On-demand provisioning of computing resources in

computing power network with mixed CPU and

GPU
Yahui Wang, 1 Yajie Li, 1* Jiaxing Guo, 1 Yingbo Fan, 1 Ling Chen, 1 Boxin Zhang, 1 Wei Wang, 1 Yongli
2023 21st International Conference on Optical Communications and Networks (ICOCN) | 979-8-3503-4350-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICOCN59242.2023.10236419

Zhao, 1 Jie Zhang 1

1
Beijing University of Posts and Telecommunications, Beijing, China
* [email protected]

Abstract—We propose an on-demand computing II. NETWORK MODEL AND PROBLEM STATEMENT
provisioning method in computing power network with mixed
CPU and GPU. Simulation results validate that more computing A. CPN with CPU and GPU
tasks can be accommodated by optimizing the usage of Due to the dynamic nature of computing tasks, the traffic
heterogeneous computing resources. has the feature of high burstiness and high throughput in CPN.
Optical networks have become the key infrastructure in large-
Keywords—computing power network, CPU and GPU,
scale computing networks, taking into account the tremendous
computing resource provisioning.
bandwidth in fibers. The elastic optical network (EON), with
I. INTRODUCTION its flexible bandwidth allocation, has been developed to
efficiently provision services with diverse bandwidth
Computing power network (CPN) is a novel information requirements[5]. Therefore, EON has become a promising
infrastructure that enables on-demand allocation and flexible architecture in computing power networks.
scheduling of computing resources, storage resources, and
network resources between the cloud, edge, and end [1]. Transmission node Unavailable computing resource CPU CPU Resource
Heterogeneous computing resources, such as CPUs and GPUs, Computing node Unavailable FS Available FS GPU GPU Resource
are coordinated and scheduled in CPNs. The CPN with CPU
and GPU can better satisfy the computing resource demands CPU GPU

of various scenarios, including high-performance computing

and artificial intelligence [2]. A B
Computing tasks have various degrees of affinity towards
different computing resources in different scenarios, such as
CPU and GPU. CPU resources are suitable for general data E CPN
F
processing while GPU resources are utilized for tasks that CPU

require parallel processing, such as graphics rendering. There

are two typical types of computing tasks, namely general C D
computing tasks and GPU-intensive computing tasks. General Fig. 1. CPN with CPU and GPU
computing tasks can utilize either CPU or GPU resources and
We model the CPN as shown in Fig. 1, which consists of
the latter can provide computing acceleration. GPU-intensive
nodes and links. The nodes are divided into computing nodes
computing tasks are tasks that require a significant amount of
GPU resources. and transmission nodes. Computing nodes refer to nodes that
Computing tasks rely on high-performance computing provide CPU and GPU resources for computing tasks. The
resources to meet the Quality of Service (QoS) requirements. links refer to network resources, which provide frequency
If a service provider fails to properly assess needs of slots (FS) for transmission. It is necessary for service
computing tasks, it may lead to excessive or insufficient providers to consider the required network and computing
resource allocation. Moreover, it can lead to waste of resources in order to effectively deploy computing tasks.
resources due to inappropriate resource scheduling. Therefore, Computing tasks often rely on high-speed and stable network
service providers need to carry out comprehensive demand connections for data transmission. In addition, the deployment
analysis and resources allocation for tasks to improve node for a computing task must provide sufficient computing
computing efficiency and reduce resource waste. In addition, resources to meet demands.
service providers need to consider the cost efficiency of B. Problem statement
providing computing resource, ensuring that the provided
resources meet demand with a low cost [3][4]. Fig. 2 shows an example of computing resources
In this paper, the concept of resource investment costs provisioning for requests. Fig. 2(a) illustrates the network
(RIC) is proposed to quantify the usage efficiency of CPU and model and the current state of computing and spectrum
GPU in CPN. An on-demand provisioning of computing resources in CPN. Note that each CPU or GPU is subdivided
resource (OPCR) algorithm is designed for processing into 5 computing units (CUs). It is important to note that CPUs
multiple applications in a cost-effective approach. Simulation and GPUs possess distinct computing capabilities.
results validate that more tasks can be accommodated by Specifically, in this particular scenario, we assume that the
reducing 3.6% blocking ratio through optimizing the usage of performance of GPUs is two times that of CPUs.
heterogeneous computing resources. The computing task is denoted as = { , , , , } ,
where s is source node, d is date size ,b is bandwidth required,

