Load Balancing of Elastic Traffic in Heterogeneous Wireless Networks

Abdulfetah Khalid, Samuli Aalto and Pasi Lassila

23.01.2013

Outline

• Introduction• Statement of the research problem• Optimal static (probabilistic) allocation• Dynamic policies• Simulation results• Conclusions

LTE Advanced: Heterogeneous Networks

Heterogeneous server model

• Assumptions:– A single macro-cell – n microcells– Poisson arrival process of

elastic flows (such as TCP downloads)

– General flow size (service requirement) distribution

– Single cell modeled as Processor Sharing(PS) queue

Research problem

• How to balance the traffic load between a macrocell and microcells?

• Target: To find an optimal load balancing policy which minimizes the mean flow level delay

• Mean flow delay implies how long it, on average, takes to transfer a file

Load balancing policies

• Apply dispatching (load balancing) policy

• Optimal Static Policy – Analytical approach– State independent policy– Used as a base line to compare the

performance of other policies

• Dynamic Policies– State dependent policy– Reacts to instantaneous changes in the

system– JSQ, Modified JSQ, LWL, Myopic– Simulations used to study performance

•

Analytical approach: optimal probabilistic allocation• Allocating the incoming

arrivals to– the micro cells with optimal

probability (pi*)

– the rest to macro cell with prob. (1- pi *)

• Objective: is to find this optimal probability values so that the mean flow delay is minimized

Analytical approach: optimal probabilistic allocation

• Given arrival rates, λi, and mean service rates, µi,

• Mean flow delay is minimized by finding optimal allocation probabilities, pi*

• For probabilistic allocation the mean flow delay, E[T], is given by

Analytical approach: optimization problem

• It can be stated as a mathematical optimization problem of the form

• Since the objective function, E[T], and constraints are convex

• Optimization problem is treated as convex optimization problem

• So, convex optimization techniques are used

Dynamic policies

• JSQ: Join the shortest queue– allocate arriving flows to

server with fewest # jobs

• MJSQ: Modified join the shortest queue– the # of flows in the server

is scaled with the service rate of server

• LWL: Least work load

– dispatch arriving flows to server with least work load

• MP: Myopic

– allocate the arriving flows to the server with least additional cost.

– additional cost =additional delay in the system experienced by all flows

Simulation: Two server case

• Assumptions– Two microcells

• Dedicated arrivals to macrocell (λ0)

• flexible arrivals to microcells (λ1 and λ2)

– Service rate of microcells (µ1 and µ2) is larger than macrocell (µ0)

– Performance is studied for • both exponentially distributed and • bounded Pareto distributed flows

– Used to model traffic that consists of heavy-tailed flow sizes

Simulation: Symmetric traffic scenario

• Two microcells– No dedicated arrivals to

the macrocell • With service rate µ0 =1

– Variable and identical arrival rates to both microcells with

• Arrival rates λ1 = λ2 = λ

• Service rates µ1 =µ2 = 2

Simulation results: Symmetric traffic scenario

Ratio of the number of flows in the system between the dynamic and base line optimal static policies

bounded Pareto distributed flowsexponentially distributed flows a=2

Asymmetric traffic scenario

• Two microcells– Dedicated arrivals to

macrocell with• With variable arrival rate λ0

= λ

• Service rate µ0 =1

– Constant and variable arrival rates macrocells

• Arrival rates λ1 =1 and λ2 = 2

• Symmetric Service rates µ1 =µ2

= 2

Simulation results: Asymmetric traffic scenario

bounded Pareto distributed flowsexponentially distributed flows

• Ratio of the number of flows in the system between the dynamic and base line optimal static policies

a=2

Simulation results: Effect of number of microcells

bounded Pareto distributed flowsexponentially distributed flows a=2

Simulation results: Effect of flow size variation

bounded Pareto distributed flows

exponentially distributed flowsbounded Pareto distributed flows

bounded Pareto distributed flows a=2

a=3

a=1.5

Conclusions

• As expected, dynamic policies perform better than the optimal static policy

• MP and MJSQ were best policies• Highest performance gain is achieved when the load of the

system is high• Implemented dynamic policies show near insensitivity

property to the flow size variation– Except the LWL policy

• Its performance gain decreases as flow size variation increases.

• Similar performance gain was achieved with MP and MJSQ – Most striking observation– MJSQ is a robust policy

Future work

• Study the system performance considering the arrival process to consist of both elastic and streaming flows– Only elastic flows was considered

• Modifying the basic model used in the thesis– Specify the service rate of the servers from radio model

• Is it possible to optimize the implemented policies?– with the help of Markov Decision Process (MDP)

• Study system performance with other metrics– Only single metric was considered, i.e mean flow level delay– Fairness, throughput,..

Thank You !

Any Comments or Questions?