| Component | Specification |
|-----------|----------------|
| Hardware Platform | 2 × Intel Xeon Gold 6230R (20 cores each, 2.1 GHz base, 3.9 GHz boost) – hyper‑threaded, total 40 physical cores / 80 logical threads. |
| Memory | 32 GB DDR4‑2666 ECC (4 × 8 GB modules) |
| Storage | 2 × Samsung 970 Pro NVMe SSD (1 TB each) in RAID 0 for raw I/O throughput. |
| GPU | NVIDIA RTX 4090 (24 GB GDDR6X) – used for optional hardware‑accelerated encoding (disabled for this run). |
| Network | Dual‑port 25 GbE NIC (LACP bonded) – used only for telemetry streaming. |
| Operating System | Ubuntu 22.04 LTS (kernel 5.15.0‑1057‑generic) |
| Java Runtime | OpenJDK 21 (HotSpot) – 64‑bit, with -XX:+UseG1GC -Xms16g -Xmx24g. |
| JAVHD Build | Version v3.4.2‑beta (commit a3f7c2b). Includes the new adaptive bitrate module and updated memory‑pool manager. |
| Benchmark Harness | Custom FPRE harness (fpre‑080‑rm‑javhd.jar) – orchestrates input generation, processing, and metrics aggregation. |
| Input Media | Synthetic 4K‑60 fps HDR stream (YUV 4:2:2, 10‑bit) generated on‑the‑fly; total duration 1 hour. |
| Monitoring Stack | Prometheus 2.48 + Grafana 10.2 (scrape interval 5 s); custom exporters for JVM, cAdvisor, and Intel RAPL. |
All components were verified to be at the latest stable firmware/driver levels prior to the run.
The adaptive bitrate engine holds a rolling window of 2 seconds of frame metadata (motion vectors, histogram data). At 4K 60 fps, each frame’s metadata occupies ≈ 150 KB, yielding a ~18 MB per‑second buffer. The G1GC heap sizing (24 GB max) comfortably contains the combined workload, but the non‑heap usage spikes during encoder initialization due to native buffers.
Risk: Should the pipeline be extended to support 8K streams (≈ 4× data), the memory consumption would approach or exceed the 32 GB limit, leading to potential Out‑Of‑Memory (OOM) errors.
| Phase | Heap Used (GB) | Non‑Heap (GB) | Total (GB) | |-------|----------------|---------------|------------| | Ingest | 2.4 | 0.3 | 2.7 | | Pre‑Process | 4.1 | 0.5 | 4.6 | | Core Processing | 10.2 | 0.9 | 11.1 | | Encoding | 5.3 | 0.6 | 5.9 | | Peak Total | 19.8 | 2.3 | 22.1 (including OS) |
The G1GC collector kept pause times < 5 ms, with a young‑generation collection every 1‑2 seconds. No Full GC events occurred.
Deliver an automated, timestamped short-form video preview generator that produces a single 1:59 (one minute 59 seconds) highlight clip for a given video asset every day at 01:59:59 (local time).
The encoding stage accounts for ~62 % of total CPU cycles, as confirmed by perf top. The remaining cycles are split between HDR tone‑mapping (≈ 18 %) and adaptive bitrate calculations (≈ 12 %).
Observation: The encoding library (x264 Java binding) runs close to the theoretical maximum of the CPU’s SIMD units. Moving to a hardware‑accelerated encoder (NVENC) would likely reduce CPU load by ~30 % and free headroom for additional processing (e.g., AI‑based upscaling).