阅读目的:学习eMBB和URLLC共存下的资源分配/调度等问题
point:
和m/g/1模型关系?—考虑队列时延
risk概念?—考虑了对URLLC delay的细粒度描述,刻画了tailed distribution.
平均URLLC包时延的闭合表达
Introduction
- related work
回顾了eMBB和URLLC共存下的调度方案[3-9],指出上述文献没有考虑URLLC packet的队列时延,且说明文献[10]中指出只要在高SINR和较低的URLLC包到达率下,该时延可以忽略。但是文献[11]中指出,由于bursty URLLC traffic,且可能有资源短缺,因此会产生blocking,从而导致严重的包的queueing delay。这一情况的存在使得上述文献中的算法不能适用于特定信道低SINR或URLLC packet arrival rate不是那么小的情况。
文献[12][13]中指出,由于潜在的queueing delay, 使用M/G/1模型最小化复用两种业务时的带宽。文献[12]提出了基于delay cycle(idle period and bysy period)的频谱计算方法。[13]中推导了基于URLLC包时延闭合表达式的variable packet length。然而,[12]中仅仅分析eMBB的包时延,[13]中仅针对基于URLLC包平均时延满足URLLC的时延需求,不是其最坏情况(violation case).为了刻画URLLC业务的平均时延,[14]中考虑了M/G/
∞
\infty
∞模型及sotchastic optimization problem,最大化eMBB throughput。
- Motivation
最后,作者总结已有文献中,缺乏对时延的的细粒度metrics,从而影响了调度机制的适用性。基于此以及文献[4]中的risk-resistant方法,本文提出了下行的资源分配的risk-resistant机制,构建了M/G/1模型下violation case的URLLC平均包时延。通过联合优化eMBB吞吐量和平均URLLC包时延,旨在降低RILLC传输的风险(控制额外的URLLC delay抖动),在最大化eMBB吞吐量的同时满足URLLC业务的时延需求。
- Contribution
1)risk-resistant model (引入CVaR:conditional value at risk)作为risk metric.
- a joint optimization problem of both eMBB and URLLC schedulers,non-convex problem si transormed to two convex problems via Taylor’s theorem.
System Model
standard 5G NR terms, OFDMA downlink cellular networks
single cell system, covers a radius of V meters
U
=
{
1
,
2
,
.
.
.
U
}
\mathcal{U}=\{1,2,...U\}
U={1,2,...U}:URLLC users
M
=
{
1
,
2
,
.
.
.
M
}
\mathcal{M}=\{1,2,...M\}
M={1,2,...M}:eMBB users
B
=
{
1
,
2
,
.
.
.
B
}
\mathcal{B}=\{1,2,...B\}
B={1,2,...B}:a set of RBs in an eMBB TTI duration, the minimum unit of resorces in the time-frequency domain
Z
Z
Z: URLLC packet size
λ
(
a
r
r
i
v
a
l
s
/
s
l
o
t
/
u
s
e
r
)
\lambda(arrivals/slot/user)
λ(arrivals/slot/user): URLLC arrival following Poisson process
Λ
=
Z
⋅
λ
⋅
U
⋅
τ
\Lambda=Z\cdot\lambda\cdot U\cdot \tau
Λ=Z⋅λ⋅U⋅τ: URLLC traffic load
assignment of RBs to eMBB users and to randm URLLC traffic
分配给eMBB用户的RB是基于PF
assumption: full buffer traffic model
f
R
B
f_{RB}
fRB:bandwidth of each RB
Data rate model for eMBB
eMBB用户的数据速率损失和URLLC puncturing的数量呈现线性关系。
ω
m
(
t
)
\omega_m(t)
ωm(t):weight that the m-th eMBB user is punctured at the t-th time slot
ω
(
t
)
=
(
ω
1
(
t
)
,
ω
2
(
t
)
,
.
.
.
