SRS gets performance tools that load and profile SRS in Docker for every
protocol: a lean load tool, an internal skill that runs an improvement
round, and a measurement method that holds up to a 10% verdict. The same
tools then measure the Go proxy. The first runs found one SRS fix and
three Go proxy fixes.
The load tool, a standalone Go module in cmd/perf (bin/srs-perf):
- Per-core worker processes shard the clients and report counters each
second. Lean clients for every protocol: RTMP publish and play,
HTTP-FLV, HTTP-TS, HLS, DASH, RTSP, SRT, WHIP and WHEP. No decode;
frames and lag come from timestamps.
- 17 scenarios: each protocol's fan-out, publishers (rtmp-pub,
whip-pub), many-streams, cross-protocol bridges (srt-rtmp, srt-whep,
whip-rtmp, whip-flv), and delay probes for the RTMP pause, forward and
edge.
- run ramps by doubling to a capacity, then holds at 50% and 90% of it.
compare checks that the setups match and reports capacity and CPU per
client at the 10% line.
- make perf-image builds an optimized SRS with symbols and frame
pointers, with perf, FlameGraph and FFmpeg. SRS runs on one pinned core
and is configured only by env files, one per scenario.
The measurement method:
- A flat CBR source, checked to ±3% per second, since SRS CPU follows
the bitrate.
- SRS CPU from its cgroup once a second. Holds have no profiler in the
measuring window; each count gets its own separate profile hold.
- perf-ab.sh builds two revisions and runs them side by side on separate
cpusets. compare gives the 95% CI of the paired per-second difference.
- perf-verify.sh checks every run: capacity, healthy steps, profiles and
symbols, noise, and what bound the run.
The internal-perf-for-srs skill:
- SKILL.md holds the rules, the shared tools and the round workflow. A
Target Router loads references/cpp-server.md for the C++ media server or
references/go-server.md for the Go proxy.
- Each reference covers its tools, topology, scenarios, measuring, how
to read its profiles (Linux perf, or Go pprof and trace), A/B, the
verified scenarios with their capacity and top bottleneck, and how to
start a round.
- srs-develop routes performance, benchmark, load test and profile tasks
to it. perf-skill-test.sh checks the text against the scenario registry.
- Reports are git-ignored, since their numbers hold only for the machine
that made them.
- The skills now name each server by its job: the C++ media server and
the Go proxy, with no "generation".
C++ media server, one core, Linux arm64 in Docker:
- Fix: an empty SrsSharedPtr holds no reference count, so each fan-out
copy of a message does one malloc instead of three. Covered by
srs_utest_ai05.cpp.
- Capacity: rtmp-rtmp, rtmp-flv, rtmp-hls, whip-rtmp and whip-flv
12,800; rtmp-dash and srt-rtmp 6,400; rtmp-ts and rtmp-rtsp 3,200;
whip-whep 2,000 (cap); rtmp-whep, srt-whep, srt-srt and rtmp-pub 1,280;
whip-pub 1,000 (cap); many-streams 640 streams.
- Top bottlenecks found, not fixed: the RTMP fan-out copy,
_st_epoll_dispatch with EAGAIN reads, 4 KB reads and writes for HLS and
DASH segments, one write per message for HTTP-TS and two per packet for
RTSP, the SRTP replay window in libsrtp, and libsrt for SRT.
- Delays: the paused RTMP player's 500 ms sleep, and the forward's and
edge publish's 500 ms receive timeout, each adding about 220 ms of lag.
The Go proxy, cmd/proxy, on one core in front of 1 to 4 SRS origins:
- perf-up.sh -target proxy, run -target proxy, which samples the proxy
and each origin, Go profiles in each profile hold (CPU, allocations,
mutex, block, trace, GC), and proxy A/B.
- Fix: HTTP-FLV, HTTP-TS and HLS relay the origin's error status,
headers and body, rather than a 500 with the origin's address.
- Fix: a WebRTC STUN binding routes by its ufrag, not a cached client
address. WHIP and WHEP failed from 160 clients on three origins.
