Quality of Service and the End to End

  • Slides: 63
Download presentation
Quality of Service and the End to End Argument IEEE Network November/December 2007 Gunnar

Quality of Service and the End to End Argument IEEE Network November/December 2007 Gunnar Karlsson, School of Electrical Engineering of KTH, the Royal Institute of Technology Ignacio Mas, IPTV development at Ericsson Research Presented by Guan-Wei, Chen 9/25/2020 OPLAB_IM_NTU 1/63

About Authors - Gunnar Karlsson Ø Ø Ø receive his M. S. degree from

About Authors - Gunnar Karlsson Ø Ø Ø receive his M. S. degree from Chalmers University of Technology in Gothenburg, Sweden(1983) receive his PH. D. degree from Columbia university(1989), New York research related to quality of service and wireless content distribution 9/25/2020 OPLAB_IM_NTU 2

About Authors - Ignacio Mas u u u receive his M. S. degree in

About Authors - Ignacio Mas u u u receive his M. S. degree in electrical engineering from the Royal Institute of Technology (Stockholm, Sweden) now he is a doctoral student at KTH his research interests are in multimedia networking , quality of service, and operating systems 9/25/2020 OPLAB_IM_NTU 3

Abstract u u u Audio-visual service are now popular in the internet earlier base

Abstract u u u Audio-visual service are now popular in the internet earlier base on batch downloading, now real-time interactive and streaming services grow rapidly can benefit from Qo. S if it widely provided Quality is obtained by means of probe-based admission control then implementation on Application Layer, Transport Layer, Network Layer Compared and discussed the possibility deployment 9/25/2020 OPLAB_IM_NTU 4

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 5

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 6

Introduction u multitude of quality of service, the traffic defined Ø Asynchronous transfer mode

Introduction u multitude of quality of service, the traffic defined Ø Asynchronous transfer mode architecture Ø Integrated service architecture Ø Differentiated service architecture u u u most proposed Qo. S solution within one of three architecture limited Qo. S offering from Internet Service Provider lock of deployed Qo. S solutions 9/25/2020 OPLAB_IM_NTU 7

Introduction u u u little attention Qo. S researchers have given the culture of

Introduction u u u little attention Qo. S researchers have given the culture of network operations to avoid complexity and anticipate failure addition of Qo. S mechanisms would introduce new potential failure mode it might be difficult to debug and has to be weighed against the uncertain benefit of providing Qo. S support Qo. S function traditionally been associated with the network (Ex: admission control , scheduling, and policy ) 9/25/2020 OPLAB_IM_NTU 8

Introduction u u provide some form of Qo. S without changes to the network

Introduction u u provide some form of Qo. S without changes to the network operation or network lifting the Qo. S mechanisms out of the network to higher protocol layers would have advantages: Ø Qo. S is provided from host to host Ø The Qo. S mechanisms may develop independent of the network Ø the management and operation of a network remains unaffected u assigning functions to the layers of a protocol architecture relied on end-to-end argument 9/25/2020 OPLAB_IM_NTU 9

Introduction u u Functions places at low levels of a system may be redundant

Introduction u u Functions places at low levels of a system may be redundant or of a little value when compared with the cost of providing them at that low level Application Layer solution can be without standardization, but gives the least benefit (Ex: prevent setting up session when congestion) Transport Layer solution needs standardization and provides differentiation into two traffic class Network Layer solution can reinforce differentiation 9/25/2020 OPLAB_IM_NTU 10

Introduction -Quality of Service u u u Qo. S is the collective effect of

Introduction -Quality of Service u u u Qo. S is the collective effect of service performances which determine the degree of satisfaction of a user of the service Qo. S is often associated performance metrics are packet loss and delay, and also restricted scope Affect user satisfaction: traffic control behavior、route changes、 sensitivity to service attack and latency in fault handling…. 9/25/2020 OPLAB_IM_NTU 11

Introduction -Quality of Service u TCP attempts to share the rate of a bottleneck

