.png){kind=icon}
.png){kind=icon}
Return
ICCP: Toward Congestion Control Agent via Controlling Logic Decoupling and Algorithm Integration
.png){kind=icon}
PRE
AI
X
B
徐
J
DOI：10.1109/tnsm.2026.3714992.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address the limitations of single congestion control algorithms (CCAs) in dynamic and heterogeneous network environments, selecting an appropriate algorithm from a pool of existing ones has become a widely adopted strategy. Existing mechanisms, however, are typically constrained by the Linux kernel’s unified abstractions, which limit the flexibility of selecting from a small set of in-kernel CCAs. Learning-based CCAs further increase the deployment cost because their inference logic is often compute-intensive and can block concurrent flows when executed within a synchronous control path. In this paper, we present ICCP, a unified congestion control framework that supports both heuristic and compute-intensive algorithms for concurrent flows. Rather than introducing a new reinforcement learning method, ICCP provides a three-layer, decoupled runtime framework consisting of the protocol stack, the user-space algorithm library, and the congestion control agent. ICCP uses asynchronous request handling, a shared proxy, and a “zero-copy” serialization-based RPC path to support both batch and single-inference modes with controlled communication overhead. We implement three distinct reinforcement learning-based congestion control algorithms within ICCP, including Sage, Orca, and DTCC, to evaluate the framework using representative compute-intensive CCAs. Simulations and real-world experiments demonstrate that ICCP maintains robust and efficient communication and inference performance as the number of concurrent flows increases. Overall, ICCP provides a practical runtime framework for integrating, evaluating, and deploying heterogeneous congestion control algorithms in multi-flow environments.
Keywords:
Congestion control
reinforcement learning
control plane
batch inference
Journal
IF:
5.4
Papers:
509
Citations:
9.2K
Organization
Cited Papers