equation : ji (t , t + τ ) j j (t , t + τ ) where parameters = ∆j ∆i ∆ i are the Jitter Differentiation Parameters and ji (t , t + τ ) are the average queuing delay jitter of class i packets (class i is better than class j if ∆ i > ∆ j ). delay between all classes at any time, is a proportional jitter scheduler with the same Differentiation Parameters. Assumed that jitter of one packet in a queue is the difference of queuing delay of this packet and the preceding packet in this class. It is easy to realize that when the current delays between two classes are proportioned with its Delay Differentiation Parameters, its difference will stay proportional with its Delay Differentiation Parameters, too. In addition, along the work of the Differentiated Services Working Group of the IETF all streams refer to unidirectional streams. And although WTP could provide only average proportional delay between different classes and is not an optimal proportional delay scheduling algorithm, these features lead us to further interesting ideas of using WTP and RJPS in different positions of a network (egress or core routers) and compare its performance to know which topology is the best under which conditions. 4. The model of network Suppose that the routers in our network use WTP or RJPS for scheduling packets proportionally between different classes i (i is from 1 to N), each class i has a weight ∆ i . Through our network, each packet number n of class i suffers a network delay D network ,i ,n . At the receiver, this packet belong to class i is buffered, Property: For providing proportional jitter between different classes in a DiffServ Network, it is just necessary to have one proportional jitter scheduler at the Egress router. The others could be FIFO or other scheduling algorithms. For providing proportional delay between different classes, it is necessary to have Proportional Delay Scheduler at all the routers of the network. It is easy to see that delay has accumulated property, that means delay of a class or flow through a network is the sum of the queuing delay at each router and the time of transmission. Hence there is a need to implement proportional delay scheduler at every router in a DiffServ Network for receiving proportional delay between different classes. Jitter, however, could not be accumulated and it is only necessary to have a proportional jitter scheduler at the Egress of the network to receive proportional jitter between different classes. This property pointed out that for providing proportional jitter in a DiffServ Network there is only a need to implement a proportional jitter scheduling algorithm (as RJPS), and for providing proportional delay between different classes in a DiffServ Network it is necessary to implement proportional delay scheduling algorithms (as WTP) in all routers of this network. Furthermore, we should note that an optimal proportional delay scheduler, which produces precisely proportional and its play out buffer delay is called D Playoutbuffer ,i , n is depending on D Playoutbuffer ,i , n . delay variation J network ,i (t ) , loss rate Li and the window w over it the PDD is calculated and defined by Concord algorithm. Playoutbuffer D Playoutbuffer ,i , n = f PDD ( w, J network ,i (t ), Li ) The end-to-end delay of this packet is calculated as the sum of network delay and play out buffer delay: endtoend ,i , n TED= Di = D network ,i ,n + D Playoutbuffer ,i , n = D network ,i ,n + Playoutbuffer f PDD ( w, J network ,i (t ), Li ) The average end to end delay of one class i is calculated as follows: D endtoend ,i = ∑D endtoend ,i , n n n Performance Comparison Criteria: Because the DiffServ philosophy is to realize Relative Differentiated Service, which seeks to provide per-hop per-class relative services, there are no absolute guarantees due to the lack of admission control or resource reservations. Consequently, the network cannot provide worst-case bounds for any service metric. Instead, each router only guarantees that the service invariant is locally maintained, even though the absolute service might vary with network conditions. And whether to implement a model of Relative Proportional Differentiated Service in terms of delay or delay jitter in a network, the objective of our works is receiving better end to end quality of service, and in this case, a better end-toend delay (the sum of network delay and play out buffer delay). Hence we choose end-to-end delay as performance metrics for comparing the quality of different network topologies. Supposing we have N classes, each has a weight of Network endtoend , 0 ( D NTj Topology endtoend ,1 , D NTj number j endtoend , N −1 ,..., D NTj ∆ i and produces ) end-to-end delay for N classes when the loss rate is L%. Under the same condition (the same load, the same loss rate, the same load distribution between different classes) the Network Topology number k produces endtoend , 0 ( D NTk endtoend ,1 , D NTk endtoend , N −1 ,..., D NTk ) end-to-end delay for these N classes. We believe that it is complicated to establish a set of performance criteria for comparing the performance of proposed network topologies, because it is depending on pricing structure, on the importance of each class...One possibility is introducing some cost for each byte in one class, and comparing the gain we receive. Some could introduce also the reward for decreasing the delay in the higher class and some loss for the decrease of the delay in the lower class. For simplicity, our proposition is using normalized end-toend delay for comparing the quality of these topologies. routers in the network we could not assure we will always have proportional delay because IP packets traverse network through different routes whose number of hops is different and hence its queuing delays will not stay proportional with each other. It is a big disadvantage of Proportional Delay DiffServ Network. Due to the lack of the page numbers of this paper, we will just focus on some simple network topologies as follows: Topology of Network Property Name Only RJPS Proportional Jitter between different classes Proportional Delay between different classes Proportional Jitter only between different classes Proportional Jitter only between different classes NT1 Only WTP WTP at core router, RJPS at egress FIFO at core routers, RJPS at egress FIFO at core routers, WTP at egress RJPS as core routers, WTP at egress Table 1. Different Network Topology Pk = ∑i =1 N D endtoend ,i NTk ∑ N i =1 RJPS RJPS RJPS RJPS Network Topology 1 WTP WTP * ∆i ∆i WTP WTP WTP WTP WTP RJPS Network Topology 3 FIFO We say that the Network Topology, whose normalized end-to-end delay is smaller, is better. Network Topologies: We examine the following network topologies and its performance under the same context. In Section 3 we have concluded that it is only necessary to have RJPS at egress router for receiving proportional jitter between different classes, the others could be FIFO, WTP, or RJPS. But for providing proportional delay between different classes in such a network we need to have WTP at all routers, and even though when we have WTP at all NT3 NT4 NT5 NT6 For example, these different network topologies are illustrated in Figure 3. Network Topology 2 For the Network Topology k, we define the normalized end-to-end delay Pk of this network as: NT2 FIFO FIFO RJPS Network Topology 4 FIFO FIFO FIFO WTP Network Topology 5 RJPS RJPS RJPS WTP Network Topology 6 Figure 3. Different Network Topologies 5. Simulation Our simulation study uses the ns-2.1b7a Simulator [12]. The simulation model is as follows. The routers use packet sources of type on-off traffic. The topology used is shown in Figure 4. The links are 6Mb, 3Mb, 1.5 and 0.75Mbps. There is a total of 3 classes 0, 1, 2. The weights of class 0, 1, 2 are 2.0, 1.0, and 1.5 respectively. We will compare the end-to-end delays of different topologies NT1 to NT6 listed in Table 1. The Concord algorithm is used at the receivers. It is necessary to note that we should maintain a window of packets in order to calculate the PDD function. In our simulations, we use PDD taken over a moving packet window of size 3000 packets for Concord algorithm. C1 C1 C1 C1 C0 R1 R2 R3 FIFO C1 C0 C0 Source of class i The result in the Table 2 is the average normalized end-toend delay calculated for NT from 1 to 6, with different loss rate probabilities from 5% to 15%. The graph 6 plots the quality of these Network Topologies and puts it in order of the decreasing quality. It is easy to see that in this case the quality of the networks which contain RJPS at its routers (NT1, NT3, NT4, NT6) increases when the loss rate increases, and compared to others topologies which use WTP at its routers (NT2, NT5), it achieves higher performance. This result leads us to a conclusion that in certain cases, the network which uses only RJPS at its egress router receives better performance quality than the network which uses WTP at all of its routers. On the other hand, the implementation of RJPS at the boundary of the network improves the quality of the network while reducing the cost of implementation extremely. C0 C0 Ci It is very interesting to note that although NT4 requires only RJPS at its egress router, it could generate better