2.3 KiB
After run-ahead: there is no second lever
Client GL thread, inproc, ranked by SELF time (the same capture as PERF-PROFILE-7c6f6886.md):
| self | symbol |
|---|---|
| 31.58% | SessionProducer::WaitForAppliedOrEventBacklog |
| 3.72% | MGPipeValidateForVerb |
| 3.43% | MGPipeTracker::Update |
| 2.77% | Client::ReadDrawBindings |
| 1.78% | MGPipeSlotAllocator::FindByLifetimeId |
| 1.55% | BufferState::GetBindingSlot |
| 1.36% | MGPipeShaderBufferEmitter::EmitClass |
| 1.30% | change_protection (kernel: the persistent-map dirty-page tracking) |
| 1.26% | MGPipeFillAccess::CopyField |
| 1.26% | __memcpy_aarch64_nt |
| 1.13% | PipeWireEncoder::EncodeRecord |
One symbol is 31.6%. The next is 3.7%. Everything after the wait is a long flat tail of one-to-three-percent functions.
That has two consequences, and they are the honest answer to "optimise inproc until it matches monolith":
- The projection in
PERF-CORRECTION.mdis sound. There is no hidden concentration for the estimate to have missed — removing the wait really does leave a thread whose cost is the sum of many small things, which is what a ~3.9 ms/frame client looks like. - Going PAST parity has no single lever. After P5e there is no 10% function to delete. Reaching,
say, 1.5x monolith would mean a long grind across the tail: the tracker walk, the draw-binding
read, the slot allocator's lifetime lookup, the shader-buffer emitter, the encoder, and the
persistent-map page tracking that shows up here as kernel
change_protection. Each is worth one to three percent of one thread.
So the correct thing to promise is parity, and the correct thing to do after landing it is to re-measure rather than to plan the grind in advance: with the wait gone, the tail's shares all change denominator and the ranking above is not the ranking that will matter.
One more caution for whoever does re-measure. change_protection at 1.30% is kernel time inside the
mprotect-based dirty-page tracking P5d introduced for persistent maps, and MOBILEGL_IPC_PERSISTENT_HASH_SUPPRESS=0
plus MOBILEGL_IPC_PERSISTENT_BLOCK_KB are its A/B knobs. It is the one item in the tail with a
ready-made control arm, so it is the cheapest thing to attribute properly and the easiest to
mis-attribute by staring at self-percent.