Differentiated Services

Presentation on theme: "Differentiated Services"— Presentation transcript:

1 Differentiated Services
2002년 5월 28일 길홍렬

2 Outline Introduction Drop Strategies and Loss-rate Differentiation
Integrated Services Differentiated Services Drop Strategies and Loss-rate Differentiation On the Impact of Policing and Rate Guarantees in Diff-Serv Network Dynamic Class Selection Reference

3 Introduction Quality of service Present service: best effort service
A set of service requirement to satisfied Delay , Jitter , Reliability … Present service: best effort service Service quality에 대한 demand 증가 High reliability for their web site Low delay and jitter for Internet Telephony, Video Conferencing Internet QoS 제공하기 위해 제안된 framework Integrated Services / RSVP Differentiated Services

4 Integrated Services Resource reservation에 의한 QoS 지원
data 전송 시작 전에 RSVP를 이용하여 path setup, resource reserve Router는 flow별로 state를 유지해야 한다

5 IntServ Components(1) The signaling protocol (e.g. RSVP)
To create and maintain flow-specific state in endpoint hosts and in routers along the path of a flow The admission control routine When router receives a RESV message Decide whether a request for resources can be granted Routing Agent Reservation Setup Agent Management Traffic Control DB Routing DB Admission Control

6 IntServ Components(2) The classifier The packet scheduler
When router receives a packet Perform a Multi-Field classification and put the packet in a specific queue The packet scheduler When router forwards a packet Schedule the packet to meet its QoS requirement Packet Scheduler Input Driver Internet Forwarder Output Classifier

7 IntServ Problems router마다 유지해야 하는 state 정보가 flow 에 비례하여 커진다.
특히 flow의 수가 많은 core router에서 큰 문제 모든 router마다 IntServ를 위한 component를 전부 구현해야 한다.

8 Differentiated Services
A relative-priority scheme Marking the DS field of packets differently, and handling packets based on their DS field Premium Service: low jitter, low delay Assured Service: better reliability Customer have a SLA with its ISP When packet enters one domain from another re-marked as determined by SLA between two domain ISP의 border router에서 incoming packet에대 packet 을 classification, marking, shaping core router는 class 별로 service

9 DiffServ Example Packets are classified, policed, and possibly shaped
packets are processed based on DS field DS field of Packets are marked

10 DiffServ Border Router
Classification Multi-Field classification Marking setting the DS field in a packet token bucket을 통해 traffic의 rate과 burstiness 제한 Policing handling out of profile traffic Shaping delaying packets within traffic stream to cause it to conform to some defined traffic profile

11 DiffServ Marker Main difference between Premium and Assured
Token bucket controls both the rate and the burstiness Bucket is filled to the maximum packet size (peak rate) (shaping or dropping) Bucket is filled to the configured rate and burst Parameter (No dropping)

12 DiffServ Core Router For assured service For premium service
In and out packets use same queue(Assured Queue) Queue is managed by RIO(RED with In and Out) RIO In-packets are less-dropped (more reliable) Maintains two RED algorithm: for in and for out-packets Queue size is between two threshold, only out packets are randomly dropped. Queue size is exceed second threshold, in and out packets are randomly dropped For premium service All packets marked with P-bit enter a PQ(Premium Queue) Packets in the PQ will be sent before packets in the AQ

13 DiffServ Advantage Number of service classes is limited
Amount of state information is proportional to the number of class (c.f number of flow in Integrated services) Scalable Sophisticated operations are only needed at boundary of the network Less cost Simple core routers can forward packets very fast Incremental deployment is partially possible

14 Drop Strategies and Loss-rate Differentiation

15 Motivation Assured Forwarding PHB에서 Differentiating queue mechanism을 위해 사용되는 대부분의 drop strategy는 ‘drop-arrivals’ drop-arrivals strategy는 drop 이 필요할 때, 새로 들어온 packet에 대해서만 drop 여부를 결정 drop이 필요할 때, 새로 들어오는 packet의 drop precedence level이 모두 낮다면 (in-profile packet) drop 지연 지연된 drop은 burst한 loss와 더 큰 jitter를 야기