979-8-3503-4350-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE CAMPINA GRANDE. Downloaded on October 21,2024 at 15:07:38 UTC from IEEE Xplore. Restrictions apply.
 c is computing workload, and is latency of QoS. In this of computing resource (KSP-PCR) algorithm as benchmark
illustration, we assume that request as = { , , 2,24,10}, algorithm.
and the spectrum resource from node E to computing nodes A OPCR algorithm
and D meets the requirements for request transmission. Fig. Input: Give link spectrum resource and node types
2(b) present three computing resources provisioning options
G(V, E), request = ( , , , , ) ∈ {1,2, … }, as
for the general computing tasks at computing nodes A and
D. In Fig. 2(b), the requires 4 CUs when using only CPU well as CPU and GPU resource capacity 99A<, 9BA< .
resources in node A, but 6 CUs in node D due to a longer Output: the selected computing node C9 ,C ∈ D and
transmission time. Moreover, computing tasks requires 2
CUs when using GPU resources in node A. The CUs required provisioning of computing resource for r.
of computing tasks are calculated in (1). The resource 1: for each in
provisioning solutions for GPU-intensive computing tasks are 2: search for transmission paths set A{1,2, … E}
also similar. 3: if A is empty
Transmission node Unavailable CUs CPU CPU Resource CPU Unit
4: r=(s, d , ,c , ) is blocked; break
Computing node Unavailable FS GPU GPU Resource GPU Unit
5. else if A is not empty
Computing resource
calculate CUs required of in each C9 ;
CPU GPU

A CPU/GPU 6:
Node N CPU N GPU Capacity/a CPU(GPU)
A B A
E
3
4
3
0
5 Unit
5 Unit
7: then exclude C9 in D9 with insufficient CUs ;

E F D = 200km Tr = 2 time units 8: if there is no D9 available

tt = 4 At =6
c
r={E,d,2,24,10} tt + tc ≤ tQ
9: r=(s, d ,b ,c , ) is blocked; break
CPU E
C D 10: for each available C9 in set D9
r={s,d,b,c,tQ } tt = 6 C D
D = 300km t =4
c 11: for each E in A
(a) Network model and resource state
12: calculate RIC in (2) and add it in set F ; end
A Option I A Option II D Option III 13: end
CPU GPU
14: sort all the provisioning schemes of computing
CPU
resource in ascending order of F;
15: deploying tasks with the first scheme in F;

IV. SIMULATION SETUP AND RESULT

CU CPU = c = 4 CU = c =2 CU CPU = c =6
tc GPU tc RP tc
RIC = (200 × 2 × 0.03 RIC = (200 × 2 × 0.03 RIC = (300 × 2 × 0.03
A. Simulation setup
+4 × 2 × 5) ⋅ 2 = 104 +2 × 4 ×1) ⋅ 2 = 40 √ +6 × 2 ×1) ⋅ 2 = 60 1 11 600 12
1500

(b) computing resource provisioning options for service

2 4 8
7 9 13
5
Fig. 2. Illustration of computing resource provisioning for computing tasks.
1800 14
Equation (1) comprises transmission time and computing 3
6
10
time, where n is FS number, m is modulation formats, b is Transmission node Node with CPU and GPU Node with CPU

bandwidth of each FS, which is 6.25 Gbps, D is distance from Fig. 3. NSFNET topology
source node to computing node, and is transmission speed As shown in Fig.3, NSFNET is used as the simulation
which is 5 / . The aforementioned three resource topology. We consider computing tasks in the form of data
provisioning options are all available for request deployment. flow, and their parameters value are set in one unit of time.
If we deploy computing tasks using a computing resource The request arrival and duration time follow Poisson and
allocation based on the K-shortest path (KSP) algorithm while exponential distributions, respectively. Specific parameters
disregarding the RIC of computing tasks, resources may be are shown in TABLE I. 95 6 and 9G 6 represent the capacity
wasted due to inefficient utilization. RIC is calculated using of CPU and GPU resource, where (10,15CUs) refers to 10
(2), where is transmission cost, is computing resource CPUs, with each CPU having 15 CUs in computing node.
cost, T is type of computing resource, is duration time,
and ) is cost of using inactive CPU or GPU resources. σ is 1 TABLE I. PARAMETER
when the CPU or GPU is active. Therefore, when deploying a Parameter Value
computing task, computing resources should be provided on- d [1,8] × 10H bit
demand to reduce resource waste and deployment costs. [2,5] FS
, 4 c [4,8] × 10I FLOPS
* =+ +2× 3+ (1) [10,20] ms
-×/×0 567
FS 80
89 = (2 × ; × + 9<= × × )) × (2)
99A< 3TFLOPs(10,15CUs)
III. ALGORITHM 99A< 6TFLOPs(10,15CUs)

We propose an on-demand provisioning of computing 0.03 unit/ (FS*Km*time unit)

resources (OPCR) algorithm for deploying computing tasks. 2 unit/ (CUs*time units)
9A<
The specific steps are shown in the pseudocode of OPCR BA<
4 unit/ (CUs*time units)
algorithm. Additionally, we use the KSP-based provisioning ) 5 unit

