Modeling load balancing in carrier aggregation mobile networks R-M. Indre Joint work with F. Bénézit, S. E. El Ayoubi, A. Simonian IDEFIX Plenary Meeting, May 23 rd 2014, Avignon
What is carrier aggregation? One of the most important features of HSPA+ and LTE Why? Improves user throughput and cell capacity How? Data traffic is aggregated on 2 or more carriers With Carrier Aggregation Joint Scheduler Carrier 1 Carrier 2 Example: Dual carrier HSDPA -> 2 carriers of 5 MHz each
Scope A dual carrier cell: Some devices are single carrier (SC) Some devices are dual carrier (DC) Downlink data channel Our claim: Joint scheduling improves performance for SC and DC users Joint Scheduler Carrier 1 Carrier 2
Roadmap 1 Single Carrier devices only 2 Dual Carrier devices only 3 Mix of Single and Dual Carrier devices 4
Single Carrier devices only Compute the flow throughput How can we model the system? Input Flow arrivals (e.g. streaming flows) Flows characterized by volume and inter-arrival times Scheduler: shares the channel between flows Scheduler Server = Carrier Carrier C Clients = Flows
Scheduling on a single carrier n flows sharing a carrier of capacity C Poisson flow arrivals of rate Mean flow size σ λ Equal sharing (i.e., each flow obtains an instantanenous rate of C/n) Processor sharing (PS) model λσ C Traffic intensity: (Mbit/s) λ n PS discipline C Load: 6 Traffic intensity Capacity
Scheduling on two carriers n flows sharing two carrier of capacities C Poisson flow arrivals of rate Mean flow size Equal sharing σ λ C1 C2 λ SC C1 C2 λ σ Traffic intensity: (Mbit/s) SC Load: ρ = C SC PS discipline λ σ + C 1 2 7
How to select the carrier? Which scheduling policy? Option 1: Bernoulli scheduling Simple Does not account for the state of the system Option 2: Join the Shortest Queue (JSQ) Accounts for the system state Poor performance if C2 >> C1 Option 3: Join the Fastest Queue (JFQ) C1 C2 > n + 1 n + 1 1 2 n 1 C1 8 C2 C1 > n + 1 n + 1 2 1 n 2 C2
System Model We model our system by means of the Markov process n( t) = ( n ( t), n ( t)) 1 2 n 1 C1 λ SC n 2 C2 For C1 = C2 =C => analytical results for JSQ policy, Flatto & McKean For C1 > C2 => numerical resolution of balance equations for JFQ
Carrier aggregation gain Carrier aggregation improves throughput and capacity C1=C2=1
JFQ vs. JSQ JFQ accounts for the capacity of the servers and the system state Improves throughput performance If C2>>C1, more significant increase of throughput Close to 1 C1=1 C2=2
Roadmap 1 Single Carrier devices only 2 Dual Carrier devices only 3 Mix of Single and Dual Carrier devices 12
Dual Carrier devices only Compute the user throughput Flows share two carriers of total capacity C1 + C2 Poisson flow arrivals of intensity Mean flow size σ λ DC Scheduler equally shares the channel between flows Joint Scheduler C1 λ DC Carrier 1 Carrier 2 C2 13
Volume balancing for DC flows How to split DC traffic among the 2 carriers to maximize throughput? Split the DC traffic s.t. completion times are equal σ σ 1 σ 2 T1 C1 C2 =T2 T1 =T Volume balancing T2 Simple implementation 14 As soon as a carrier finishes transmitting a frame, the scheduler feeds the carrier with a new frame
Volume balancing for DC flows How to split DC traffic among the 2 carriers? Split the DC traffic s.t. completion times are equal σ 1 σ 1 2 T2 C1 =T2 T1 =T σ = σ + σ T0 σ 2 C2 For any time interval in which the state n(.) of the system is constant T1 15
Volume balancing performance Flow throughput with respect to the load VB realizes ideal load balancing C1+ C2 Throughput C2 γ DC = ( C1 + C2 )(1 ρ) λ DC PS discipline C + C 1 2 1 ) 16 Load ρ 1
Roadmap 1 Single Carrier devices only 2 Dual Carrier devices only 3 Mix of Single and Dual Carrier devices 17
Mix of dual and single carrier SC and DC flows share two carriers of total capacity C1 + C2 Joint Scheduler λ DC C1 Carrier 1 Carrier 2 λ SC C2 18
System Model We model our system by a Markov process n( t) = ( n ( t), n ( t), m( t)) 1 2 λ DC n, m 1 C1 λ SC n, 2 m C2 Numerical resolution of balance equations integrating JFQ policy for SC Volume Balancing policy for DC
System stability Necessary stability condition Offered traffic < system capacity ( λ SC + λ DC ) σ < C + C 1 2 Sufficient condition for a single zone (formal proof) Stability condition for the multi-zone case No. of areas Proba to be in area j j ( λ + λ ) σ < SC, j DC, j C λ DC C1 λ SC C2
Throughput performance Joint Scheduler Carrier 1 Carrier 2 Presence of SC users slightly improves throughput of DC users Presence of DC users slightly decreases throughput of SC users Performance is quasi insensitive to the % of SC traffic C1 = C2 C =1 Only SC Only DC Mix SC DC (50%) 21
Throughput performance Joint Scheduler Carrier 1 Carrier 2 C2>C1 (C1=1, C2=1.3, Dual Band HSDPA) Results are fairly close to γ SC C (1 2 ρ ) and γ DC = ( C1 + C2 )(1 ρ) Still quasi-insensitive to the % of SC traffic C1 = 1 C2 =1.3 Only DC Only SC =1.3 Mix SC DC (50% SC) x Mix SC DC (10% SC) 22
Impact of DC traffic Joint Scheduler Carrier 1 Carrier 2 DC flow throughput slightly decreases as the % of DC traffic increases C1 = 1 C2 = 2 23
Application to HSDPA Case study for DC and DB HSDPA 2 areas with different radio conditions DC HSDPA: C11=C21=14 Mbit/s DB HSDPA: C11=14 Mbit/s C21=10 Mbit/s Rates at the edge are 10 times lower 50% of users in cell Center Traffic intensity such that mean throughput at the cell edge is 1 Mbit/s 24
Conclusion
Conclusion Carrier aggregation Scheduling policy for SC and DC users Mix of SC and DC 1 JFQ and VB realize quasi ideal load balancing 2 Performance is almost insensitive to the % of SC traffic Joint Scheduler Carrier 1 Carrier 2
Future and ongoing work Take into account gains due to channel-aware scheduling New scheduling policy for SC users: dynamically adjust to the system state Stability condition for multi-zone LTE extension: multiple carriers Joint Scheduler Carrier 1 Carrier n