Introduction -Quality of Service u TCP attempts to share the rate of a bottleneck link fairly over all ongoing TCP sessions u Throughput variations when sessions enter or leave and when the total data rate of the sessions changes due to additive-increase multiplicativedecrease congestion control u One possible is complementary policy to block new real time session when congestion, to allow admitted sessions to remain unresponsive 9/25/2020 OPLAB_IM_NTU 12

Introduction -Quality of Service u u u the policy base on human perception of

Introduction -Quality of Service u u u the policy base on human perception of real time streaming interactive service with audio-visual signals There is a trade-off between the distortion and the resulting bit rate Lower data rate, higher the distortion Noticeable quality variations reduce user satisfaction even average quality is high, so our interesting in limiting the throughput variations Use probe-base admission control 9/25/2020 OPLAB_IM_NTU 13

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 14

Probe-Based Admission Control u u u base on transmission of a probe which consists

Probe-Based Admission Control u u u base on transmission of a probe which consists of a stream of probe packets used to infer the state of a network path lasts a few seconds and constitutes small addition to total session length Probe rate (rpr) is constant, and session rate is not allowed to exceed it the mean of the data rate of the session and distribution can’t know in advance when the rate is variable 9/25/2020 OPLAB_IM_NTU 15

Probe-Based Admission Control 9/25/2020 OPLAB_IM_NTU 16

Probe-Based Admission Control 9/25/2020 OPLAB_IM_NTU 16

Probe-base admission control Receiver measures the loss and possibly the delay for the probe

Probe-base admission control Receiver measures the loss and possibly the delay for the probe u Admission control classified in u Per-hop and End-point (Probe-base ) u Need queue to save probe ü ü u Double queues Threshold-queue The acceptance decision given that 9/25/2020 OPLAB_IM_NTU 17

Probe-base admission control u u u The backoff randomizes new setup attempts if the

Probe-base admission control u u u The backoff randomizes new setup attempts if the failure was too many simultaneous probes Probe packets contain session and control data Control data include parameters for the configuration of the receiving application Session data can be a greeting or ring tone Probe packet stream can use for clock synchronization and allowing the jitter removal system to settle into steady state Use data rate of the session for decision rather than probe rate , the to leave group more reliable and to remain group less reliable 9/25/2020 OPLAB_IM_NTU 18

Probe-base admission control u An important of end-to-end admission control is the stability of

Probe-base admission control u An important of end-to-end admission control is the stability of the route for an accepted session u Route changes occur two conditions: ü ü 9/25/2020 Load balancing scheme reducing the load of the link topological changes application notice a disruption in the data flow OPLAB_IM_NTU 19

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 20

Different Layer Qo. S Support (Application Layer Qo. S Support) u u Self-Admission Control

Different Layer Qo. S Support (Application Layer Qo. S Support) u u Self-Admission Control measures the packet loss ratio of the first few seconds and estimate the packet loss probability Two-way sessions – Initiated by the first received probe packet – Carried out by the called application u u Relation between the loss ratio of an initial part of the call and the loss ratio of the whole duration Initial interval increases , the points group closer around the line Y=X 9/25/2020 OPLAB_IM_NTU 21

Different Layer Qo. S Support (Application Layer Qo. S Support) 9/25/2020 OPLAB_IM_NTU 22

Different Layer Qo. S Support (Application Layer Qo. S Support) 9/25/2020 OPLAB_IM_NTU 22

Different Layer Qo. S Support (blocking threshold) u denote lp : the loss ratio

Different Layer Qo. S Support (blocking threshold) u denote lp : the loss ratio of an initial interval lt : the loss ratio of the total call ε: a stochastic variable , is lt - lp la : pre-established loss rate Use statistics method: Strict policy : higher risk than 10% => la-2. 6% Relax policy : more than 90% => la+1. 81% 9/25/2020 OPLAB_IM_NTU 23

Different Layer Qo. S Support (Application Layer Qo. S Support) Ø Different kinds of

Different Layer Qo. S Support (Application Layer Qo. S Support) Ø Different kinds of calls based on the initial estimation and the final outcome 9/25/2020 OPLAB_IM_NTU 24

Different Layer Qo. S Support (Application Layer Qo. S Support) u Advantage: – Not

