Virtualization-enabled dynamic functional split and resource mapping in optical data center networks [Invited]

Bo Tian; Shanting Hu; Qi Zhang; Xiaofei Huang; Lei Zhu; Huan Chang; Xiaolong Pan; Xiangjun Xin

doi:10.3788/COL202523.050001

Chinese Optics Letters, Volume. 23, Issue 5, 050001(2025)

Virtualization-enabled dynamic functional split and resource mapping in optical data center networks [Invited] On the Cover

Bo Tian^1、*, Shanting Hu^1、**, Qi Zhang², Xiaofei Huang³, Lei Zhu², Huan Chang¹, Xiaolong Pan¹, and Xiangjun Xin¹

Author Affiliations

¹School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China

²School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China

³Beijing Institute of Electronic System Engineering, Beijing 100584, China

show less

Figures & Tables(11)

Fig. 1. Heterogeneous optical switching-based DCN architecture.

Download full size

View in Article

Fig. 2. Internal structure of an optical ToR.

Download full size

View in Article

Fig. 3. Fine-grained functional split options.

Download full size

View in Article

Fig. 4. Resource mapping framework in HOS-DCNs.

Download full size

View in Article

Fig. 5. Comparison of throughput with increasing computational capacities using DRL-BFM, RBM, LCBM, and LBFO.

Download full size

View in Article

Fig. 6. Comparison of throughput with increasing bandwidth capacities using DRL-BFM, RBM, LCBM, and LBFO.

Download full size

View in Article

Fig. 7. Comparison of throughput with increasing ratios of high-latency services using DRL-BFM, RBM, LCBM, and LBFO.

Download full size

View in Article

Fig. 8. Load balancing performance using DRL-BFM, RBM, LCBM, and LBFO.

Download full size

View in Article

Table 1. Hyperparameters in the Learning Process

View table

View in Article

Table 1. Hyperparameters in the Learning Process

Parameter	Value	Parameter	Value
Learning rate	10⁻⁴/10⁻⁵	Discount factor	0.9
Hidden layers	3	Neurons	256/layer
Minibatch size	128	Number of epochs	10
Optimizer	Adam	Activation function	ReLU(·)
Output function	Softmax	Coefficient of L_t^CLIP(θ)	2
Coefficient of L_t^VF(ϕ)	0.05	Coefficient of H (π_θ, s_t)	0.01

Table 1. PPO-based Training Process

View table

View in Article

Table 1. PPO-based Training Process

Input: Request

R

, latency threshold

t_{r}^{t h}, r \in R

g_{r, k}

c_{r, k}, r \in R, k \in K

Initialize: Critic network

V_{ϕ}^{n} (s (t))

, actor network

π_{θ}^{n}

\forall n \in N

1: for iteration = 1, 2, … do

2: Initialize experience replay buffer

B = Ø

3: while the computational capacities of all racks are non-empty

4: for

L_{r} \in R

f_{r, 1}

is processed in rack

Δ_{r, 1}

6: update computational capacity of

Δ_{r, 1}

7: for

k = 2, 3, \dots, K

8: identify the current agent

n \in N

9: obtain network state

S_{r, k}

10: update computational capacity in

Δ_{r, k}

11: if

Δ_{f, k - 1} \neq Δ_{f, k}

12: update fiber bandwidth capacity between

Δ_{f, k - 1}

and

Δ_{f, k}

13: end if

14: observe outcome reward

R_{r, k}

15: collect

D_{r, k} = ⟨ S_{r, k}, a_{r, k}, R_{r, k}, S_{r, k}^{'} ⟩

16: store into the replay buffer

B = B \cup D_{r, k}

17: end for

18: end for

19: end while

20: for

x = 1, 2, \dots, B_{n}

21: randomly sample a minibatch of

b

from

B

22: for

t = 1, 2, \dots, | b |

23: calculate the advantage function

{\hat{A}}_{θ} (t)

24: calculate

π_{θ} (a_{t} | s_{t})

under old policy

θ

25: for

e p o c h = 1, 2, \dots, P

26: calculate

π_{θ^{'}} (a_{t} | s_{t})

under new policy

θ^{'}

27: calculate the ratio

r_{t} (θ)

28: calculate the strategy gradient

L_{t}^{C L I P} (θ)

29: update

θ

Δ L_{t}^{C L I P} (θ)

30: calculate the value function estimator

L_{t}^{V F} (ϕ)

31: update

ϕ

Δ L_{t}^{V F} (ϕ)

32: end for

33: end for

34: end for

35:

π_{θ} \leftarrow π_{θ}^{'}

36: end for

Table 2. DRL-BFM

View table

View in Article

Table 2. DRL-BFM

Input: Request set

R

, trained DRL parameters

π_{θ}^{i}, i \in N

1: for

r \in R

2: for

k = 2, 3, \dots, 7

3: identify the current cluster

i

4: calculate the network state

S_{r, k}

5: run policy

π_{θ}^{i}

to select an action

a_{r, k}

6: if

a_{r, k}

is valid

7: map

f_{r, k}

onto VM

# k

and process it in

Δ_{r, k}

8: update computational capacity in

Δ_{r, k}

9: if

Δ_{f, k - 1} \neq Δ_{f, k}

then

10: if

Δ_{r, k - 1}

and

Δ_{r, k}

are in the same cluster then

11: map the virtual link between

f_{r, k - 1}

and

f_{r, k}

onto VL

# A_{n}

Δ_{r, k - 1}

12: update the bandwidth capacity of VL

# A_{n}

13: else

14: map the virtual link between

f_{r, k - 1}

and

f_{r, k}

onto VL

# E_{n}

Δ_{r, k - 1} .

15: update the bandwidth capacity of VL

# E_{n}

16: end if

17: end if

18: else

19: map

f_{r, k}

onto VM

# k

and process it in

Δ_{r, k - 1}

20: end if

21: end for

22: end for

Tools

Get Citation

Copy Citation Text

Bo Tian, Shanting Hu, Qi Zhang, Xiaofei Huang, Lei Zhu, Huan Chang, Xiaolong Pan, Xiangjun Xin, "Virtualization-enabled dynamic functional split and resource mapping in optical data center networks [Invited]," Chin. Opt. Lett. 23, 050001 (2025)

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Special issue on optical interconnect and integrated photonic chip technologies for hyper-scale computing

Received: Sep. 6, 2024

Accepted: Mar. 26, 2025

Published Online: May. 9, 2025

The Author Email: Bo Tian (tianbo@bit.edu.cn), Shanting Hu (hushanting@bit.edu.cn)

DOI:10.3788/COL202523.050001

CSTR:32184.14.COL202523.050001

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology