Flow-Level Analysis of Load Balancing in HetNets and Dynamic TDD in LTE

Transcription

1 Flow-Level Analysis of Load Balancing in HetNets and Dynamic TDD in LTE Pasi Lassila (joint work with Samuli Aalto, Abdulfetah Khalid and Prajwal Osti) COMNET Department Aalto University, School of Electrical Engineering HEWINETS seminar, October 24, / 26

2 Outline Introduction to flow-level models Overview of 2 problem settings Load balancing in HetNets (between one macrocell and one/many microcells) Intercell coordination between 2 base stations in Dynamic TDD Conclusions HEWINETS seminar, October 24, / 26

3 What are flow-level models? Flow-level models are used to model the system from the point of view of elastic data traffic Elastic data traffic corresponds to, e.g., TCP file transfers The transmission rate varies during the lifetime of the flow Flow-level models are dynamic Flows arrive according to some stochastic process (Poisson) Flows have random service requirements (file sizes) Instantaneous service rate of a flow depends on state of the system (e.g., number of other flows) How this dependence is modeled typically contains idealizations HEWINETS seminar, October 24, / 26

4 What are flow-level models? Performance at the flow level Performance is expressed as throughput or flow delay (file transfer delay) For example, mean flow delay would describe how long file transfers on the average last Importance Users do not care about delays of individual packets, but only about the total time to transmit a file of a given size Flow-level models characterize the system at the timescale where users experience the performance HEWINETS seminar, October 24, / 26

5 M/G/1-PS queue Consider a single link with constant capacity C In a single server processor sharing (PS) queue the capacity C of the server is equally shared between the customers in system If there are n flows in system each receives service at the rate C /n Flows are served in parallel (there is no queueing!) Round robin (RR) service can be approximated by PS when service quantum in RR is small relative to flow durations (cf. time slot length in 3G/LTE systems) M/G/1-PS queueing model Arrival process is Poisson with intensity l Flow sizes X obey a general distribution with mean E[X] Service discipline is PS Well-known model with many nice analytical properties (robust) HEWINETS seminar, October 24, / 26

6 General goal Model wireless systems at flow-level Involves abstractions (simplifications) Objective To learn what queueing/scheduling theory can teach us about resource allocation and dynamic traffic HEWINETS seminar, October 24, / 26

7 Load balancing in HetNets HEWINETS seminar, October 24, / 26

8 Scenario Single macro cell (index 0) with multiple outband and separate micro cells (1,,n) 2 No interference between cells Traffic consists of elastic DL data flows 1 0 n Resources of each cell time-shared uniformly between the active flows HEWINETS seminar, October 24, / 26

9 Queueing Model n Single macro cell (index 0) with multiple micro cells (1,,n) Traffic: elastic DL data flows Poisson arrivals in each cell Generally distributed flow sizes Cells modeled as M/G/1-PS servers Assumption: Micro cells faster than the macro cell mi ³ m 0 "i HEWINETS seminar, October 24, / 26

10 Dispatching Policy n Dispatching policy decides for each arriving flow (belonging to any traffic class i) whether it should be served by the local micro cell i or the global macro cell 0 Maximal stability region: n 0 + å i = 1max{ li - mi,0} l < m 0 HEWINETS seminar, October 24, / 26

11 Optimal Dispatching Problem n Optimal dispatching policy minimizes the mean flow delay Static probabilistic policies state-independent analytical/numerical approach optimal static policy used as a baseline in performance comparisons Dynamic policies state-dependent JSQ, MJSQ, LWL, MP, and FPI performance evaluation based on simulations better performance? HEWINETS seminar, October 24, / 26

12 Traffic Scenarios Exp flow sizes (2nd experiment: bounded Pareto) n = 2 (3rd experiment: n = 2,,10) m 0 = 1 m 1 = m 2 = 2 (4th experiment: m 1 = m 2 = 4) Scenario Fixed Varied 1 (symmetric) l 0 = 0 l 1 = l 2 = l 2 (symmetric) l 1 = l 2 = 2 l 0 = l 3 (asym.) l 0 = 0, l 2 = 2 l 1 = l 4 (asym.) l 1 = 1, l 2 = 2 l 0 = l HEWINETS seminar, October 24, / 26

