Deadline-Constrained Traffic Scheduling Algorithm

for Collocated Wireless Networks

Tashrif Billah
Columbia University
Department of Electrical Engineering
ELEN 6909: Network Algorithm and Dynamics Course Project
Spring 2017
 Collocated Network
 A base station
 Some pairs of users interacting over the base station: Links
 All the pairs interfere with one another
 Only one pair can talk to each other at a time
Traffic Nature
• Non real time traffic: Emails
• No deadline
• Exists until scheduled or dropped from buffer

• Real time traffic: Live video streaming

• Expires after a certain time

• An algorithm is required to ensure certain fraction of the arrived packet transmission

 Literature Review

• QOS Satisfaction: achieving transmission of a certain fraction of the arriving

packets to any link, weighted or unweighted links, fair policy [1][2]
[1] A Fast-CSMA Algorithm for Deadline-Constrained Scheduling over Wireless Fading Channels, Bin Li et al, 2013
[2] On the Performance of Largest-Deficit-First for Scheduling Real-Time Traffic in Wireless Networks, Lei Ying, 2016

• Cost optimality: minimizing the cost incurred due to drop of unscheduled packets
owing to deadline expiry, weighted links [3]

Traffic and Network Model, Optimization Problem, Algorithm, Proof Sketch, Limitation
 Fast-CSMA, Collocated Networks, Bin Li et al, 2013 [1]
Traffic and Network Model
• Packet arrives to links with iid probability over links and over time: 𝝀𝒍
• There are frames of length T, multiple packets can arrive at a time
• All packets arrive at the beginning of the frame and are due by the end of the
frame
• Un-transmitted packets are dropped at the end of each frame: 𝝆𝒍 is the
maximum allowable packet drop rate due to deadline expiry
• Channel has a transmission capacity 𝑪𝒍 that is subject to fading and varies with
time

Optimization Problem
• Ensuring quality of service i.e. (𝟏 − 𝝆𝒍 ) fraction of the arrived packet transfer
from individual link
 Fast-CSMA, Collocated Networks, Bin Li et al, 2013 [1]

• Evolution of virtual queue length 𝑿𝒍 𝒌 + 𝟏 𝑻 = 𝑿𝒍 𝒌𝑻 + 𝑫𝒍 𝒌𝑻 − 𝑰𝒍 𝒌𝑻 + 𝑼𝒍 [𝒌𝑻]

• 𝑋𝑙 [𝑘𝑇] is the virtual queue length at the beginning of frame k

• D𝑙 𝑘𝑇 − 𝐼𝑙 𝑘𝑇
• Upon arrival of a packet, it increases by 1 with probability (𝟏 − 𝝆𝒍 ) and 0 otherwise
• It decreases by 1 whenever a packet is schedule provided that channel condition is good

• 𝑈𝑙 [𝑘𝑇] has the same meaning as max (0,•) i.e. it prevents virtual queue length from being
negative

Virtual queue length (or deficit) is the measure of how far a link is from achieving
required delivery rate
Fast-CSMA, 2013 [1] • R[t] should change slowly so that CSMA Markov chain stay
close to its stationary distribution
log •
• CSMA (Carrier Sensing Multiple Access) • If R[t] is chosen as log log • function [5] of the queue length
Sense carrier in the interference neighborhood (real or virtual), the Markov chain behaves in an adiabatic
manner i.e. its parameters does not change significantly
If no one is transmitting
over time
Transmit
Else Limitation-1: Large mixing time
Back off Limitation-2: Influence of rapidly varying channel condition
Wait a duration equal to an exponentially
distributed random variable R[t] • Idea is to exploit fast convergence property of hitting time
Start over i.e the duration after which the Markov chain stays in a non
End if zero state
Transmit for a duration exponentially
distributed with mean 1

• Fast-CSMA
Same as above except, the link that
captures the channel transmits data
until the end of the time slot
Fast-CSMA, Collocated Networks, Bin Li et al, 2013 [1]

