Memory & Telemetry

Live Peak Memory Sampling

Standard operating system utilities such as resource.getrusage only report process-lifetime high-water marks (ru_maxrss).

When an agent executes multiple audio operations in a single session or long-lived container, subsequent operations report the previous high-water mark rather than their own memory consumption. This obscures memory spikes and memory leaks.

The PeakMemorySampler Thread

audio-tools includes a background polling thread that samples resident set size every 100 milliseconds throughout the separation run:

• Linux Hosts: Directly queries /proc/self/status for the live VmRSS value.
• macOS Workstations: Queries system process memory descriptors.

This records accurate task-specific metrics:

• peak_rss_mb: Peak resident set size observed during this specific operation.
• memory_delta_mb: Difference between peak memory and baseline memory at task start.

Real-Time Factor ($RTF$)

The Real-Time Factor measures compute throughput relative to audio duration:

$$RTF = \frac{\text{Inference Duration (seconds)}}{\text{Audio Duration (seconds)}}$$

• An $RTF < 1.0$ indicates that inference runs faster than real-time playback.
• A standard 3-minute song ($180\text{s}$) processed on an 8-core CPU typically achieves an $RTF$ between $0.35$ and $0.55$, completing in under 90 seconds.