16 Weighted RED Drop precedence level별로 서로 다른 확률로 drop 여부를 결정

17 Drawback of Weighted RED
drop precedence level이 0인 packet 만 도착하여 average queue length가 계속 증가하는 상황 loss-rate이 급격하게 높아 질 수 있다 average queue length가 max_th(1) 를 초과한 상황에서는 drop precedence level이 1인 packet 이 burst하게 drop될 수 있다 burst한 drop은 queue oscillation을 야기

18 WRED with the gentle modification
high loss-rate 문제를 완화할 수 있다 문제점 queue oscillation 문제를 해결할 수 없다 long queue length 야기

19 Drop-from-queue Mechanism
새로 도착하는 packet 뿐만 아니라 queue 내에 있는 packet도 drop 할 수 있다. drop the arriving packet if it is an out packet otherwise, drop the last out packet present from within in the queue congestion 상황에서, queue 안에 drop precedence level이 높은 packet (out-of-profile packet) 이 최소한 하나라도 있다면 즉시 drop 할 수 있다. drop-arrivals scheme에 비해 avg_ql가 더 적게 증가하므로 queue oscillation 문제가 완화된다

20 Simulation Setup link R3-R4 is congested

21 Simulation with TCP Reno and Standard WRED (1)
Loss rates

22 Simulation with TCP Reno and Standard WRED (2)
Throughputs

23 Simulation with TCP Reno and Standard WRED (3)
Average queue length

24 Simulation with TCP Reno and Gentle WRED (1)
Loss rates

25 Simulation with TCP Reno and Gentle WRED (2)
Throughputs

26 Simulation with TCP Reno and Gentle WRED (3)
Average queue length

27 Summary 일반적인 drop scheme은 새로 도착한 packet에 대해서만 drop 여부를 결정한다
high loss-rate, queue oscillation 야기 Gentle WRED high loss-rate 을 완화 queue oscillation은 미해결 increase the long-term average queue length Dropping packet from the queue queue oscillation 해결 lower loss-rates and less bursty loss pattern a

28 On the Impact of Policing and Rate Guarantees in Diff-Serv Network : A Video Streaming Application Perspective

29 Introduction General goal application에 따라 많은 차이 존재 method
DiffServ가 제공하는 network guarantee와 DiffServ를 사용하는 application의 actual benefit과의 관계 이해 application에 따라 많은 차이 존재 application을 streaming video로 한정 method Expedited Forwarding PHB 에서 token rate, token bucket depth 에 따른 simulation을 통해 user가 인지하는 video quality를 측정

30 Streaming Video Technologies
Streaming Video Server IBM Video Charger support MPEG-1, MPEG-2 burstiness가 적은 편 Windows Media Technologies support ASF, WMV burstiness가 큰 편

31 Video quality Measurement(1)
no standard procedure ITS Video Quality Measurement tool을 사용 원본 frame과 수신된 frame sequences에서 video quality를 측정 원본 frame의 quality와 수신된 frame의 quality를 비교하여 video quality parameter를 계산 video quality parameter를 통해 quality score를 산출 video quality parameters that is highly correlated with the subjective assessments of human viewer panels

32 Video quality Measurement(2)
Storing Received Video Frames to File Video quality measurement 의 overhead가 커서 real-time 으로 video streams을 처리하지 못하기 때문 video client는 Windows platform 사용 Directshow architecture를 통한 video, audio contents 재생 Renderer를 storage filter로 교체

33 Video quality Measurement(3)
Capturing Network Dynamics Information frame을 binary file에 저장할 때, frame마다의 arrival time과 target presentation time을 별도의 ASCII file에 저장한다 이러한 timing information은 network에 의해 야기된 손상을 emulation하게 된다 대부분의 renderer는 lost되거나 delayed frames으로 인한 손상을 완화하기 위한 technique 사용 buffering, repeating last frame

