OCS Core Working Mechanism: Photonic Path Scheduling Through Control and Data Plane Decoupling

As AI foundation models, high-performance computing (HPC), and hyperscale AI clusters continue to expand, conventional electrical switching-based data center networks are facing increasing pressure in terms of bandwidth, power consumption, and traffic management.

In particular, AI training work""s generate massive amounts of east-west traffic between GPU nodes. Collective communication operations such as AllReduce and All-to-All require sustained, high-bandwidth data exchange across large numbers of computing nodes. Traditional electrical switches must parse, buffer, queue, and forward packets, creating increasing power and processing overhead as network speeds move toward 800G, 1.6T, and beyond.

Optical Circuit Switching (OCS) offers a fundamentally different approach: instead of continuously processing individual packets, it establishes dedicated end-to-end optical paths and allows high-volume traffic to travel directly through the optical domain. The core value of OCS is therefore not simply making light travel faster. Rather, it dynamically reconfigures the physical path of optical signals through an optical switching matrix. Once an optical circuit is established, traffic can travel directly through the optical path without requiring the switch to parse packets or perform conventional electrical switching operations.

At the heart of OCS is a simple architectural principle:
· The control plane decides where the optical signal should go, while the data plane provides the physical optical path for the signal to travel through.
· This separation between control and data planes is fundamental to understanding how OCS works and why it is becoming increasingly relevant to AI data center networks.

The entire process can be summarized in four stages: connection establishment → optical switching → transparent transmission → connection release.

1. Connection Establishment: The SDN Controller “Builds the Road”
When an AI cluster starts a training task, the upper-layer AI scheduling system can identify communication requirements between different GPU nodes in advance. For example, collective communication operations such as AllReduce may require a high-bandwidth connection between Groups A and B.

The AI scheduler sends a connection request to the SDN controller. Based on the network topology and available port resources, the controller calculates the required path, reserves the corresponding OCS input and output ports, and sends configuration commands to the optical switching hardware. After receiving the command, the OCS establishes the specified physical optical path.

Importantly, this stage mainly involves connection configuration rather than business data transmission. Control signals are transmitted through an independent management channel and remain separated from the optical data traffic.

2. Optical Switching: The Optical Matrix Changes the Physical Path
Once the optical signal enters the OCS, the internal optical switching matrix establishes the required path. OCS architectures can use different technologies, including MEMS, silicon photonics, and piezoelectric switching technologies.

Taking MEMS OCS as an example, the controller adjusts the angles of an array of micro-mirrors, causing the optical signal entering one port to be reflected toward the target output port and forming an end-to-end physical optical path.

The key point is: There is no O-E-O (Optical-Electrical-Optical) conversion.

The signal remains in the optical domain throughout the switching process. The OCS does not analyze packets, buffer traffic, or replicate data. Its primary function is simply to change the direction of the optical signal.

3. Transparent Transmission: Direct High-Speed Optical Connectivity
Once the optical path is established, business traffic can travel continuously through the dedicated connection.Because OCS primarily manages optical connectivity rather than processing data content, it provides a high degree of rate, protocol, and modulation-format transparency. Whether the traffic uses 400G, 800G, or future higher-speed interfaces, the optical switching hardware itself does not need to process the underlying data, provided that the optical interfaces and system specifications are compatible. This means that as network speeds evolve, operators can focus on upgrading the optical modules and endpoint equipment without necessarily replacing the entire optical switching infrastructure, helping extend equipment lifecycles.

However, this transparency also creates limitations. OCS is not inherently optimized for broadcast, multicast, or large numbers of short-lived connections, because it establishes dedicated optical paths rather than flexibly replicating and forwarding individual packets.

4. Connection Release: Reclaiming Optical Network Resources
When an AI training task is completed, or when nodes no longer require a dedicated connection, the upper-layer system sends a release request to the SDN controller. The controller issues a disconnect command, and the OCS optical switching matrix restores the corresponding switching state. The existing connection is then released, and the occupied input and output ports return to the resource pool.These ports can subsequently be assigned to new tasks, enabling dynamic scheduling and reuse of optical network resources.

Key Advantages and Limitations of OCS
Based on its operating mechanism, OCS offers three major advantages:
① Control and data plane decoupling
Network scheduling and business traffic transmission are separated, preventing control signaling from directly participating in data forwarding.
② All-optical transmission without O-E-O conversion
OCS reduces the power consumption and latency associated with optical-electrical conversion and electrical-domain data processing.
③ High service transparency
OCS can support the evolution from 800G and 1.6T to future higher-speed optical interconnects without requiring the optical switch to process the underlying traffic.

At the same time, OCS has several inherent limitations:
· A connection must be established before data transmission. Mainstream MEMS OCS typically operates at millisecond-level switching times, making it unsuitable for instantaneous microbursts.
· Limited broadcast and multicast capabilities. OCS lacks the flexible data replication capabilities of electrical packet switches.
· Dedicated resource al. Once an optical path is established, the associated ports generally remain occupied until the connection is released.

Therefore, OCS is particularly well suited to high-bandwidth, long-duration, and relatively predictable traffic, such as AI model training, HPC, and DCI.

The more realistic future architecture is not “OCS replacing electrical switching,” but rather a hybrid network based on OCS + EPS + SDN/AI scheduling.

In this architecture, electrical packet switches handle flexible, low-volume, and complex packet traffic; OCS provides dedicated high-bandwidth optical paths for large and persistent flows; while the AI/SDN control plane dynamically allocates optical network resources according to computing work""s.

From this perspective, the value of OCS goes beyond simply “switching light.” Its greater significance lies in transforming optical networks from static connectivity into dynamically scheduled resources driven by computing work""s.