𝑓 𝑥 𝑓 𝑥+𝑎
• 𝑔 𝑥 = {𝑒 | lim = 1}; f(x) = x, 𝑒 𝑥
𝑥→∞ 𝑓 𝑥

Number of remaining packets at time slot t of frame k

Channel capacity

Proof idea for throughput optimality

𝑾𝒍 𝒕 = 𝐥𝐨𝐠 𝑹𝒍 𝒕 𝑎𝑛𝑑 𝑾∗𝒍 = 𝐦𝐚𝐱 𝑾𝒍 [𝒕] , 𝐹𝐶𝑆𝑀𝐴 algorithm picks a link 𝒍 satisfying
𝒍
Fast-CSMA, Collocated Networks, Bin Li et al, 2013 [1]

• Proof idea for stability

Lyapunov function 𝑉 𝑋 = σ𝐿𝑙=1 ℎ(𝑋𝑙 ) 𝑤ℎ𝑒𝑟𝑒 ℎ′ 𝑥 = 𝑓 𝑥 , ∃ 𝛾 > 0 𝑎𝑛𝑑 𝐻 ≥ 0 such that

V. Q. length at the beginning of frame K Change of V.Q. length in frame k

Fast-CSMA, Collocated Networks, Bin Li et al, 2013 [1]

• Limitation
o Frame based traffic
o The link is always full or always empty over a frame
o Practical traffic has probabilistic arrival all over the frame
o Practical traffic has different deadlines, not the same one for all links
Efficiency Ratio of Largest Deficit First, Lei Ying, 2016 [2]
Traffic and Network Model
• Ad hoc network
• Packet arrives to links with iid probability 𝒂𝒍 over links and over time
• There is no frame, multiple packets can arrive at a time and they arrive at any time
• Each packet is associated with a deadline according to a probability distribution over
[𝝉𝒎𝒊𝒏 , 𝝉𝒎𝒂𝒙 ]
• Un-transmitted packets are dropped at the end of each frame: 𝝆𝒍 is the required
fraction of packet transmission
• Perfect channel condition is assumed: transmission takes place with 100% confidence
when scheduled

Optimization Problem
• Ensuring quality of service i.e. 𝝆𝒍 fraction of the arrived packet transfer from
individual link
Efficiency Ratio of Largest Deficit First, Lei Ying, 2016 [2]
𝝁
• Deficit queue: 𝑫𝒍 (𝒕), deficit of link 𝒍 at time 𝒕 under scheduling policy 𝝁
Largest Deficit First Algorithm
• Schedule a set of links in descending order of their deficits
• As long as they do not interfere with each other
• There is at least one packet in the input of the selected links
• Arbitrarily break all ties
Some Definition
• 𝑴 ∈ 𝟎, 𝟏 𝑳 𝒙 𝑹 is the maximal matching
• 𝑴𝑱 (𝑭) ∈ 𝟎, 𝒂𝒎 𝑳 𝒙 𝑻′
is the service matrix where there are L rows and 𝑻′ ≤ 𝑻
columns
• 𝒂𝒎 is the maximum number of packet arriving to any link over frame F and T is
the frame length for analysis purpose
Efficiency Ratio of Largest Deficit First, Lei Ying, 2016 [2]
• ᴧ𝑳𝑫𝑭 is the stability region of LDF, ∧ is the maximal stability region
of the system i.e. set of arrival rates supportable by any scheduling
policy

• Connecting ᴧ𝑳𝑫𝑭 to ∧ is difficult, that is why an intermediate region

∧ (𝐹) is needed

𝜎 ∗ 𝑖𝑛𝑡 ∧ ⊆ lim inf ∧ 𝐹 ⊆ ∧𝐿𝐷𝐹

𝐹→∞
Efficiency Ratio of Largest Deficit First, Lei Ying, 2016 [2]

• Real time local-pooling factor 𝜎 ∗ (𝐹) = sup 𝜎 ∈ 0,1 ∶ ∀ 𝜙1 , 𝜙2 ∈ 𝐶𝐻 𝑀𝐽 (𝐹) 𝜎𝜙1 ⊁ 𝜙2 }

• 𝜎 ∗ = lim inf 𝜎 ∗ (𝐹)
𝐹→∞

• LDF achieve at least 𝝈∗ 𝒊𝒏𝒕(ᴧ)