Different Layer Qo. S Support (Application Layer Qo. S Support) u Advantage: – Not be loaded by new sessions when congestion – Interactive session is not disturbed by session that risks being aborted due to poor quality – Receiver may avoid starting to listen to or view a session that is likely to be rendered useless by poor throughput – Although it is much simpler to implement than feedback congestion control, it might be desirable to provide probe-based admission control at the transport layer as a generic service to all applications that might benefit from it 9/25/2020 OPLAB_IM_NTU 25

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u Provide

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u Provide two generic traffic classed : Batch controlled by TCP and Streaming use probe-based admission control All sessions to enter the network are congestion controlled probe-based admission control makes the aggregate of streaming sessions responsive to load variations by blocking new sessions to prevent the load from increasing 9/25/2020 OPLAB_IM_NTU 26

Different Layer Qo. S Support (Transport Layer Service differentiation ) ØFair sharing of bottleneck

Different Layer Qo. S Support (Transport Layer Service differentiation ) ØFair sharing of bottleneck capacity between sessions of the two classed by choosing the admission criterion appropriately 9/25/2020 OPLAB_IM_NTU 27

Different Layer Qo. S Support (Transport Layer Service differentiation ) u acceptance decision Wmax

Different Layer Qo. S Support (Transport Layer Service differentiation ) u acceptance decision Wmax : maximum window size RTT : the round-trip time b : the number of packets acknowledged by a received ACK T 0: timeout value If rpr <= r. TCP , then accept the session 9/25/2020 OPLAB_IM_NTU 28

Proof u Reference: Modeling TCP Reno Performance: A Simple Model and Its Empirical validation

Proof u Reference: Modeling TCP Reno Performance: A Simple Model and Its Empirical validation Yi=αi+ Wi -1 9/25/2020 OPLAB_IM_NTU 29

Windows size Proof Other Yi calculated (E[β]=E[W] /2) 9/25/2020 OPLAB_IM_NTU 30

Windows size Proof Other Yi calculated (E[β]=E[W] /2) 9/25/2020 OPLAB_IM_NTU 30

Proof Use above estimation combine E[X] and E[A] 9/25/2020 OPLAB_IM_NTU 31

Proof Use above estimation combine E[X] and E[A] 9/25/2020 OPLAB_IM_NTU 31

Proof B(P)=Mi/Si 9/25/2020 OPLAB_IM_NTU 32

Proof B(P)=Mi/Si 9/25/2020 OPLAB_IM_NTU 32

Proof Using L’Hopital rule, can calculate Q R is geometric distribution, so E[R] T

Proof Using L’Hopital rule, can calculate Q R is geometric distribution, so E[R] T 0 (time out) is most 64 T 0, E(ZTO) is 9/25/2020 OPLAB_IM_NTU 33

Proof Substituting expressions for Q, E[S], E[R] , and E[ZTO] for B(P) Where 9/25/2020

Proof Substituting expressions for Q, E[S], E[R] , and E[ZTO] for B(P) Where 9/25/2020 OPLAB_IM_NTU 34

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u An

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u An issue concerns the quality admitted streaming sessions Session will be disturbed by new streaming session that probe the network for admission The admitted sessions should be protected by forward error correction with error control codes to withstand the loss by the disturbing sessions The loss estimate from the probing can be used to calculate the appropriate level of redundancy necessary for insulting the session in this way 9/25/2020 OPLAB_IM_NTU 35

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u Probe-based

Different Layer Qo. S Support (Transport Layer Service differentiation ) u u u Probe-based admission control can be used for multicast sessions An admission policy based on estimation of RTTs Require each receiver to estimate it , for example, by means of pings to a rendezvous point The solution is acceptable for a low arrival rate of new receivers An admission policy independent of RTTs is preferable 9/25/2020 OPLAB_IM_NTU 36

Different Layer Qo. S Support (Transport Layer remarks ) u u Batch and streaming

Different Layer Qo. S Support (Transport Layer remarks ) u u Batch and streaming classes can’t be argued which class is better TCP provides a lossless service where throughput variations and retransmissions lead to uncontrolled delay probe-based admission control allows the sender to use a fixed amount of network capacity, and thereby controls delay and loss while introducing uncontrollable blocking Incorporated in the datagram congestion control 9/25/2020 OPLAB_IM_NTU 37

