1. IP traffic and QoS control : the need for flow aware networking Jim Roberts http://perso.rd.francetelecom.fr/roberts France Telecom R&D NSF-COST Workshop Hania, 23-25 June 2003

2. 2 QoS and the statistical nature of demand  assuring QoS relies on understanding the traffic performance relationship demand capacityperformance l volume l characteristics l bandwidth l how it is shared l response time l loss, delay

3. 3 IP traffic is variable rate  IP traffic is "self-similar"...  variability at all time scales  data, video, telephone,... ... because of  heavy-tailed size distributions  (bandwidth sharing by TCP)  consequences  the failure of Poisson modelling?  the failure of the token bucket Tspec! Ethernet traffic, Bellcore 1989

4. 4 Prefer characterization at flow/session level  packets are part of "flows"...  a flow: all packets corresponding to one instance of a given application... ... having the same identifier and occuring within a short time ... flows are part of "sessions"  a succession of flows and "think times"  relating to some homogeneous user activity (e-commerce, mail,...) flow arrivals start of session end of session think times

5. 5 Prefer characterization at flow/session level  packets are part of "flows"...  a flow: all packets corresponding to one instance of a given application... ... having the same identifier and occuring within a short time ... flows are part of "sessions"  a succession of flows and "think times"  relating to some homogeneous user activity (e-commerce, mail,...)  modelling assumption: sessions occur as a Poisson process  in the busy period (like telephone calls!) flow arrivals start of session end of session think times

6. 6 Traffic classification: streaming & elastic flows  streaming traffic  real time audio and video applications  need signal conservation, i.e., "negligible" packet loss and delay  elastic traffic  document transfers (as fast as possible): files, mail, Web, p2p,...  need for throughput conservation, i.e., "negligible" rate reduction with respect to external limits (access line, server capacity,...)  this classification is robust  even as new applications emerge user-network interface network-network interface user-network interface

7. 7 Performance of elastic traffic: an isolated bottleneck  processor sharing model of bandwidth sharing  assume instantaneous fair shares   Pr [n flows in progress] =  n    E [throughput] = C   performance is insensitive to traffic characteristics  only necessary traffic assumption is Poisson session arrivals!  its all a question of underload and overload  generally C   external rate limits  throughput conservation  when , processor sharing model is unstable  very bad performance transparency (   1)congestion (   1) high throughput low throughput

8. 8 Box diagram for flow throughput  simulation results: capacity 10 Mbit/s, offered load 90%  visualization of per flow throughput time occupation startend throughput 10 Mbit/s 200 sec

9. 9 Box diagram for flow throughput > 400 Kbit/s < 20 Kbit/s time 10 Mbit/s 200 sec  simulation results: capacity 10 Mbit/s, offered load 90%  visualization of per flow throughput

10. 10 Performance of elastic traffic: an isolated bottleneck  processor sharing model of bandwidth sharing  assume instantaneous fair shares   Pr [n flows in progress] =  n    E [throughput] = C   performance is insensitive to traffic characteristics  only necessary traffic assumption is Poisson session arrivals!  its all a question of underload and overload  generally C   external rate limits  throughput conservation  when , processor sharing model is unstable  very bad performance transparency (   1)congestion (   1) high throughput low throughput

11. 11 Extension to networks  notions of fairness  max-min fairness: ideal fairness (?)  maximized "utility", eg, proportional balanced, but what is utility in random traffic?  balanced fairness: an allocation that preserves processor sharing insensitivity (cf. papers by Bonald & Proutière - http://perso.rd.francetelecom.fr/bonald )http://perso.rd.francetelecom.fr/bonald  how sensitive is real bandwidth sharing?  accounting for unfairness, TCP behaviour,...  still a question of underload and overload  transparency  saturation

12. 12 Integration of streaming and elastic traffic  bufferless multiplexing for streaming traffic  ensures controlled loss, minimal delay

13. 13 Integration of streaming and elastic traffic  bufferless multiplexing for streaming traffic  ensures controlled loss, minimal delay  fair sharing of residual capacity for elastic flows  integration realized by priority queuing and TCP

14. 14 Integration of streaming and elastic traffic  bufferless multiplexing for streaming traffic  ensures controlled loss, minimal delay  fair sharing of residual capacity for elastic flows  integration realized by priority queuing and TCP  performance model of an integrated link?  exact analysis is very difficult!  sensitive performance  quasi-stationary approximation

