Design and Implementation of BMT (Bili Media Transport) Streaming Protocol for Reducing Backhaul Bandwidth

BiliBili's streaming media technology team introduced a new proprietary streaming protocol called BMT (Bili Media Transport) to address the high back‑haul bandwidth costs caused by supporting both FLV and HLS protocols.

Background : Since 2020 BiliBili has gradually migrated to a self‑developed P2P solution, saving billions of yuan in downstream bandwidth. However, because both FLV and HLS are still used for downstream delivery, CDN must fetch the same content twice, doubling the back‑haul bandwidth.

Problem : Before HLS, the live system only used RTMP/FLV, with all users watching FLV via CDN. After HLS was added, most users switched to HLS, but legacy clients still required FLV, forcing CDN to perform separate back‑haul for each protocol.

Technical selection : The team evaluated existing media protocols (WebRTC, RTP, CMAF) and concluded that a custom protocol was needed to satisfy BiliBili’s strict consistency requirements for CDN‑cached media segments. BMT had to provide:

Broad codec compatibility (including support for newer codecs such as H.265 and AV1).

Multi‑track capability (up to 16 tracks) for innovative live‑streaming scenarios.

Low payload overhead to minimize extra bandwidth.

Simplified handshake to reduce connection latency.

A comparison table shows BMT’s packet sizes and overhead are significantly lower than RTMP/FLV and CMAF:

BMT

RTMP/FLV

CMAF

Audio packet size

16 bytes

17 bytes

140 bytes

Video packet size

12 bytes

13 bytes

134 bytes

Encapsulation overhead (1‑payload‑ratio)

0.338 %

0.349 %

3.489 %

Core library development : Based on the protocol specification, the team built a C library libbmt that is used by internal streaming services. Compatibility extensions were added to popular components such as FFmpeg and MPV.

Back‑haul architecture : The new back‑haul design reuses existing assets, employing QUIC on the CDN side for transport‑layer acceleration and dynamic edge slicing to fetch only the needed segments, further reducing back‑haul traffic.

Production results : By May 2023, BMT was rolled out to roughly 90 % of live streams, saving over 100 GB of CDN back‑haul bandwidth per day. The protocol also enabled multi‑track live events, such as the 2022 “Cross‑Evening” show.

Future outlook : Beyond back‑haul, BMT will be used for upstream push streams, performing logical slicing on the client side to lower cloud CPU usage. The protocol will also carry business‑specific metadata (e.g., AI‑generated scene tags) to enable richer interactive live‑streaming experiences.