COMPASS：面向算力网络时空资源错配的算网一体化路由机制

肖微; 赵宝康; 苏金树; 黄雪锋; 时维嘉

doi:10.11999/JEIT260555

COMPASS：面向算力网络时空资源错配的算网一体化路由机制

doi: 10.11999/JEIT260555 cstr: 32379.14.JEIT260555

1.
国防科技大学计算机学院长沙 410000
2.
军事科学院北京 100850

基金项目: 国家自然科学基金 (U22B2005)，国家自然科学基金 (62372462)

详细信息

作者简介:
肖微：女，博士生，研究方向为算力网络、计算机网络，邮箱　xw@nudt.edu.cn

赵宝康：男，副教授，研究方向为计算机网络、人工智能网络优化

苏金树：男，研究员，研究方向为高性能网络、领域定制网络，邮箱　birchsu@139.com

黄雪锋：男，博士生，研究方向为算力网络、计算机网络

时维嘉：男，博士生，研究方向为计算机网络、网络安全

通讯作者:
苏金树　birchsu@139.com

中图分类号: TN915
计量
- 文章访问数: 14
- HTML全文浏览量: 2
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2026-05-03
- 修回日期: 2026-09-15
- 录用日期: 2026-09-15
- 网络出版日期: 2026-09-19

COMPASS: An Integrated Computing-Network Routing Mechanism for Spatiotemporal Mismatch in Computing Power Network

1.
College of Computer Science and Technology, National University of Defense Technology, Changsha 410000, China
2.
Academy of Military Science, Beijing 100850, China

Funds: National Natural Science Foundation of China (Grant No. U22B2005), National Natural Science Foundation of China (Grant No. 62372462)

摘要

摘要: 数字经济推动全球数据量指数级增长，对底层算力基础设施提出严苛挑战。算力网络通过算力与网络融合实现算力资源全局优化调度，但算力资源分布不均与任务到达不平衡性制约其性能。现有路由机制因缺乏计算和网络的全局视角和动态适应能力，导致任务失败率上升、系统负载不均衡。为此，本文提出基于深度强化学习的算力网络智能路由算法COMPASS。COMPASS通过学习任务、网络和计算复杂特征，融合网络拓扑、算力资源状态、任务特性的三维状态表征，采用目标算力节点与路由路径联合动作设计，并引入多目标加权奖励机制优化任务完成时间和失败率等关键指标。为验证方案有效性，本文开发了算力网络路由仿真平台并集成真实网络拓扑。实验表明，COMPASS较传统方法奖励值提升48%，负载均衡性能改善，在任务分布不均衡场景下将任务失败率从27.0%降至9.5%，展现强适应性。
- 算力网络 /
- 强化学习 /
- 路由 /
- 调度
Abstract: Objective Global data growth challenges computing infrastructure, while Computing Power Networks (CPN) face a critical limitation: separating network routing from computing offloading yields suboptimal scheduling due to inconsistent computing and network state dimensions. This spatiotemporal resource mismatch causes nearly 20% higher task failure rates in imbalanced versus uniform task arrival scenarios. Furthermore, existing DRL-based scheduling schemes lack topology generalization and flexible action spaces for arbitrary CPN nodes. This work aims to design an end-to-end integrated computing-network routing mechanism to resolve this mismatch, jointly optimize task failure rates and completion times, and enhance CPN dynamic adaptability and service quality. Methods First, we construct a centralized software-defined CPN architecture, mathematically modeling task arrivals, network bandwidth, computing resources, and service times, alongside a multi-objective function minimizing task failures and overall service time. Next, we develop the DDQN-based COMPASS algorithm featuring: (1) a 3D state representation integrating network topology, real-time computing status, and task attributes; (2) a joint, truncated action space coupling computing node selection and routing paths to reduce exploration complexity; and (3) a multi-objective weighted reward with heavy penalties for task failure to prevent poor strategy convergence. A tailored training workflow addresses experience delays caused by execution latency, incorporating routing and DRL training algorithms. To validate performance, we developed the open-source CNRSim platform based on EdgeCloudSim, integrating three real backbone topologies, four task types, and mainstream benchmark schemes. Results and Discussions Parameter tuning identified an optimal action space dimension of 20 (Fig. 5) and a learning rate of 5.0E-5, effectively balancing convergence speed and final performance (Fig. 6). Evaluated on three real backbone topologies, COMPASS outperformed traditional benchmarks, improving the average reward by 48% (Fig. 7) and achieving the lowest cumulative task failure rate (Fig. 8). Notably, under imbalanced task distributions, it significantly reduced the failure rate from 27.0% to 9.5% (Fig. 9). In-depth analysis on the GEANT2 topology revealed that COMPASS ensures superior computing load balancing across servers (Fig. 10) and processes more tasks with lower bandwidth occupancy (Fig. 11). Failure analysis indicated that jointly optimizing resources mitigates failures caused by bandwidth shortages and node overloads (Fig. 9). Furthermore, COMPASS maintained peak average rewards under dynamic workloads (Table 3). Conclusions This paper proposes COMPASS, a DRL-based integrated computing-network routing algorithm, to address CPN bottlenecks caused by spatiotemporal mismatches between resource distribution and task arrivals. By utilizing a 3D state representation, joint action design, and a multi-objective reward mechanism, COMPASS enables global collaborative scheduling. Extensive experiments on the CNRSim platform confirmed its superiority over benchmarks in reward maximization, failure reduction, and load balancing, particularly in imbalanced scenarios. This work offers a framework for CPN optimization and large-scale distributed resource scheduling. Future work will explore task fault tolerance and retransmission mechanisms, and investigate the potential of collaborative optimization by combining COMPASS with emerging intelligent technologies.
- Computing Power Network /
- Reinforcement Learning /
- Routing /
- Scheduling