Efficiency Ratio of Largest Deficit First, Lei Ying, 2016 [2]
• As long as the deficit arrival rate 𝜆𝑙 = 𝑎ഥ𝑙 𝜌𝑙 ∈ 𝜎 ∗ 𝑖𝑛𝑡 ᴧ , the drift of the Lyapunov function
𝑉 𝑡 = max 𝐷𝑙 (𝑡) becomes negative [4]
𝑙∈L

Proof Sketch
𝒅 𝒅
• 𝑲 𝒕 = {𝒍 ∈ 𝑳: 𝑫𝒍 𝒕 = 𝒎𝒂𝒙 𝑫𝒊 (𝒕)} ; LDF serves these links first
𝒅𝒕 𝒊∈𝑳 𝒅𝒕
d 𝑑 d
• Fluid limits 𝐴𝑙 𝑡 𝜌𝑙 = 𝜆𝑙 , 𝜇𝑙 = 𝜋𝑙 , 𝐷 (𝑡) = 𝜆𝑙 − 𝜋𝑙 (𝑡)
𝑑𝑡 𝑑𝑡 𝑑𝑡 𝑙

𝑑 𝑑
• ∃ 𝑙 ∈ 𝐾 and 𝜋𝑙 𝑡 ∈ 𝐶𝐻(𝑀𝐽 ) for which 𝜆𝑙 − 𝜋𝑙 (𝑡) ≤ −𝜖 i.e. drift 𝑉 𝑡 ≤ 𝐷 ≤ −𝜖
𝑑𝑡 𝑑𝑡 𝑙
• As long as 𝜆 ∈ 𝜎 ∗ 𝑖𝑛𝑡 ∧ , the largest deficit decreases i.e. all other deficits decreases in
[𝑡, 𝑡 + 𝛿]

𝜇
lim lim sup Pr(σ𝑙∈𝐿 𝐷𝑙 𝑡 ≥ 𝐶) = 0, the total deficit remains bounded
𝐶→∞ 𝑡→∞
Interference Degree and Efficiency Ratio
• Interference degree 𝜷: max # of links interfering with a single link that can be scheduled
simultaneously
•𝜷=𝟏 𝜷=𝟐

𝟏
• For LDF, 𝝈∗ ≥
𝜷+𝟏
• Proof intuition and outline: divide any arbitrary traffic
to two groups

• Limitation: It does not have any relation with deadline, random tie breaking may not be
the right decision
 Our Model
• A set of 𝑳 collocated links, only one link i.e. pair of users can active at a time
• A complete graph of vertex degree L i.e. a conflict graph of interference degree 1
• Each link has a packet arrival rate {𝝀𝒍 : 𝝀𝒍 ∈ 𝟎, 𝟏 | σ𝑳 𝝀𝒍 < 𝟏}
• Time is slotted, one packet arrives every time slot w.p. 𝝀𝒍 or not w.p. 𝟏 − 𝝀𝒍
• Links have deadlines (𝝉𝟏 , 𝝉𝟐 , … , 𝝉𝑳 ) 𝝉 ∈ ℕ, ∀ 𝒊, 𝒋 ∈ 𝑳, 𝝉𝒊 ≠ 𝝉𝒋 }, all packets arriving to a link
have the same deadline
• Perfect channel condition is assumed: transmission takes place with 100% confidence when
scheduled

Problem: Satisfying QOS

• Links have a required quality of service 𝜌1 , 𝜌2 , … , 𝜌𝐿 𝜌 ∈ ℝ, 𝜌 ∈ (0,1)}
• QOS fraction of arriving packets needs to be transmitted before deadline expiry
• Goal is to meet QOS of every links
 Formulating Optimization Problem for QOS Satisfaction
min 0
Subject to
σ𝑇 (𝑎 (𝑡)−𝑠𝑙 𝑡 )
lim 𝑡=1σ𝑇 𝑙 ≤ 1 − 𝜌𝑙 ∀ 𝑙 ...... Delivery ratio is satisfied
𝑇→∞ 𝑡=1 𝑎𝑙 (𝑡)

𝑇
1
⇒ lim 1 − ෍ 𝜇𝑙 𝑋𝑡 ≤ 1 − 𝑝𝑙 ∀𝑙
𝑇→∞ 𝜆𝑙 𝑇
𝑡=1
σ𝐿𝑙=1 𝑠𝑙 𝑡 = 1 ∀ 𝑡 ...... Only one link is scheduled at a time