15. 15 Integration of streaming and elastic traffic  bufferless multiplexing for streaming traffic  ensures controlled loss, minimal delay  fair sharing of residual capacity for elastic flows  integration realized by priority queuing and TCP  performance model of an integrated link?  exact analysis is very difficult!  sensitive performance  quasi-stationary approximation  a problem of local instability  when  A e > C –  s  but not with admission control... A e = offered elastic traffic  s = current streaming load

16. 16 Packet level performance  signal conservation means negligible packet loss and delay local arrival process

17. 17 Packet level performance  signal conservation means negligible packet loss and delay  delay in one queue  M/D MTU /1  as worst case limit of  D/D/1 local arrival process

18. 18 Packet level performance  signal conservation means negligible packet loss and delay  delay in one queue  M/D MTU /1  as worst case limit of  D/D/1  accumulation of jitter  the "NJ conjecture": M/D MTU /1  remains worst case in successive queues local arrival process

19. 19 Traffic control  admission control for streaming traffic  to ensure signal conservation  no a priori descriptors for variable rate flows  must use measurement based admission control (cf. Grossglauser &Tse, 2003)

20. 20 Traffic control  admission control for streaming traffic  to ensure signal conservation  no a priori descriptors for variable rate flows  must use measurement based admission control (cf. Grossglauser &Tse, 2003)  admission control for elastic traffic  to avoid congestion collapse (and ensure throughput conservation)  MBAC is easy!  using implicit control (on the fly flow identification, no signalling, no reservation) congestion (   1)admission control (   1)

21. 21 Traffic control  admission control for streaming traffic  to ensure signal conservation  no a priori descriptors for variable rate flows  must use measurement based admission control (cf. Grossglauser &Tse, 2003)  admission control for elastic traffic  to avoid congestion collapse (and ensure throughput conservation)  MBAC is easy!  using implicit control (on the fly flow identification, no signalling, no reservation)  admission control on integrated link  facilitates MBAC for all flows congestion (   1)admission control (   1)

22. 22 "Cross protect": avoiding explicit differentiation  signal conservation for flows of rate < fair rate  throughput conservation by implicit admission control forwarding decision forwarding decision priority fair queueing switch fabric priority fair queueing measurements for admission control a cross protect router

23. 23 Traffic performance in multiservice wireless networks?  flows in wireless have a spatial traffic component  in addition to traffic characteristics relevant in wireline networks  wireless resource consumption depends on user position and mobility  the resource is not just bandwidth!  users consume power and contribute interference

24. 24 Traffic performance in multiservice wireless networks?  flows in wireless have a spatial traffic component  in addition to traffic characteristics relevant in wireline networks  wireless resource consumption depends on user position and mobility  the resource is not just bandwidth!  users consume power and contribute interference  little work accounting for random elastic traffic; see however:  S. Borst, User-Level Performance of Channel-Aware Scheduling Algorithms in Wireless Data Networks, Infocom 2003  T. Bonald & A. Proutière, Wireless downlink data channels: User performance and cell dimensioning, Mobicom 2003

25. 25 Conclusion  understanding the traffic performance relation  capacity – demand – performance  characterize traffic at flow (and session) level  streaming and elastic flows  Poisson arrivals

26. 26 Conclusion  understanding the traffic performance relation  capacity – demand – performance  characterize traffic at flow (and session) level  streaming and elastic flows  Poisson arrivals  modelling elastic bandwidth sharing  processor sharing model and extensions  a question of underload and overload!

27. 27 Conclusion  understanding the traffic performance relation  capacity – demand – performance  characterize traffic at flow (and session) level  streaming and elastic flows  Poisson arrivals  modelling elastic bandwidth sharing  processor sharing model and extensions  a question of underload and overload!  integration of streaming and elastic traffic  difficult to model: sensitive and locally unstable  negligible jitter for streaming flow packets

28. 28 Conclusion  understanding the traffic performance relation  capacity – demand – performance  characterize traffic at flow (and session) level  streaming and elastic flows  Poisson arrivals  modelling elastic bandwidth sharing  processor sharing model and extensions  a question of underload and overload!  integration of streaming and elastic traffic  difficult to model: sensitive and locally unstable  negligible jitter for streaming flow packets  traffic control  measurement-based, per-flow, implicit admission control  facilitated in an integrated system  a new proposal: the cross protect router