Different Layer Qo. S Support (network support for service differentiation) u u u Scheduling

Different Layer Qo. S Support (network support for service differentiation) u u u Scheduling in the network can provide further differentiate services allow an operator to control the blocking probability for the streaming class control the average throughput for the batch class 9/25/2020 OPLAB_IM_NTU 38

Different Layer Qo. S Support (Queuing system) Denote that Cl : link capacity Cb

Different Layer Qo. S Support (Queuing system) Denote that Cl : link capacity Cb : for elastic batch transfers Cs : for admission controlled streams 9/25/2020 OPLAB_IM_NTU 39

Different Layer Qo. S Support (Queuing system) u u u A streaming session is

Different Layer Qo. S Support (Queuing system) u u u A streaming session is admitted if there is enough capacity available for its probe within link capacity There must be a single target loss level for the admission of all hosts using the service class no need to consider TCP fairness in the admission decision because of classes separated by scheduling 9/25/2020 OPLAB_IM_NTU 40

Different Layer Qo. S Support (Advantage of network support) Ongoing streaming sessions are disturbed

Different Layer Qo. S Support (Advantage of network support) Ongoing streaming sessions are disturbed by neither batch transfers nor probes u The operator can control the blocking probability for streaming sessions by adjusting the capacity allocation, Cs u The delay may be reduced by configuring short buffers for streaming class u Multicast admission control u 9/25/2020 OPLAB_IM_NTU 41

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 42

Comparison of the Proposals u u u A bottleneck link with 10 Mb/s capacity

Comparison of the Proposals u u u A bottleneck link with 10 Mb/s capacity shared between batch and streaming 10 batch sessions and peak rate limited to 500 kb/s Streaming sessions consist of a two-state Markov chain (on and off) with 250 kb/s in the on state , with an average rate of 100 kb/s Probing time is 2 s at the peak rate Corresponding to 976 probe packets of 64 bytes Each case has been run 30 times 9/25/2020 OPLAB_IM_NTU 43

Comparison of the Proposals 9/25/2020 OPLAB_IM_NTU 44

Comparison of the Proposals 9/25/2020 OPLAB_IM_NTU 44

Comparison of the Proposals (Results) u u u (no blocking) streaming sessions are freely

Comparison of the Proposals (Results) u u u (no blocking) streaming sessions are freely admitted and experience severe packet loss (Self) The loss in streaming sessions has been reduced at the offer expense of blocking (TP 1) bounds the loss through the formula for equivalent TCP throughput (TP 2) provides a loss target, which is met, but blocking goes up and TCP is favored (TP+NP) network support allows the blocking to be controlled in relation to the TCP share of the link capacity 9/25/2020 OPLAB_IM_NTU 45

Comparison of the Proposals (Results) u u u Self-admission control is useful since the

Comparison of the Proposals (Results) u u u Self-admission control is useful since the admitted streaming sessions meet the target of at most 1 percent loss Streaming load and particular decision level can balance between streams and batch classes Provide two generic service rather than each streaming application having its own admission control 9/25/2020 OPLAB_IM_NTU 46

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 47

Deployment and Operator Application u u u End to end argument applied to Qo.

Deployment and Operator Application u u u End to end argument applied to Qo. S is the separation of roles between network operation and host function Incognizant of the differentiation at the transport layer Admission control at the application layer Self-admission is incorporated into streaming and conversational application without standardization Weakness: can’t mandate and reinforced by provider Admission control protect user establishing session during congestion 9/25/2020 OPLAB_IM_NTU 48

Deployment and Operator Transport u u u Transport layer differentiation provide two service classes:

Deployment and Operator Transport u u u Transport layer differentiation provide two service classes: batch and streaming Large batch is efficiently made at a fixed rate Small batch could be inefficient because of substantial probe phase Admission control decision policy Datagram congestion control Operators enforce the traffic entering networks is congestion controlled by hosts 9/25/2020 OPLAB_IM_NTU 49

Deployment and Operator Transport u u u u Operator’s role is further affected by

Deployment and Operator Transport u u u u Operator’s role is further affected by Qo. S because of service class differentiation in the network Operators manage the capacity allocation for the streaming class Allocation to control the blocking probabilities of the link If to keep blocking probability below a declared value, users need to pay charge for guarantee Hosts manage the admission decisions of session Decisions are time critical, fully distributed, good scaling Network control blocking probabilities of the link, occurring less frequently than admission decision 9/25/2020 OPLAB_IM_NTU 50

Deployment and Operator Network u u Network support Qo. S affects peering between operators

Deployment and Operator Network u u Network support Qo. S affects peering between operators Agree on the policy for resource allocation about a guaranteed upper limit for blocking or fairness between classes offering low delay for streaming, isolating the class well, and control capacity sharing Offering low delay for streaming by configuring short buffers , being able to control capacity sharing, especially when Qo. S solutions in higher layers used widely 9/25/2020 OPLAB_IM_NTU 51

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 52

Relate Work u Probe-base admission control is not the only one possible , Ex:

Relate Work u Probe-base admission control is not the only one possible , Ex: ECN( Explicit congestion notification) u ECE (Echo) CWR (Congestion Windows Reduced) is variable to manage sliding windows u 9/25/2020 OPLAB_IM_NTU 53

Relate Work(ECN) 9/25/2020 OPLAB_IM_NTU 54

Relate Work(ECN) 9/25/2020 OPLAB_IM_NTU 54

Relate Work u u u ABE (alternative best effort) defined two type packet ,

Relate Work u u u ABE (alternative best effort) defined two type packet , delay-sensitive mark green and throughput-sensitive mark blue Green packet are more likely to be dropped than blue packet Blue packet accepted if its duplicated was accepted in the virtual queue Green packet accepted if it pass green acceptance test router implementation based on DSD (Duplicate Scheduling with Deadlines) 9/25/2020 OPLAB_IM_NTU 55

Relate Work(ABE) 9/25/2020 OPLAB_IM_NTU 56

Relate Work(ABE) 9/25/2020 OPLAB_IM_NTU 56

Relate Work(DSD) 9/25/2020 OPLAB_IM_NTU 57

Relate Work(DSD) 9/25/2020 OPLAB_IM_NTU 57

9/25/2020 OPLAB_IM_NTU 58

9/25/2020 OPLAB_IM_NTU 58

Relate Work u u u End to end argument has been discussed by Moors,

Relate Work u u u End to end argument has been discussed by Moors, saying congestion is phenomenon and endpoint can’t limit congestion and network responsible for endpoint offer excessive traffic This paper counterargument is that congestion is how endpoints inject traffic into the network Probe-based admission control and TCP reduced the risk of congestion 9/25/2020 OPLAB_IM_NTU 59

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different

Today’s Agenda 1 -Introduction About Quality of Service 2 -Probe-Based Admission Control 3 -Different Layer Qo. S Support 4 -Comparison of the Proposals 5 -Deployment and Operator 6 -Related Work 7 -Conclusion 9/25/2020 OPLAB_IM_NTU 60

Conclusion u u u Capacity planning is a necessary long-term task Traffic volumes have

Conclusion u u u Capacity planning is a necessary long-term task Traffic volumes have to be measured and future demands forecast Necessary capacity increases in the network Additional capacity must be leased or procured and installed Mistakes in forecasting make Ø the network susceptible to long-term congestion Ø the operation uneconomical due to unnecessary investments u The cost of controlled admission is the probing phase 9/25/2020 OPLAB_IM_NTU 61

Conclusion (Solution) 1) by application developers implementing selfadmission control 2) by the standardization of

Conclusion (Solution) 1) by application developers implementing selfadmission control 2) by the standardization of congestion control for TCP including operating systems’ protocol stacks 3) by service class differentiation in the network 9/25/2020 OPLAB_IM_NTU 62

Thanks for your listening!!! 9/25/2020 OPLAB_IM_NTU 63

Thanks for your listening!!! 9/25/2020 OPLAB_IM_NTU 63