σ𝑡𝑡−𝜏𝑙 +1(𝑎𝑙 𝑡 − 𝑠𝑙 (𝑡)) ≥ 0 ∀ 𝑙 ...... There exists packet in the input

An algorithm will decide which link to schedule based on state 𝑋𝑡 = 𝑓(𝑄 𝑡 , 𝐷(𝑡), 𝜏𝑙 , 𝜆𝑙 )
Evolution of queue length 𝑄 𝑡 = (𝑄 𝑡 − 1 + 𝑎 𝑡 − 𝑠(𝑡))+
Evolution of deficit length 𝐷 𝑡 = (𝐷 𝑡 − 1 + (1 − 𝜌)𝑎 𝑡 − 𝑠(𝑡))+
 Proposed Algorithm
• Earliest deadline first with largest deficit tie-breaking (EDF-LDF):
• State 𝑋𝑡 = 𝑓(𝐷(𝑡), 𝑡𝑎 + 𝜏𝑙 − 𝑡)

for t= 1...T
If (there exist links with packets due)
Find out the links with earliest deadline due (EDD) packet
If (there is more than one EDD links)
Find out the link 𝑙 with largest deficit
If still tied
Randomly pick up one link 𝑙 from the tied ones
end if
𝑠 𝑙, 𝑡 = 1 // Link 𝑙 is scheduled
𝐷 𝑙, 𝑡 = (𝐷 𝑙, 𝑡 − 1)+ // Updating deficit
End if
End if
End for
Proposed algorithm: EDF with LDF tie-breaking
 Proposed Algorithm
*** EDF+LDF ***
Achieves better QOS than the rest
Total deficit remains bounded
Existing Well-Known Algorithms
• LDF+EDF
Find out the largest deficit links and then break ties using EDD packets

• LDF
Find out the largest deficit links and then break ties arbitrarily
 Suffers from poor QOS
• EDF  Total deficit becomes unbounded
Always schedule the EDD packets
Easy to implement
• Randomized
Randomly choose one link among the ones that have packets
Achieves nearly same QOS for all the links
QOS and Deficit Comparison of All Algorithms

EDF+LDF satisfies the QOS and the total deficit is the lowest
 ∗ 1 ∗ 1
𝜎 = ,𝛽 = 1, Achievable capacity region, 𝜎 ≥
𝛽+1 2

72-77% 77-82%
Existing Algorithm (LDF) Proposed Algorithm (EDF+LDF)
Work In Progress over the Summer

QOS satisfying problem

 [2] claims and proves a lower bound of the achievable QOS region, we
are trying to relate that bound to deadlines
 Trying to develop a rigorous proof that EDF with LDF tie braking
performs better over the other algorithms
 Thank You!

This work is being done under the supervision

of
Prof. Javad Ghaderi
Columbia University

Tashrif Billah
MS/PhD Track Student
Columbia University
Department of Electrical Engineering

Disclaimer: A number of notational simplicities have been introduced like dropping subscripts, and also rigorous mathematical proofs have
been modified to suit larger audience. Please refer to the list of references to get accurate details.
 List of References

 1. B. Li and A. Eryilmaz, Optimal Distributed Scheduling under Time-Varying Conditions: A Fast-CSMA

Algorithm with Applications, IEEE Transactions on Wireless Communications, Vol. 12, No. 7, 2013
 2. X. Kang, W. Wang, J. J. Jaramillo, and L. Ying, On the Performance of Largest-Deficit-First for
Scheduling Real-Time Traffic in Wireless Networks, IEEE/ACM Transactions on Networking, Vol 24, No. 1,
2016
 3. R. L. Givan, E. K. P. Chong, and H. S. Chang, Scheduling Multiclass Packet Streams to Minimize
Weighted Loss, Queueing Systems, 41, 2002
 4. C. Joo, X. Lin, and N. B. Shroff, Performance Limits of Greedy Maximal Matching in Multi-hop Wireless
Networks, In proceedings of 46th IEEE Conference on Decision and Control, New Orleans, LA, 2007
 5. J. Ghaderi and R. Srikant, On the Design of Efficient CSMA Algorithms for Wireless Networks, CDC,
2010