13 Effect of Nmbr of Microcells (Scenario 2) HEWINETS seminar, October 24, / 26

14 Conclusions All dynamic policies improve significantly the flow-level performance compared to the optimal static policy best performance gain achieved with high load gain increased when more micro cells Among the implemented dynamic policies, myopic MP appears to be systematically the best; MP may even be close to optimal in minimizing the mean flow delay; more robust MJSQ is typically able to achieve almost the same performance; FPI policies are not able to give any essential improvements over MJSQ; classical JSQ typically performs worst Performance gain of dynamic polices (except LWL) approximately insensitive with respect to the flow size distribution HEWINETS seminar, October 24, / 26

15 Publications 1. A. Khalid, P. Lassila and S. Aalto, Load Balancing of Elastic Data Traffic in Heterogeneous Wireless Networks, in Proceedings of 25th International Teletraffic Congress (ITC-25), A. Khalid, Load Balancing of Elastic Data Traffic in Heterogeneous Wireless Networks, Aalto University School of Electrical Engineering, 2013, Master's Thesis. HEWINETS seminar, October 24, / 26

16 Optimal Intercell coordination in Dynamic TDD HEWINETS seminar, October 24, / 26

17 Scenario Dynamic TDD-LTE Possible to tune fraction of time used for uplink/downlink in a cell Multiple cells inter-cell interference downlink-downlink downlink-uplink uplink-uplink Intercell coordination Possible to control transmission rate (power) of both cells at fast time scale HEWINETS seminar, October 24, / 26

18 Model (1) (Basic Setting) 2 neighboring cells 2 classes of users per cell/station uplink and downlink users Symmetric service rates (between stations): UL service rate: DL service rates: HEWINETS seminar, October 24, / 26

19 Model (2) (Interference model) Uplink power << Downlink power Low downlink interference Service rate vectors HEWINETS seminar, October 24, / 26

20 Model (3) Traffic assumptions Poisson arrivals of flows with intensities Exponential file sizes with mean Elastic traffic, PS queues with completion rates: Four M/G/1-PS queues with coupled service rates depending on modes {uu, du, ud, dd} Objective: Study how much dynamic policies improve performance (delay) over static (optimal) policies Literature: MaxWeight policy asymptotically optimal, stochastic optimality of on/off policy for 2 base stations (but no TDD) HEWINETS seminar, October 24, / 26

21 Static policies Static (probabilistic) policy: each mode selected with fixed probability Four independent M/G/1-PS queues Maximal stability region Optimal static policy Determine optimal probabilities to minimize delay Numerically or sometimes explicit result HEWINETS seminar, October 24, / 26

22 Dynamic policies We define several priority based policies HU/HD: generalizations of earlier stochastically optimal on/off policy E.g., HU is good if uplink rate is high compared with downlink HU/HD may be unstable sometimes => H3/H4 policies Other dynamic policies Policies based on MDPs (Markov Decision Processes) FPI, MDP Max Weight policy MW = Select the mode with the maximum weight Weight =Σ (service rate) x (queue length) HEWINETS seminar, October 24, / 26

23 Stochastic optimality Policy p is stochastically optimal if Two optimality results HEWINETS seminar, October 24, / 26

24 Example from numerical results Simulations in many scenarios of dynamic policies to evaluate gain over static optimal policy HEWINETS seminar, October 24, / 26

25 Conclusions Objective to study achievable gains from dynamic coordination of two base stations using dynamic TDD Priority policies Can be stochastically optimal but may also run into stability problems depending on parameters No unique, robust priority policy MW and FPI are robust for all the cases In our scenarions dynamic policies perfom better than the static optimal policy by as much as 50 60% HEWINETS seminar, October 24, / 26

26 Publications 1. P. Osti, S. Aalto and P. Lassila, Optimal dynamic TDD intercell coordination for elastic traffic, 2013, submitted. 2. P. Lassila, A. Penttinen and S. Aalto, Flow-level modeling and analysis of dynamic TDD in LTE, in Proceedings of Eighth Euro-NF Conference on Next Generation Internet (NGI 2012), P. Osti, P. Lassila and S. Aalto, Optimal intercell coordination for multiple user classes with elastic traffic, in Proceedings of Eighth Euro-NF Conference on Next Generation Internet (NGI 2012), HEWINETS seminar, October 24, / 26