Quality of Service An Introduction

Transcription

1 An Introduction Mike Weaver 1

2 Session Objectives The new CCNP-ONT course introduces the concept of providing a Quality of Service framework in a converged IP (packet switched) network This session aims to introduce you to some of the fundamental components associated with this technology, so that.. at the completion of this session, you will be able to:- Explain the necessity for Quality of Service (QoS) in converged networks and describe and define some of the fundamental terms associated with incorporating QoS (e.g. bandwidth, delay, packet loss, queues, queuing, etc.) 2

3 Beyond best effort what is QoS? QoS is a term used to refer to technologies used to minimise some of the negative effects associated with congestion in a packet switched network Packet switched networks rely on queue (or buffer) availability at all ingress and egress points of network connected devices to smooth the bursty nature of computer generated traffic These buffers and the associated queues were the essential components of packet networks that allowed a far better performance for this type of traffic than traditional circuit switched networks 3

4 Dealing with congested networks the problems However, if these networks experience congestion at any node, they must queue traffic packets until there is sufficient capacity to allow onward transmission, leading to variable delays in delivery In early switched networks, all traffic was treated equally, regardless of its degree of elasticity Device buffers are finite, leading to packet loss if a buffer (queue) is full and cannot accept further packets In addition, queue length grows exponentially as the packet arrival rate approaches the transmission rate 4

5 Dealing with congested networks the solutions There are two fundamental techniques that can be used to attempt to alleviate these congestion problems in a converged IP network 1. Over provision the network. If there is no congestion, then essentially there are no problems! 2. Deploy QoS in the network, differentiating between inelastic traffic which must be delivered with minimal (and largely predictable) delay, and elastic traffic that can usually be queued without difficulty 5

6 A joint approach In reality, both approaches over provisioning and configuration of network devices to incorporate QoS techniques are often combined in modern networks QoS should generally be deployed in those parts of the network that are permanently or occasionally overloaded It is common that peripheral parts of a network are overloaded, and the backbone is over-provisioned 6

7 Delay in a perfect network t1 t time time d1 d2 d3 t1 t2 d1=d2=d3 7

8 Delay in a practical network T time time d1 d2 d3 d6 T1' d1 d2 d3 d6 T1 T1' 8

9 Review Questions The Practical Network Why are the delays different? What are some of the causes of packet loss? What effects will overall delay have on a traffic flow? Voice call? Streaming media? FTP traffic? Telnet traffic? What about variations in delay within the same flow? How can this variation be handled? What is delay a function of? 9

10 End to end packet delivery delay It can be seen that total packet delivery delay is a random value There will always be a minimum delay due to various physical characteristics of the network signal propagation, serialisation etc the End to End Delay 10

11 End to End Packet Delivery Delay End/End Delay = (P1 + Q1 + P2 + Q2 + P3 + Q3 + P4) ms Processing and Queuing Delay Q1 Processing and Queuing Delay Q2 Processing and Queuing Delay Q3 ` Propagation and Serialisation Delay P1 Propagation and Serialisation Delay P2 Propagation and Serialisation Delay P3 Propagation and Serialisation Delay P4 11

12 Delay of packet delivery There will also be a variable component due to additional buffering; queuing, etc and the amount of congestion within the network To analyse the delay, we need to create a distribution histogram a method that will allow us to compute the delay variability within the transmission of a very large number of packets ( ) 12

13 Delay histogram Delay Distribution The greater the number of packets measured, the more precisely the histogram reflects network delay For example, using this data allows us to state that the probability of packets having a delay < 54mS is 60% # of packets Delay (ms) 13

14 Delay distribution density function End to end delay for a typical connection smoothed over an infinite number of packets 14

15 Review Questions Delay characteristics Are delay characteristics important in an underprovisioned network that includes both elastic and inelastic traffic? If so, which characteristics? Why does this need to be a statistical analysis? 15