34 Local Testbed video server 와 video client 를 연결하는 몇 개의 DiffServ capable routers 로 구성

35 QBone Testbed inter-domain DiffServ testbed

36 Video Clip

37 Experimental Configurations

38 QBone Testbed Results(1)
Lost clip/1.5 Mbps encoding

39 QBone Testbed Results(2)
Lost clip/1.0 Mbps encoding

40 QBone Testbed Results(3)
Dark clip/1.5 Mbps encoding

41 QBone Testbed Results(4)
Dark clip/1.0 Mbps encoding

42 QBone Testbed Results(5)
Frame loss and Relative Quality for Dark clip Token bucket depth : 3000 bytes

43 Local Testbed Results Lost clip/1.0 Mbps encoding

44 Summary DiffServ EF PHB를 제공하는 network에서 다양한 configuration 과 video quality와의 관계를 실험적으로 고찰 frame loss로는 정확한 video quality 측정 불가 video type의 차이는 network configuration과 연관이 적다 server type과 encoding rate이 video quality와 frame loss에 영향을 미친다 encoding rate보다 token rate이 커야 video의 high quality를 보장할 수 있다 token bucket depth를 약간만 증가시켜도 video quality를 상당히 개선할 수 있다

45 Dynamic Class Selection: from Relative Differentiation to Absolute QoS

46 Introduction Relative Differentiation Model
N개의 service class 제공 높은 class의 서비스가 delay와 loss 측면에서 더 좋음을 보장 no admission control or resource reservation 자신이 원하는 quality가 만족되지 않으면 application level에서 class를 하나 증가시킨다. strict QoS guarantee를 요구하는 application을 제외한 나머지를 위한 model Dynamic Class Selection (DCS) 개개의 user들이 독립적으로 자신의 delay requirement를 만족하는 최소의 class를 선택

47 DCS Model User perform class transitions
user j는 현재 사용하고 있는 class cj 를 통해 겪고 있는 average queueing delay 를 구한다 If and cj < N, user moves to cj +1 If and cj > 1, user moves to cj -1

48 DCS Algorithm(1) User specifies a requirement Dmax on the maximum Round-Trip Delay (RTD) Sender timestamps each packet k before transmitting it Receiver returns the timestamps back to the sender Sender measure the RTD Dk minimum RTD Dmin is also measured exponential running-average

49 DCS Algorithm(2)

50 Simulation Topology

51 Simulation Result(1) Static vs DCS for flow with Dmax=100msec
fI=1, fD=4, k=0.9

52 Simulation Result(2) Controlling DCS parameter
to meet a per-packet RTD requirement

53 Simulation Result(3) Effect of the CT delay requirements on a DCS flow

54 Summary Relative differentiation DCS Advantage – simplicity
Disadvantage – don’t provide users with absolute QoS DCS dynamically search for an acceptable class provides absolute QoS and end-to-end service under certain conditions

55 References Xipeng Xiao, Lionel M. Ni, “Internet QoS: the Big Picture”, IEEE Network 1999 Ulf Bodin, Olov Schelen, “Drop Strategies and Loss Rate Differentiation”, IEEE ICNP 2001 W. Ashmawi, R. Guerin, S. Wolf, M. Pinson, “On the Impact of Policing and Rate Guarantees in Diff-Serv Networks: A Video Streaming Application Perspective”, SIGCOMM 2001 Constantinos Dovrolis, Parameswaran Ramanathan, “A Case for Relative Differentiated Services and the Proportional Differentiation Model”, IEEE Network, 1999 Constantinos Dovrolis, Parameswaran Ramanathan, “Dynamic Class Selection: From Relative Differentiation to Absolute QoS”, IEEE ICNP, 2001