1. A Dynamic Buffer Management Scheme for End-to-end QoS Enhancement of Multi-flow Services in HSDPA Suleiman Y. Yerima, Khalid Al-Begain Integrated Communications Research Centre Faculty of Advanced Technology University of Glamorgan

2. 2 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Outline  HSDPA overview  TSP queuing and BM schemes  Time-space priority BM  Dynamic Time-space priority BM  Simulation model  Results  Conclusions

3. 3 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 HSDPA Overview  High Speed Downlink Packet Access (HSDPA):  3GPP Enhancements to UMTS (3G) RAN  Higher Peak data rate: up to 14Mbps  Lower connection and response times  3-5 X capacity increase Three interacting domains:  Core Network  Radio Access Network (RAN)  User Equipment (UE) Radio Network Controller Iub interface User Equipments Core Network HSDPA CELL Node B External Network

4. 4 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 HSDPA Overview  New PHY & MAC enhancements in Node B:  New MAC-hs layer in Node B  High-Speed downlink Shared Channel  Link adaptation (AMC)  Packet Scheduling (PS)  L1 retransmissions (HARQ)  Shorter Transmission interval (2ms) Radio Network Controller Iub interface User Equipments Core Network HSDPA CELL Node B External Network

5. 5 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 motivation:  Emergence of multiple flow traffic profile per user/connection  Existing HSDPA QoS mechanisms single flow based E.g. MAC-hs Packet Scheduling (PS) Node B buffering No BM schemes in standards- open issue

6. 6 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 motivation (contd.)  The most challenging multiple flow scenario is connections with RT and NRT flows (conflicting QoS requirements) e.g. Voice + file download RT flow -> delay, jitter sensitive & loss tolerance NRT flow-> loss sensitivity & delay, jitter tolerance  Hence our proposed MAC-hs BM schemes based on Time-Space Priority (TSP) queuing model

7. 7 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 TSP queuing model Transmission to user terminal TSP threshold RTC flow NRTC flow Service process Typical priority queuing models are either loss or delay differentiated Our Time-space priority queuing model (TSP):  Single queue with hybrid differentiation  Loss differentiated  delay differentiated  RTC packets:  High priority delay  Low priority loss  NRTC packets  High priority loss  low priority delay Hence RTC => preferential transmission NRTC =>preferential buffer admission

8. 8 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008  TSP advantages  Efficient buffer utilization  Most viable for joint RTC and NRTC QoS control compared to typical priority queuing approaches  Hence we designed an efficient TSP-based BM scheme for HSDPA  Exploiting existing mechanisms in HSDPA specs.

9. 9 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 TSP-based Buffer Mgt. in HSDPA Enhanced TSP with flow control thresholds  RNC sends MAC PDUs over Iub interface  Employs Iub signalling with credit allocation algorithm to mitigate buffer overflow  Employs Discard Timer for RTC  Higher layer protocol (ARQ, TCP) performance improvement HARQ ARQ TCP CN & EXTERNAL IP TCP Source RNC Node B UE Iub interface Air interface UE 3 UE 2 UE 1 RT flow UE 1 UE 1 MAC-hs buffer Capacity Request {Priority, User Buffer Size (UBS)} RNC UE 1 NRT flow R H L N Iub flow control Packet Scheduling Capacity Allocation {Priority, Credits} HS-DSCH Frame {Priority, UBS, PDU size, #PDUs} RNCNode B

10. 10 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 TSP-based BM in HSDPA Credit based FC algorithm: C Total = C NRT + C RT C RT = (λ RT / PDU_size) ∙ TTI C NRT = min { C NRTmax, RNC NRT } C NRTmax = (λ’ NRT /PDU_size) ∙ TTI, N’ T < L β ∙ (λ’ NRT /PDU_size) ∙ TTI, L ≤ N’ T ≤ H, 0 < β < 1 0, N’ T > H Where λ' NRT = α ∙ λ' NRT-1 + (1- α) ∙ λ NRT is EWMA of Scheduler NRT data rate and N’ T = θ ∙ N T-1 + θ ∙ N T is an EWMA of total queue size

11. 11 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008  Enhanced TSP improves e2e NRTC throughput without compromising RT QoS  Problem: TSP has static delay prioritization  Potential NRTC bandwidth starvation  Non optimal ARQ and TCP performance  A possible solution:  Exploit possible RTC delay tolerance Hence we extend the TSP BM with Dynamic Time priority switching

