[whoisaldo]

Why

On a hybrid AMD laptop, the first session after about 20-40 s idle had its first picture about 700 ms late. The discrete card had powered off entirely (D3cold):

Idle since the card's last use	dGPU and the PCIe ports above it (dGPU, downstream, upstream, root)	D3D11 device on the dGPU
about 4 s	D0	about 160 ms
8-16 s	the dGPU in D3, then each port above it, about 5 s apart	160-163 ms
24 s and more	all in D3, root port included: power removed	780-830 ms

Whatever touches the card first pays for the wake. With a D3D11 device created first, the AMF open that follows took 74-145 ms; without one, the AMF open itself took about 830 ms.

The iPad looks for PCs over mDNS on its connect screen, seconds before it sends HELLO. That is early enough to start the wake.

What

• discovery::watch_browsers: reads mDNS on the shared port alongside the daemon, and sends nothing. A PTR or ANY question for _eternaldisplay._udp.local from another device calls on_browse.
  • The daemon's own Respond events can't be used: it answers its own probes at each registration and at the 60 s re-advertisement.
  • Queries from the PC's own addresses and loopback are ignored.
  • It reads up to mDNS's 9000 bytes per datagram and never stops on a receive error. Errors that pass (an oversized datagram, a reset, an interruption) go on at once, a few in a row; other or repeated errors wait longer each time, up to a minute.
• gpu_wake: on a browse, creates a D3D11 device on the encoder's DXGI adapter on a thread of its own and drops it at once.
  • Only the card of the encoder that streams. Each encoder open, fallbacks included, publishes it in SharedControl::wake_adapter: the pipeline GPU's adapter for a hardware encoder of its family, None for the CPU or another card's encoder (NVENC or AMF failing over to Quick Sync). Adopting a late GPU detection publishes it too, until the new pipeline's encoder opens. The browse callback reads only that.
  • Never while a viewer is connected, and never more than one at a time.
  • The wake thread checks again, just before it creates the device, that no session has started since the browse (the session count), and drops the wake if one has.
  • Holding the device for 10 s made no measurable difference to the first picture, so it isn't held.
• WakeGate: at most one wake every 15 s. Each wake that no session followed doubles the wait for the next, up to 10 min, so a device that browses and never connects causes a few wakes an hour. A session since the last wake restores the 15 s wait before the elapsed-time check, so a viewer after a long backoff doesn't wait it out (Session::sessions_started).

Tests

• discovery: a browse for the service is told from other mDNS traffic (other service types, A questions, responses, case, a second question through a compression pointer). Malformed, truncated and self-pointing packets are not browses.
• discovery, receive errors: an oversized datagram, a reset and an interruption go on at once, then back off if they keep coming. Other errors back off from the first, up to a minute, and the watch never stops.
• gpu_wake:
  • At most one wake per interval, and none in a session; one that paid off keeps the short interval.
  • A session after a ten-minute backoff restores the short interval at once. This failed before the fix: "the session's browse waited out the old ten-minute wait".
  • Only the card of the encoder that streams is woken: not for the CPU, nor for another card's encoder.
  • encoder: an open publishes the card to wake, and a Quick Sync fallback from NVENC clears it.
  • supervisor: an adopted late detection becomes the card to wake, after startup encoded on the CPU.
  • A wake is dropped when a session started first.
  • A device that browses every 20 s and never connects gets increasing gaps, settling at one wake every 10 min.
• Locally (macOS): cargo test -p eternal-host --lib passes (1231). cargo clippy --all-targets -D warnings and cargo fmt --check are clean.
• Windows: cargo clippy -p eternal-host --all-targets -D warnings is clean.

Behaviour before and after

Measured at f271721. RX 7600S hybrid laptop, h264_amf, headless synthetic still desktop. The first session after 45 s idle, 6 trials each, client time from HELLO_ACK to the first decoded frame. The client browsed for the service over mDNS (loopback only, in a local test copy) the given time before HELLO.

	trials	median
before, no browse	883, 946, 844, 944, 871, 891 ms	about 885 ms
after, no browse	869, 865, 851, 887, 956, 844 ms	about 865 ms
after, browse 1.5 s before HELLO	377, 340, 354, 802, 379, 364 ms	about 365 ms
after, browse 0.3 s before HELLO	1900, 585, 599, 628, 782, 620 ms	about 625 ms

• A warm session (3-5 s gaps) takes about 130 ms. After a wake the AMF reopen still took about 280 ms, against about 60 ms warm.
• Cost of a wake that no session follows: the card stayed in D0 for about 13.5 s and was back in D3cold 28.6 s after the wake. The laptop was on AC power, so battery drain was not measured.