Add time-based execution, megapixel scoring, and fast-jobs-first ordering to VideoTranscodeBench (#565)

[marziehlenjaniMeta]

Summary:

Motivation
The existing video_transcode_bench_svt_mini job uses --sample-rate 0.01
to reduce runtime, but this approach has fundamental problems on
high-core-count machines:

1. Core underutilization: With only ~1% of clips sampled, the total
   number of encode jobs (clips x resolutions x CRF values) can be fewer
   than available cores. On a 72-core machine, many cores sit idle for the
   entire run — the benchmark measures a fraction of the machine's
   capacity.
2. Score instability: The throughput-based score (GB/s) varies
   significantly with sample-rate because different subsets of clips have
   different total sizes and encoding characteristics. A 1% sample gives a
   different score than a 10% sample, making cross-run comparisons
   unreliable.
3. Unrepresentative workload: Sampling removes clips rather than
   shortening the run, so the remaining workload may not reflect the full
   distribution of resolutions and content types.

A time-based approach solves all three problems: use the full dataset
(sample-rate=1.0) so all cores stay saturated, and cap runtime with a
time limit instead. This ensures every machine — regardless of core
count — runs at full utilization for a consistent, predictable duration.

Additional challenges required further changes:

4. Score depends on which jobs complete: With a time limit, only a
   subset of jobs finish. The GB/s throughput metric is biased by
   resolution (1080p jobs contribute more bytes than 144p for similar
   compute). A new megapixel-based metric (MPx/s) normalizes across
   resolutions, making the score stable regardless of which jobs complete.
5. Slow drain phase: Jobs are ordered large-to-small by default. With a
   short time limit, all slots fill with slow 1080p jobs, few complete
   before the deadline, and in-flight jobs take 30-60s to drain. A new
   --fast-jobs-first option reverses the order, maximizing completions and
   reducing drain to 1-3s.
6. Inflated total_data_encoded: The previous calculation counted all
   input clips regardless of whether they were actually encoded. Now
   derived from joblog data — only successfully completed, deduplicated
   input files are counted.

Summary

Adds a time-bounded execution mode to VideoTranscodeBench that caps
benchmark runtime via SIGTERM-based parallel job control, along with a
resolution-normalized megapixel scoring metric and an option to
prioritize fast jobs.

Key changes:

• timed_parallel_feeder.sh (new): Wraps GNU parallel with a time limit.
  When max_time > 0, runs parallel in the background and sends SIGTERM
  after the deadline, allowing in-flight jobs to finish gracefully while
  preventing new jobs from spawning. Produces a joblog for post-hoc
  analysis. When max_time = 0, behaves identically to the original flow.
• Megapixel score mode (--score-mode megapixel): Computes throughput as
  MPx/s by extracting resolution and frame count from input filenames in
  the joblog. This normalizes across resolutions, making scores stable
  regardless of which job subset completes in timed runs. Baseline derived
  mathematically from existing throughput baseline: 86.12 MPx/s.
• Effective time from joblogs (megapixel mode only): When
  score_mode=megapixel, uses sum(JobRuntime for successful jobs) /
  num_parallel_slots instead of wall-clock time. The default throughput
  mode continues to use wall-clock time, preserving backward compatibility
  with existing baselines.
• total_data_encoded fix: Now derives encoded data size from unique
  input files of successful jobs (via joblog), instead of counting all
  clips regardless of completion. Deduplicates across CRF values with sort
  -u.
• --fast-jobs-first: Reverses command file ordering (via tac) so
  small/fast jobs run first. For short timed runs this maximizes job
  completions and minimizes the drain phase (1-3s vs 30-60s).
• Two new job definitions in jobs.yml:
  • video_transcode_bench_svt_timed: 600s limit, megapixel scoring,
    fast-jobs-first
  • video_transcode_bench_svt_timed_mini: 15s limit, megapixel scoring,
    fast-jobs-first
• Backward compatible: Existing jobs are unchanged — max_time=0 means no
  limit, throughput GB/s scoring with wall-clock time, original job
  ordering.

Reviewed By: YifanYuan3

Differential Revision: D98639355

[meta-codesync]

@marziehlenjaniMeta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D98639355.