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE CAMPINA GRANDE. Downloaded on October 21,2024 at 15:07:38 UTC from IEEE Xplore. Restrictions apply.
 B. Results analysis The results of spectrum utilization ratio and blocking ratio
In this section, we evaluate the performance of KSP-PCR are shown in Fig. 5. In Fig. 5(a), we can see that OPCR
algorithm and our proposed OPCR algorithm in terms of CPU consumes 10.8% more spectrum resources than KSP-PCR
utilization, GPU utilization, spectrum utilization, blocking algorithm under the same traffic load. This is because KSP-
ratio and the sum of RIC in 350 Erlang. PCR has the shortest average path length, while the OPCR
0.8 algorithm may select a longer path to minimize the RIC. The
OPCR second factor influencing spectrum utilization is the blocking
KSP-PCR 67.3%
12.7% ratio. As shown in Fig. 5(b), KSP-PCR has a higher blocking
CPU Utilization

0.6
54.6% ratio, 3.6% higher than OPCR with 600 Erlangs.
7x104
0.4 OPCR

The sum of RIC[units]

6x104 KSP-PCR
0.2 5x104

4x104
0.0 3x104 50.53%
(a)
0 200 400 600 800 2x104
Task ID 1x104
1.0 OPCR 0
KSP-PCR 93.2%
GPU Utilization

14.8% 0 200 400 600 800

0.8
78.4% Task ID
0.6
Fig. 6. The sum of RIC
0.4 Fig. 6 shows the sum of RIC, where the sum of RIC in
0.2
OPCR is approximately 50.53% of KSP-PCR. This is because
KSP-PCR does not consider the RIC when provisioning
0.0
（ b）
resources for computing tasks, whereas OPCR provides
0 200 400 600 800 computing resources on-demand with the minimum RIC.
Task ID CONCLUSION
Fig. 4. (a). CPU Utilization (b) GPU Utilization We propose the concept of RIC to quantify the usage
Fig. 4 shows the results of CPU utilization and GPU efficiency of CPU and GPU in CPN. An OPCR algorithm is
utilization ratio when there are 800 requests in 350 Erlang. It then designed for processing computing tasks. It can achieve
is clear that as the number of requests increase, more CPUs the minimum RIC of computing tasks under the condition of
and GPUs are required to meet the demands of computing resource capacity constraints. When the traffic load is 600
tasks. We can observe that OPCR consumes 12.7% more CPU Erlang, the algorithm can accommodate more computing tasks
resources than the KSP-PCR algorithm. Furthermore, for the by reducing 3.6% blocking ratio through optimizing the usage
same traffic load and number of requests, the KSP-PCR of heterogeneous computing resources.
algorithm occupies 14.8% more GPU resources than OPCR.
This is attributed to OPCR achieving a lower blocking ratio ACKNOWLEDGMENT
compared to KSP-PCR. Additionally, the OPCR algorithm This work is supported in part by Beijing Natural Science
enables on-demand provisioning of computing resources in Foundation (4232011), the Project of Jiangsu Engineering
the lower RIC for computing tasks. Research Center of Novel Optical Fiber Technology and
0.8 Communication Network, Soochow University (SDGC2117),
OPCR 72.6%
the Fundamental Research Funds for the Central Universities
Spectrum Utilization

KSP-PCR 10.8%
0.6 61.8% and NSFC (61831003, 62021005, 62101063).
REFERENCES
0.4
[1] X. Tang et al., "Computing power network: The architecture of
convergence of computing and networking towards 6G requirement,"
0.2 China Communications, vol. 18, pp. 175-185, Feb. 2021
[2] D Du, Q Liu, X Jiang, Y Xia, B Zang, and H Chen. “Serverless
0.0 computing on heterogeneous computers”. In Proceedings of the 27th
（ a） ACM International Conference on Architectural Support for
0 200 400 600 800 Programming Languages and Operating Systems (ASPLOS '22).
Task ID Association for Computing Machinery, New York, NY, USA, 797–
813. 2022.
0.08 OPCR
KSP-PCR [3] L. Liu, J. Feng, X. Mu, Q. Pei, D. Lan and M. Xiao, "Asynchronous
Deep Reinforcement Learning for Collaborative Task Computing and
Blocking Ratio

0.06 3.6% On-Demand Resource Allocation in Vehicular Edge Computing," in

IEEE Transactions on Intelligent Transportation Systems.
0.04
[4] X. Chen, W. Li, S. Lu, Z. Zhou and X. Fu, "Efficient Resource
0.02
Allocation for On-Demand Mobile-Edge Cloud Computing," in IEEE
Transactions on Vehicular Technology, vol. 67, no. 9, pp. 8769-8780,
0.00
Sept. 2018.
（ b） [5] O. Gerstel, M. Jinno, A. Lord and S. J. B. Yoo, “Elastic optical
100 200 300 400 500 networking: a new dawn for the optical layer?” IEEE Commun. Mag.,
Traffic load vol. 50, no. 2, pp. s12-s20, Feb. 2012.
Fig. 5. (a). Spectrum Utilization (b) Blocking Ratio

Authorized licensed use limited to: UNIVERSIDADE FEDERAL DE CAMPINA GRANDE. Downloaded on October 21,2024 at 15:07:38 UTC from IEEE Xplore. Restrictions apply.