16 Bandwidth Transmission rates can be measured between any two interface elements in a network However, the network bandwidth is always equal to the path that has minimum bandwidth To increase bandwidth, it is necessary to deal with the slowest elements, or bottlenecks 16

17 QoS Techniques Now that we are aware of some of the factors to be considered in analysing network congestion, we can investigate how congestion problems can best be minimised for traffic that requires predictable characteristics...also called inelastic traffic Providing a known, or predefined, specific quality of service to inelastic packet streams is an essential tool in modern converged networks 17

18 QoS Support In most converged networks, there will be multiple information flows at any given point in time Each flow will need servicing according to some prerequisite QoS requirement As a minimum, each flow will be serviced by two queues at each network device A switch process queue ( input buffer ) An output interface queue ( output buffer ) To ensure that the required QoS can be achieved, it is essential that the utilisation coefficient of each resource serving the specific flow does not exceed the predefined value 18

19 Utilisation vs delay Maintaining utilisation < 90% for elastic traffic is considered congestion control and has been refined over many years To achieve a lower utilisation for inelastic traffic requires a separate queue for each resource at the output We Ws ρ 19

20 Queuing Models Departures Arrivals Queue (Waiting Area) Link (Server) Single FIFO Queue High Priority Queue (Waiting Area) Departures Arrivals Classify Link (Server) Priority Queuing Model Low Priority Queue (Waiting Area) 20

21 First In First Out (FIFO) queue This is the traditional queue or buffer also sometimes called First Come First Served (FCFS) Packets are placed into a single common queue in the order they arrive and retrieved in the same order This has generally been the default queuing algorithm used in packet switching devices, but it affords no opportunity to differentiate between different types of information flow All traffic types are treated equally 21

22 Priority Queuing A priority queuing mechanism divides traffic into a small number of classes, and assigns some priority characteristic to each class before placing the assigned packets in separate queues Each traffic class (queue) can then be treated according to this characteristic The actual mechanism used to classify the traffic is a separate process If there are packets in the higher priority queues then they are serviced before the lower priority queues using a scheduling algorithm 22

23 Priority Queuing If the traffic flow assigned a high priority is a small component of the overall traffic say a single VoIP conversation this approach works well It will nearly always provide the high priority flow the bandwidth it requires with minimal delay Lower priority (elastic) traffic will hardly be impacted, as the delay whilst the (relatively small) VoIP flow is serviced is largely transparent 23

24 Priority Queuing High 24

25 Weighted Queuing aka Custom Queuing An alternative to using a priority queue is to weight the queues such that they guarantee some minimum percentage of the available bandwidth to different classes of traffic when there is network congestion This is similar to priority queuing except each queue is allocated a percentage of available bandwidth rather than a specific priority 25

26 Weighted queuing 26

27 Weighted queuing The queues are serviced in a round robin fashion within a specific time period Assume the servicing cycle is 1s. Then the queues in the diagram would be activated for 150ms; 150ms; 400ms and 300ms If the output interface b/w is 100Mb/s, then 15Mb of data would be retrieved from the first queue during its service interval; 40Mb from the third queue etc Each queue receives guaranteed minimum b/w, which in most cases is more acceptable than suppressing low priority traffic 27

28 Weighted queuing Note that since the data is retrieved as packets, and not as bits, the actual value of allocated bandwidth will vary The variation will be a function of the queue cycle time a longer time means that the actual bandwidth tends towards the theoretical value but longer cycle times mean longer delays in transmission 28

29 Weighted queuing When a weighted queuing algorithm is used, the utilisation coefficient for a traffic class significantly influences the delay and delay variation for that class of packets For example, if 10% of the bandwidth of 100Mb/s is allocated to the queue, and the average rate of flow is 3Mb/s, then the utilisation coefficient is 3/10 = 0.3 Delays would be insignificant at such a value. However, if the flow was 9Mb/s, the queue would grow significantly! 29

