Skip to content

Troubleshooting

Export is slow

What decides the time of an export, and what to check before you change anything.

Last updated:

You will learn#

  • where an export spends its time;
  • what to check and what to change;
  • what is measured and what is not.

The idea#

An export has four steps: it plans the frames, composes each frame in Go, feeds the frames to FFmpeg, and verifies the file. Measured on the project's own benchmark, the limit is composing the frames (FFmpeg decoding the sources and converting them to RGBA), and the encoder barely matters. A hardware encoder therefore shortens only a small part of the time. The time grows with the number of frames, the frame size and the number of layers with effects on screen.

The numbers on a table of the repository, docs/perf/, are those of one machine at one commit, and a busy machine measures several times slower than an idle one. Measure your own machine before you decide something is wrong.

Quick start#

  1. Check what else the machine is doing. Exports share the CPU with ingests, transcriptions and your other programs.
  2. Read the job, which reports the encoder used and the duration:
Terminal
vukem-video job:list --project prj_7k2m9x4q1b --state all
vukem-video job:show job_4h8k2m6p9t
vukem-video doctor --full
  1. Export a short range to compare: vukem-video export --project prj_7k2m9x4q1b --start 0s --end 5s.
CheckWhat to do
The encoder is libopenh264 and a hardware encoder existsdoctor --full shows why the hardware one failed; fix the driver, or force one with VUKEM_VIDEO_ENCODER. The gain is small
The project is larger than the output needsset the frame size to what the platform shows; 1080 x 1920 is the usual vertical size
The sources are very large or of a heavy codecthe import makes a proxy for the editor, but the export reads the original: re-encode a smaller source
Several exports are queuedVUKEM_VIDEO_MAX_EXPORTS is 2 per workspace and 1 per user or key; the rest wait
The disk is almost full or slowthe workspace and tmp/ live on one volume; free space and use a fast disk
A container with a small CPU limitgive it more CPUs

Common mistakes#

  • Changing the encoder to speed up an export that is limited by composing.
  • Starting an export while a long transcription runs on a small machine.
  • Cancelling and restarting a long export: the work is lost. Use job:show --wait to follow it.

Next steps#