- Capacity on three origins: rtmp-flv and rtmp-hls 12,800; rtmp-rtmp and
rtmp-ts 6,400; whip-whep 2,000 (cap); srt-srt and rtmp-pub 1,280;
whip-pub 1,000 (cap); many-streams 1,280 streams.
---------
Co-authored-by: chatgpt-codex-connector[bot] <199175422+chatgpt-codex-connector[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
- Fix a goroutine leak on the WHEP path: the backend→client reader was
being spawned on every inbound client packet (STUN keepalives + RTCP
feedback), leaking tens of thousands of goroutines under steady-state
load. Now spawned exactly once per connection via `sync.Once` on both
the RTC and SRT proxies. Listener and reader receive buffers are also
reused across iterations.
- Make the legacy SRS `/rtc/v1/play/` and `/rtc/v1/publish/` APIs work
end-to-end through the proxy. Those endpoints wrap the SDP in a JSON
envelope (`{"sdp":"v=0\r\n..."}` where `\r\n` is the literal 2-byte JSON
escape, not real CRLF), so ICE parsing previously absorbed the rest of
the body into the ufrag. Added `unwrapSDPEnvelope` for ICE extraction
and tightened `ParseIceUfragPwd`'s value class to stop at `\`. The bytes
forwarded to the client and the in-body candidate-port rewrite still
operate on the raw envelope.
- Enable `net/http/pprof` endpoints when `GO_PPROF` is set (blank import
in `internal/debug/pprof.go`) and add `docs/perf/proxy-whep.md` walking
through CPU/alloc/heap/goroutine/trace collection and `pprof -base`
before/after diffs for the WHEP workload (1 publisher + N players).
- Tighten `SRTHandshakePacket.UnmarshalBinary` to
`bytes.Clone(ExtraData)` so decoded handshakes kept on the connection
(`handshake0`, `handshake2`) stay valid once the receive buffer is
reused.
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Move build output from `./srs-proxy` to `bin/srs-proxy` following Go
project conventions, updating Makefile, .gitignore, and all
documentation references
- Replace third-party `godotenv` dependency with a custom `.env` parser
that supports comments, `export` prefix, quoted values, escape
sequences, and inline comments — with full unit tests
- Remove `ignore-worklog.md` and update `README.md` with skill-based AI
prompts
- Bump copyright year from 2025 to 2026 across all source files
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Reorganize the SRS (Simple Realtime Server) repository to
follow a conventional Go project structure, setting the stage for a
progressive transition from a C++ project to a Go project. The proxy,
which was once contained within its own `proxy/` subdirectory, will now
be converted into the initial Go module located at the root of the
repository, serving as a template for subsequent Go modules.
- **Go module at repo root:** `go.mod` moved to repo root, module
renamed from `proxy` to `srsx`. The repo is now a proper Go project with
`cmd/` and `internal/` at the top level.
- **Elevation of Proxy Code:** Move the proxy code from
`proxy/cmd/proxy-go/` to `cmd/proxy/`, and from `proxy/internal/` to
`internal/`. The proxy serves as the inaugural application; subsequent
modules (for instance, `cmd/origin`) will mimic this arrangement.
- **Documentation Restructured:** Transfer the documentation from
`proxy/docs/` to `docs/proxy/`, revise the main README to endorse
OpenClaw as the preferred AI tool, and update `proxy/README.md` to point
to the new documentation locations.
- **Build and config:** `Makefile` moved to root, `PROXY_STATIC_FILES`
default path corrected for the new layout, `.gitignore` consolidated.
- **Cleanup:** removed standalone `proxy/LICENSE` (repo-level license
applies), all internal imports updated to `srsx/internal/...`.
- **OpenClaw workspace:** added community bot info, git workflow
conventions, and support group behavior guidance.
This restructuring was performed by OpenClaw orchestrating Claude Code
and Codex via ACP.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: chatgpt-codex-connector[bot] <199175422+chatgpt-codex-connector[bot]@users.noreply.github.com>