ω
M
(
t
)
)
\bold\omega(t)=(\omega_1(t),\omega_2(t),...\omega_M(t))
ω(t)=(ω1(t),ω2(t),...ωM(t)):vector of the punctruing weight
高的
ω
m
(
t
)
\omega_m(t)
ωm(t)表示第m个用户更有可能被占用资源
则第m个用户在第t个时隙的下行可达速率近似为
dm:eMBB和BS距离
S
m
a
x
(
t
)
S_{max}(t)
Smax(t):maximum data flow density that system can serve for all URLLC users at the t-th time slot
Data rate model for URLLC
Problem Formulation
目标:find the eMBB scheduling strategy I ( t ) = [ I b , m ( t ) ] \bold I(t)=[I_{b,m}(t)] I(t)=[Ib,m(t)] and punctured probability vector ω ( t ) \bold\omega(t) ω(t), 最大化eMBB throughput,同时最小化risk to URLLC transmission
Delay model
只考虑 transmission delay(service delay)
X
u
X_u
Xu 和 queueing delay
Q
u
Q_u
Qu
一个包的时延定义为包到达BS到离开BS的时间间隔,=queueing+service delay
average delay of a packet based on the [18]
ρ
=
λ
E
[
X
u
]
\rho= \lambda\mathbb{E}[X_u]
ρ=λE[Xu]:service strength fo queueing system
V
a
r
(
X
u
)
=
E
[
X
u
2
]
−
E
2
[
X
u
]
Var(X_u)=\mathbb{E}[X_u^2]-\mathbb{E}^2[X_u]
Var(Xu)=E[Xu2]−E2[Xu]
(3)(4)
transmission delay:
X
u
=
Z
/
R
u
X_u=Z/R_u
Xu=Z/Ru,Z is packet size, Ru is data rate
用一阶Taylor展开近似service delay的均值和二阶均值。
(5)
(6)
为了确保上述两个值都是大于零的,Ro应该满足,且近似准确度
也依赖于R0的选择。(5)(6)代入(4)
Risk-Resistant Model
CVaR函数作为risk metric,反应平均时延超过特定阈值的机会,或者说刻画了超过Value-at-risk(VaR)d的average fraction of delay
关键参数:
β
\beta
β-VaR:定义为一个最小阈值,即平均时延不超过阈值的一个特定概率
β
∈
(
0
,
1
)
\beta\in(0,1)
β∈(0,1),[19]给出了这一参数
α
\alpha
α: 一个表示阈值的变量
D
=
∑
u
∈
U
D
u
/
U
D=\sum_{u\in\mathcal{U}}D_u/U
D=∑u∈UDu/U:estimated average delay of incoming URLLC users
ϕ
β
\phi_{\beta}
ϕβ:expectation of D under the condition
D
≥
α
β
D\ge\alpha_{\beta}
D≥αβ,这个参数就可以看做是一个对URLLC delay的细粒度描述,因为其聚焦于average packet delay in violation cases.
因此,最小化risk of URLLC,就是最小化
ϕ
β
\phi_{\beta}
ϕβ,且
括号内的表达式称为
β
\beta
β-CVaR-lime function,记为
ϕ
β
(
α
)
\phi_{\beta}(\alpha)
ϕβ(α)
- network utility function:captures both the sum-log of eMBB data rate and the risk of URLLC:
γ ∈ [ 0 , 1 ] \gamma\in[0,1] γ∈[0,1]:weight parameter
(11)求和中的第一项表示eMBB users的sum-log data rate,确保分配资源给eMBB时的比例公平,第二项表示risk function(10)
式中的 1 − γ 1-\gamma 1−γ和 γ \gamma γ用于trade off weightings given the eMBB users and URLLC users,例如一个大的 γ \gamma γ表示URLLC服务对网络效用更关键 - joint optimization problem
上述优化的目标是旨在一个给定的risk level下找到最优的资源分配策略 I ( t ) ∗ \bold{I}(t)^* I(t)∗及最优的punctruing权值vector ω ( t ) ∗ \bold{\omega}(t)^* ω(t)∗
assumption:BS做等功率分配
Resource Scheduling Strategy
(12)是mixed-integer nonlinear programming问题,decompose the problem into two subproblems
Resource Allocation to eMBB users
fix the pncturing weight vector
ω
(
t
)
\bold\omega(t)
ω(t). Remove the
ϕ
β
~
(
α
)
\widetilde{\phi_{\beta}}(\alpha)
ϕβ
(α) cause it just depends on the URLLC part which is the
ω
(
t
)
\bold\omega(t)
ω(t).
Reformulated the original optimization problem as
Puncturing Placement Problem of URLLC
given eMBB scheduling strategy
I
(
t
)
\bold{I}(t)
I(t)
基于上述文献的理论,移出non-convex term
(
x
)
+
(x)^+
(x)+,引入一个辅助变量
k
\mathcal{k}
k,及两个新的约束条件。基于期望的线性特性,总时延的期望可以表示为
式(14)可以改写为
调度决策是在t-1时隙得到,直接应用于现时隙
实际实现过程中,上述两个子问题可用迭代方式交替,见Algorithm1
定义一个任意小的正数
ϵ
\epsilon
ϵ,形成结束迭代的条件。根据文献[5,22]的收敛分析,算法converges sub-linearly in the order of
1
/
ϵ
1/{\epsilon}
1/ϵ.
Complexity Analysis of the Proposed Procedure
barrier method with gradient descent is used to solve the convex problem.
Performance Evaluation
Simulation setup
- performance:eMBB fairness, throughput, packet delay of URLLC at the outage probability of 1 0 − 5 10^{-5} 10−5
- users are distrubuted randomly within the coverage
- Rayleigh fading channel
5G-NR frame structure is implemented over one million packet communications
Performance Analysis
benchmark:
1)conventional PF scheduling
2)low-complexity packet scheduling(LC-PS) [7]
3)modified largest weighted delay first(M-LWDF) with prioritized URLLC traffic SCHEDULING【24】