HTML全文

图 1 单指标优化与多维资源协同调度冲突示意图

下载: 全尺寸图片幻灯片

图 2 资源负载不均对系统任务失败率的影响分析

下载: 全尺寸图片幻灯片

图 3 软件定义算力网络架构

下载: 全尺寸图片幻灯片

图 4 现有和本文的动作空间设计对比

下载: 全尺寸图片幻灯片

图 5 不同动作空间维度的奖励收敛曲线

下载: 全尺寸图片幻灯片

图 6 对模型学习率进行调参

下载: 全尺寸图片幻灯片

图 7 在三种拓扑下各模型的奖励对比

下载: 全尺寸图片幻灯片

图 8 三种拓扑中各策略的任务失败率及原因

下载: 全尺寸图片幻灯片

图 9 在GEANT2拓扑中评估任务失败原因

下载: 全尺寸图片幻灯片

图 10 在GEANT2下评估计算节点平均负载

下载: 全尺寸图片幻灯片

图 11 在GEANT2下评估链路平均带宽

下载: 全尺寸图片幻灯片

1 为到达算力网络的任务决策路由

In：task
Out：route
1：state ← getState()
2：action ← agent.act(state, task)
3：k_path ← network.get_k_route()
4：route ← k_path[action]
5：memory.set(state, action)
6：previousMemory.set_next_state(state)

下载: 导出CSV

2 训练智能体

In：a complete task, service_time
1：isDone ← service_time==-1 ? true ; false
2：reward ← compute_reward(isFailed, serviceTime)
3：state_id ← task.get_state_id()
4：memory.set(state_id, reward, isDone)
5：for m in memory do
6：　agent.train(m)
7：end for

下载: 导出CSV

表 1 不同类型任务的属性参数

属性	类型1	类型2	类型3	类型4
任务到达间隔(秒/s)	2	3	20	7
上传数据(千字节/KB)	1 500	20	2 500	25
下载数据(千字节/KB)	25	1 250	200	1 000
算力使用率(%)	6	2	30	10
出现频率(%)	30	20	20	30

下载: 导出CSV

表 2 实验参数定义

参数名称	取值	参数名称	取值
强化学习折扣因子γ	0.95	更新目标Q网络间隔	10
模型学习率	0.00005	模型的经验池大小	1000000
ε-Greedy策略中ε的渐变率	0.999	式11调节奖励下界$ {\alpha }_{1} $	0
ε-Greedy策略中ε的渐变最小值	0.1	式11调节奖励中任务失败率权重$ {\alpha }_{2} $	10
模型学习的批次大小	4	式11调节奖励中服务时间权重$ {\alpha }_{3} $	1

