สถาปัตยกรรม SFU เทียบกับ MCU

Imagine a video call with many people. How do all their video and audio streams connect efficiently? Direct peer-to-peer connections, common in 1-on-1 WebRTC, become complex and inefficient for larger groups.

This lesson explores two main server-side architectures for scaling multi-party calls: MCU and SFU.

MCU stands for Multipoint Control Unit. Think of an MCU as a central 'mixer' for all participants' media streams.

In an MCU architecture, every participant sends their individual audio and video stream to a central server.

The MCU server receives all individual streams, decodes them, mixes them together into a single composite stream (e.g., a grid layout of videos), and then re-encodes this single stream.

Finally, the MCU sends this *single, mixed stream* back to *all* participants.

A big advantage of the MCU model is that each client only needs to send one stream (their own) and receive one stream (the mixed stream from the server).

This significantly reduces the client's bandwidth and processing requirements, making it suitable for users with weaker internet connections or less powerful devices.

While beneficial for clients, MCUs place a heavy load on the server. The server has to decode, mix, and re-encode many streams in real-time.

• Server Intensive: High CPU and memory usage.
• Quality Compromise: Re-encoding can introduce latency and reduce individual stream quality.
• Limited Customization: Clients receive a fixed layout determined by the server.

SFU stands for Selective Forwarding Unit. Unlike an MCU, an SFU does not mix or re-encode media streams.

An SFU acts as a smart router, forwarding individual streams from one participant to all others who need to receive them.

In an SFU setup, each participant sends their stream to the SFU server. The SFU then forwards each participant's stream *individually* to all other participants.

The SFU can selectively choose which streams to forward and at what quality (e.g., sending a lower resolution to clients with poor bandwidth).

With an SFU, each client receives multiple individual streams (one from each other participant). The client then decodes and renders these streams locally.

This allows for greater flexibility in layout and individual stream control on the client side. For example, a client can choose to display only a few main speakers.

Try running this example:

SFUs are highly efficient for the server as they don't decode/re-encode, just forward. This makes them more scalable for many participants.

However, clients need more bandwidth (to receive multiple streams) and more processing power (to decode and render them all locally).

• Server Efficient: Low CPU/memory per stream.
• Client Intensive: Higher bandwidth and CPU for clients.
• Flexible: Clients can customize layout and individual stream visibility.

The choice between MCU and SFU depends on your application's specific needs:

• MCU: Best for low-bandwidth clients, fixed layouts, and when server cost for processing is acceptable.
• SFU: Ideal for high-quality, flexible layouts, and when scaling to many participants is critical, assuming clients have adequate resources.

We've explored two key architectures for multi-party WebRTC calls:

• MCU (Multipoint Control Unit): Server mixes streams, sends one composite stream to each client. Good for low-resource clients, but server-intensive.
• SFU (Selective Forwarding Unit): Server forwards individual streams to clients. Offers better quality and client flexibility, but requires more client bandwidth and processing.

Choosing between them depends on your specific application requirements and user constraints for scalability and quality.