Struct PerformanceMetrics 

pub struct PerformanceMetrics {
    pub throughput: f32,
    pub gpu_utilization: Vec<f32>,
    pub memory_usage: Vec<f32>,
    pub communication_overhead: f32,
    pub compression_ratio: f32,
    pub bandwidth_utilization: f32,
    pub step_time: Duration,
}

Performance metrics for distributed training.

Device telemetry is opt-in

gpu_utilization, memory_usage and bandwidth_utilization are derived from samples the embedder recorded with EnhancedDistributedTrainer::record_gpu_telemetry. When a device has never been sampled there is nothing to report, and this crate does not invent a number: the two vectors are then left empty and bandwidth_utilization is 0.0. Treat an empty vector as “unknown”, never as “0% utilized”.

Fields

throughput: f32

Samples per second, measured by the throughput tracker.

gpu_utilization: Vec<f32>

Per-GPU utilization in [0, 1]; empty when no telemetry was recorded.

memory_usage: Vec<f32>

Per-GPU memory usage in [0, 1]; empty when no telemetry was recorded.

communication_overhead: f32

Fraction of step time spent communicating.

compression_ratio: f32

Compression ratio achieved by the gradient codec.

bandwidth_utilization: f32

Mean interconnect bandwidth in MB/s; 0.0 when no telemetry was recorded.

step_time: Duration

Wall-clock time of the training step.

Trait Implementations

impl Clone for PerformanceMetrics

fn clone(&self) -> PerformanceMetrics

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PerformanceMetrics

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.