quality in this case than NT2, which implements WTP at all its routers. This property allows us to say that implement RJPS at the egress router in a DiffServ Network improves the performance of this network while reducing the implementation cost extremely in some cases. Ci Receiver of class i Figure 4. Topology of simulated Network We have simulated for the different NTs from 1 to 6, with a simple topology of only 3 routers. At the receiver we receive end-to-end delay, and the Figure 5 shows the variation of the normalized delay P(s) of these network topologies while the predefined loss rate at the receiver is 5%. From this graph, we have the following conclusions: • The NT1, which uses only RJPS at all the routers, produces the smallest normalized end-to-end delay, and that means the best performance. • The last topology NT6, which contains WTP at the egress router and others are RJPS, generates worse quality than NT1, NT3 and NT4, but better quality than NT2 and NT5. • The quality of NT3 and NT4 stays between the quality of NT6 and NT1. • The two other network topologies, NT2 and NT5, which use all WTP at all its routers or only at egress router, receive the biggest normalized endto-end delay. We could say NT2 and NT5 generate the worst performance in this case. 6. Conclusion In this paper a new model of DiffServ Network which could provide proportional jitter is proposed, and its performance quality is compared with some other models of DiffServ Network providing proportional delay. This comparison is based on our new performance criteria, which is called normalized end-to-end delay. Our first result showed that the model of Proportional Jitter could achieve better performance in some cases while reducing the complexity of the network extremely. References: [1] S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss, ‘’An Architecture for differentiated services,’’, RFC 2475. [2] D. Clark and W. Fang, ‘’Explicit allocation of best-effort packet delivery service,’’ in IEEE/ACM Trans. on Networking, 6(4), August, 1998, pp. 362-373. [3] K. Nichols, V. Jacobson, and L. Zhang, ‘’Two-bit differentiated services architecture for the Internet,’’, RFC 2638, July 1999. [4] V. Jacobson, K. Nichols, and K. Poduri, ‘’An expedited forwarding PHB,’’, RFC 2598, June 1999. [5] C. Dovrolis, ‘’Proportional Differentiated Services for the Internet,’’ in PhD thesis, University of Wisconsin-Madison, December 2000. [6] C. Dovrolis and P. Ramanthan, ‘’Proportional differentiated services, part II: Loss rate differentiation and packet dropping,’’ in Proc. of IWQoS 2000, Pittsburgh, PA, June 2000, pp. 5261. [7] Y. Moret and S.Fdida, ‘’A proportional queue control mechanism to provide differentiated services,’’ in Proc. of the International Symposium on Computer and Information System (ISCIS), Belek, Turkey, October 1998, pp. 17-24. [8] T. Nandagopal, N. VentikaramanR. Sivakumar, and V. Bharghavan, ‘’Delay differentiation and adaptation in core stateless networks,’’ in Proc. of IEEE INFOCOM 2000, Tel-Aviv, Israel, April 2000, pp. 421-430. [9] T. Ngo-Quynh, H. Karl, A. Wolisz and K. Rebensburg, ‘’Relative Jitter Packet Scheduling for Differentiated Service,’’ in Proc. of 9th IFIP Working Conference on Performance Modelling and Evaluation of ATM&IP Networks IFIP ATM&IP 2001. [10] T. Ngo-Quynh, H. Karl, A. Wolisz and K. Rebensburg, ‚’The Influence of Proportional Jitter and Delay on End to End Delay in Differentiated Service Network,’’ in Proc. of IEEE International Symposium on Network Computing and Application NCA 01, Cambrige, MA, USA. October 2001. [11] N. Shivakumar, C. J. Sreeman, B. Narendran and P. Agrawal, ‘’The Concord algorithm for synchronization of networked multimedia streams,’’ in Proc. Of International Conference on Multimedia Computing and Systems, 1995. [12] UCB/LBNL/VINT Network Simulator-ns (version 2), http://wwwmash.cs.berkeley.edu/ns/ns.html Normalized End-to-End Delay (s) 17,4 17,35 17,3 NT1 17,25 NT2 17,2 NT3 17,15 NT4 17,1 NT5 17,05 NT6 17 16,95 16,9 0 3,415 6,83 10,25 13,66 17,08 20,49 23,9 27,32 30,73 34,15 37,57 40,98 44,4 47,81 Time (s) Figure 5. Performance Comparison between different Network Topologies NT1 Loss 17,070029 5% Loss 15,1934032 10% Loss 13,0381438 15% NT2 17,2595945 NT3 17,1555514 NT4 17,1347128 NT5 17,2633509 NT6 17,1870795 16,8002956 15,2134038 16,1908581 17,2265406 16,3924606 15,2202508 13,1752436 13,247668 17,2633509 14,8793064 Table 2. Performance of NT1 to NT6 with different loss probabilities 18 Normalized Delay (s) 17 16 Loss 5% 15 Loss 10% 14 Loss 15% 13 12 NT1 NT3 NT4 NT6 NT2 NT5 Figure 6. Performance Comparison with different loss probabilities