下载: 导出CSV

表 3 多种任务强度下奖励评估

任务强度	COMPASS	DeepEdge	Greedy	SIH	SPFC
200	64.29±61.13	48.94±48.01	54.24±56.89	60.85±44.72	72.04±49.65
500	85.07±125.36	–9.48±151.28	56.37±144.96	–203.27±228.70	–130.89±174.80
1000	–189.97±416.35	–1331.37±773.69	–1337.55±917.71	–1913.64±1004.45	–1970.19±1066.98
1500	–3493.29±1731.12	–2565.61±1781.74	–3810.74±1604.53	–5308.04±1898.38	–4831.48±1817.08

下载: 导出CSV

参考文献(24)

[1]	HE Min, ZHANG Xiaoyu, and WANG Liangmin. U’Dedup: Updatable block-level deduplication scheme over similar data in fog-assisted cloud storage[J]. IEEE Internet of Things Journal, 2025, 12(16): 33348–33362. doi: 10.1109/JIOT.2025.3574851.
[2]	SUN Yukun, LEI Bo, LIU Junlin, et al. Computing power network: A survey[J]. China Communications, 2024, 21(9): 109–145. doi: 10.23919/JCC.ja.2021-0776.
[3]	WANG Wenzheng, XIE Renchao, TANG Qinqin, et al. Service anycast forwarding for software defined computing power network[C]. 2025 IEEE Wireless Communications and Networking Conference (WCNC), Milan, Italy, 2025: 1–6. doi: 10.1109/WCNC61545.2025.10978491.
[4]	TANG Liou, KRISHNAMURTHY P, and ABDELHAKIM M. Is machine learning the best option for network routing?[C]. ICC 2024 - IEEE International Conference on Communications, Denver, USA, 2024: 5425–5430. doi: 10.1109/ICC51166.2024.10622335.
[5]	HE Qiang, WANG Yu, WANG Xingwei, et al. Routing optimization with deep reinforcement learning in knowledge defined networking[J]. IEEE Transactions on Mobile Computing, 2024, 23(2): 1444–1455. doi: 10.1109/TMC.2023.3235446.
[6]	YAMANSAVASCILAR B, BAKTIR A C, SONMEZ C, et al. DeepEdge: A deep reinforcement learning based task orchestrator for edge computing[J]. IEEE Transactions on Network Science and Engineering, 2023, 10(1): 538–552. doi: 10.1109/TNSE.2022.3217311.
[7]	陈健, 马天瑞, 杨龙, 等. 面向移动边缘计算的协作NOMA安全卸载能耗优化[J]. 电子与信息学报, 2026, 48(2): 534–544. doi: 10.11999/JEIT250606. CHEN Jian, MA Tianrui, YANG Long, et al. Energy consumption optimization of cooperative NOMA secure offload for mobile edge computing[J]. Journal of Electronics & Information Technology, 2026, 48(2): 534–544. doi: 10.11999/JEIT250606.
[8]	冯伊凡, 吴畏虹, 孙罡, 等. 天地一体化边缘计算网络服务迁移算法研究[J]. 电子与信息学报, 2026, 48(2): 499–511. doi: 10.11999/JEIT250835. FENG Yifan, WU Weihong, SUN Gang, et al. Service migration algorithm for satellite-terrestrial edge computing networks[J]. Journal of Electronics & Information Technology, 2026, 48(2): 499–511. doi: 10.11999/JEIT250835.
[9]	LIN Xiao, WU Ruolin, MEI Haibo, et al. A game incentive mechanism for energy efficient federated learning in computing power networks[J]. Digital Communications and Networks, 2024, 10(6): 1741–1747. doi: 10.1016/j.dcan.2023.10.006.
[10]	SUN Wen, LI Zongjun, WANG Qubeijian, et al. FedTAR: Task and resource-aware federated learning for wireless computing power networks[J]. IEEE Internet of Things Journal, 2023, 10(5): 4257–4270. doi: 10.1109/JIOT.2022.3215805.
[11]	DI Zheng, LUO Tao, QIU Chao, et al. In-network pooling: Contribution-aware allocation optimization for computing power network in B5G/6G era[J]. IEEE Transactions on Network Science and Engineering, 2023, 10(3): 1190–1202. doi: 10.1109/TNSE.2022.3225292.
[12]	YU Tiankuo, YANG Hui, YAO Qiuyan, et al. Multi-visual-GRU-based survivable computing power scheduling in metro optical networks[J]. IEEE Transactions on Network and Service Management, 2024, 21(1): 1302–1315. doi: 10.1109/TNSM.2023.3314272.