30 Weighted Fair Queuing A modified approach to weighted queuing is to weight each queue equally, thus ensuring that each queue is treated equally or fairly Note that in both methods, if a queue is empty at its scheduled retrieval time, then it is omitted and the time available is distributed between all other queues according to each queue s weighting Flow based weighted fair queuing is one of Cisco s premier queuing techniques * and is the default on most serial interfaces at or below E1 * CCNP-ONT Curriculum

31 Priority vs Weighted Queuing Priority queuing ensure minimum delays for traffic of the highest priority It does not provide any guarantees for lower priority traffic flows Weighted queuing guarantees a mean traffic rate, but doesn t provide any guarantees about delay However, all flows get a guaranteed rate 31

32 Hybrid queuing To attempt to find some middle ground between the two extremes of priority vs weighted queuing, hybrid schemes have been developed The most popular scheme uses a single priority queue for the highest priority traffic, and weights each of the remaining queues (Cisco call this class based weighted fair queuing or CBWFQ) The priority queue is used for real-time traffic (i.e. voice) and the remaining queues for other traffic of varying degrees of elasticity according to need Obviously, there must be a mechanism to limit priority traffic from consuming all available bandwidth! 32

33 Questions Queuing If a router is managing two flows, one with packets of 500 bytes and the other with packets of 1000 bytes, what is the effect on bandwidth of a simple fair queue? What can be done to change this? Would a system that incorporates fair queuing have a longer overall delay than one that didn t? Why? 33

34 Queue management Using queuing algorithms is in response to congestion that is already apparent on the network congestion control mechanisms However, QoS techniques also incorporate methods that attempt to predict and prevent network congestion congestion avoidance mechanisms The aim of congestion avoidance is to prevent congestion, since it is better to transmit data at a lower rate without loss, than at a higher rate and lose packets during periods of congestion 34

35 Queue management Such mechanisms can be viewed as control processes that act on an inherently unstable ( open loop ) system, attempting to achieve a degree of stability The primary Cisco QoS tool used for congestion avoidance is known as Weighted Random Early Detection (WRED) 35

36 Weighted Random Early Detection WRED is a process by which packets are discarded randomly and earlier (i.e. before they would have been if the queue buffer became full) as congestion begins to increase at a network bottleneck The dropping of packets will cause transport protocols such as TCP (but NOT UDP) to throttle back the rate at which packets are launched into the network, (hopefully!) leading to an easing of the congestion The weighting in this process refers to preferential traffic handling for high priority packets, allowing lower priority packets to be discarded first 36

37 Classification If packets or more correctly, flows are to be identified for placing in different queues, then it is necessary to have a method of classification 37

38 Questions How much QoS is there currently on the Internet? What is the major issue with QoS on a public network? How should traffic be classified? How do we handle UDP avoidance? What is best to drop voice packets or ftp packets? 38

39 QoS Summary This presentation has looked at some of the fundamental components and issues associated with QoS The techniques described are primarily ones that are used on a hop by hop basis There are other methods not discussed here that attempt to provide guarantees from end to end However, remember that currently! the Internet is a best effort network so there is no guarantee the mechanisms will work across public links 39

40 QoS Summary And a timely last word on the arguments between those who suggest that more Internet bandwidth is the answer vs. those who suggest that traffic should be differentiated. 40

41 QoS Summary We don t need no reservation We don t need admission control All applications must be adaptive The Net works fine, so leave it alone Hey! Professor! Leave the Net alone! All we want is just flat rate pricing for all All we want is just flat rate pricing for all.. We don t need no traffic management Overprovision bandwidth for all The only true god is TCP/IP The Net isn t broken, so leave it alone Hey! Professor! Leave the Net alone! All we want is just flat rate pricing for all All we want is just flat rate pricing for all. (With thanks to Keshav, S; An Engineering Approach to Computer Networking; Addison-Wesley; 1997 and sung to Pink Floyd s The Wall ) 41

42 Questions? 42