12. 12 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 D-TSP RNCUE 1 MAC-hs buffer NRT R H L N Iub flow control Packet Scheduling Priority switching RT UE 1 Incorporates delay Priority switching to TSP DTSP priority switching algorithm: IF RT packets 0 Time Priority = NRT flow Generate Transport Block from NRT PDUs ELSE Time Priority = RT flow Generate Transport Block from RT PDUs MAX_delay = Max e2e delay – other queuing and propagation delays k= Delay budget / RTC inter-arrival time Delay budget ≤ MAX_delay Discard timer (DT) setting = MAX_delay

13. 13 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 HSPDA modelling  Detailed custom modelling with OPNET :  Multi-flow connection: VoIP source, NRT source with TCP  Fixed external and Core Network delay assumed  RNC: Packet segmentation, RLC modes  Node B: AMC Link adaptation, HARQ, MAC-hs buffers, Packet scheduler  Receiver: SINR, HARQ, RLC modes, re-assembly queues, TCP User Equipments HSDPA CELL Node B RNC Core Network voice + data connection External Network 70ms 20ms2ms

14. 14 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 simulation set up Performance metrics in test UE: End-to-end NRTC throughput Voice PDU discard ratio (Discard timer) % HSDPA channel utilization HSDPA Simulation Parameters HS-DSCH TTI2ms Path loss Model148 + 40 log (R) dB Shadow fadingLog-normal: σ = 8 dB Number of HSDSCH codes 5 CQI delay3 TTIs (6ms) HARQ processes4 HARQ feedback delay5ms Test UE position from Node B 0.2 km Packet SchedulingRound Robin RLC PDU size320 bits RNC-Node B delay20ms External + CN delays70ms TCP configMSS= 536 bytes, RWIND = 64 Flow control settingsβ = 0.5, α = 0.7, θ = 0.7 20ms 2ms 70ms Node B RNC C oncurrent VoIP + FTP for 180s MAX_delay = 250 –( 70 – 20) = 160ms BM config: R = 10, L =100, H= 150, N= 200 PDUs DTSP params: k = 2, 4, 6, 8 channel load: 1, 5, 10, 20, 30, 50 users

15. 15 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (1 user) Throughput at UE1 DTSP and TSP show similar performance Increased DB settings has marginal effect on throughput very low HS-DSCH load hence no DTSP gain

16. 16 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (5 users) Throughput at UE1 More load on HSDPA channel Increased DB settings show noticeable improvement DTSP performs better than TSP

17. 17 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (10 users) Throughput at UE1 Higher load on HSDPA channel Increased DB settings show noticeable improvement DTSP performs better than TSP

18. 18 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (20 users) Throughput at UE1 Higher load on HSDPA channel Increased DB settings show noticeable improvement DTSP performs better than TSP

19. 19 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (30 users) Throughput at UE1 Higher load on HSDPA channel Increased DB settings show noticeable improvement DTSP performs better than TSP

20. 20 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: UE1 Data download throughput (50 users) Throughput at UE1 Higher load on HSDPA channel Increased DB settings show noticeable improvement DTSP performs better than TSP performance peaks at k = 6

21. 21 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: Voice pdu discard ratio vs DB settings Voice pdu discard loss assuming max DR= 2% In 1, 5, and 10 user scenarios VoIP QoS satisfied Optimum k for 20 users = 6 optimum k for 30 users = 4 Optimum k for 50 users = 2

22. 22 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Results: HSDPA channel utilization vs DB settings UE1 HSDPA channel utilization Utilization constant in DTSP regardless of DB setting in 1 user scenario channel utilization improves with higher load due to pdu bundling DTSP has better channel utilization than TSP except at very low load ( e.g. 1 user scenario)

23. 23 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Conclusions and Summary  Conclusions:  DTSP achieves e-2-e throughput improvement for the multi-flow NRTC traffic  Better channel utilization is achieved with DTS P over TSP  Acceptable VoIP performance within QoS constraints Further work  Investigate with other Packet Scheduling  Investigate other possible multi-flow scenarios

24. 24 NGMAST ’08 International Conference, University of Glamorgan, Cardiff, Wales, Sept19th 2008 Thank You!!!