[13]	王钦定, 谭斌, 黄光平, 等. 面向AI计算服务的算网融合轻量化增量部署方法[J]. 电子与信息学报, 2026, 48(2): 512–521. doi: 10.11999/JEIT250663. WANG Qinding, TAN Bin, HUANG Guangping, et al. Lightweight incremental deployment for computing-network converged AI services[J]. Journal of Electronics & Information Technology, 2026, 48(2): 512–521. doi: 10.11999/JEIT250663.
[14]	庞冉, 易昕昕, 辛亮, 等. 算力网络路由调度技术研究[J]. 电信科学, 2023, 39(8): 149–156. doi: 10.11959/j.issn.1000-0801.2023163. PANG Ran, YI Xinxin, XIN Liang, et al. Research on routing scheduling technology of computing power network[J]. Telecommunications Science, 2023, 39(8): 149–156. doi: 10.11959/j.issn.1000-0801.2023163.
[15]	LIN Zhiyi, WANG Lingjie, NING Wenxin, et al. Heuristic-based computing-aware routing for dynamic networks[J]. Electronics, 2025, 14(18): 3724. doi: 10.3390/electronics14183724.
[16]	FENG Li, XIE Renchao, TANG Qinqin, et al. CaRCS: Joint optimization of computing-aware routing and collaborative scheduling in computing power networks[J]. IEEE Network, 2025, 39(6): 270–278. doi: 10.1109/MNET.2025.3548419.
[17]	XIE Renchao, FENG Li, TANG Qinqin, et al. Delay-prioritized and reliable task scheduling with long-term load balancing in computing power networks[J]. IEEE Transactions on Services Computing, 2024, 17(6): 3359–3372. doi: 10.1109/TSC.2024.3495500.
[18]	中国信息通信技术研究院. 中国算力发展指数白皮书 (2023年)[R]. 2023. (查阅网上资料, 未找到本条文献报告编号信息, 请确认). China Academy of Information and Communications Technology. White paper on China’s computing power development index (2023)[R]. 2023. (查阅网上资料, 未找到本条文献英文翻译信息, 请确认).
[19]	ZHANG Xinchun, KASHAF A, ZOU Yihan, et al. ResLake: Towards minimum job latency and balanced resource utilization in geo-distributed job scheduling[J]. Proceedings of the VLDB Endowment, 2024, 17(12): 3934–3946. doi: 10.14778/3685800.3685817.
[20]	ZHOU Cong, ZHAO Baokang, TANG Fengxiao, et al. Dynamic multi-objective service function chain placement based on deep reinforcement learning[J]. IEEE Transactions on Network and Service Management, 2025, 22(1): 15–29. doi: 10.1109/TNSM.2024.3446248.
[21]	BADAGERI T, HAMDAOUI B, and LANGAR R. Load-balanced multipath routing through software-defined networking[C]. 2024 International Wireless Communications and Mobile Computing (IWCMC), Ayia Napa, Cyprus, 2024: 1068–1073. doi: 10.1109/IWCMC61514.2024.10592316.
[22]	KIM B, KONG J H, MOORE T J, et al. Deep reinforcement learning based routing for heterogeneous multi-hop wireless networks[C]. MILCOM 2025 - 2025 IEEE Military Communications Conference (MILCOM), Los Angeles, USA, 2025: 618–623. doi: 10.1109/MILCOM64451.2025.11310271.
[23]	ZHANG Haijun, HUANG Miaolin, ZHOU Huan, et al. Capacity maximization in RIS-UAV networks: A DDQN-based trajectory and phase shift optimization approach[J]. IEEE Transactions on Wireless Communications, 2023, 22(4): 2583–2591. doi: 10.1109/TWC.2022.3212830.
[24]	CHEN Junyan, HUANG Xuefeng, WANG Yong, et al. ASTPPO: A proximal policy optimization algorithm based on the attention mechanism and spatio-temporal correlation for routing optimization in software-defined networking[J]. Peer-to-Peer Networking and Applications, 2023, 16(5): 2039–2057. doi: 10.1007/s12083-023-01489-7.