Class ProfilingConfig

Namespace

AiDotNet.Deployment.Configuration

Assembly

AiDotNet.dll

Configuration for performance profiling during model training and inference.

public class ProfilingConfig

Inheritance

object

ProfilingConfig

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: Profiling measures how long different parts of your ML code take to run. Think of it like a stopwatch for your code - it helps you find bottlenecks and optimize performance.

What gets tracked:

• Operation timing: How long each training step, forward pass, backward pass takes
• Memory allocations: How much memory is used during training
• Call hierarchy: Which operations call which other operations
• Percentiles: P50 (median), P95, P99 timing for statistical analysis

Why it's important:

• Find bottlenecks in your training pipeline
• Compare performance before and after optimizations
• Identify memory-intensive operations
• Understand your model's computational profile

The profiler uses production-ready algorithms:

• Welford's algorithm for O(1) streaming mean/variance
• Reservoir sampling for bounded-memory percentile estimation
• Configurable sampling for high-frequency operations

Properties

AutoEnableInDebug

Gets or sets whether to auto-enable profiling in debug builds (default: true).

public bool AutoEnableInDebug { get; set; }

Property Value

bool

Remarks

For Beginners: When true, profiling is automatically enabled in debug builds even if Enabled is false. Convenient for development without changing config.

CustomTags

Gets or sets custom tags to include with profiling data.

public Dictionary<string, string> CustomTags { get; set; }

Property Value

Dictionary<string, string>

Remarks

For Beginners: Add custom labels to profiling data (e.g., model name, experiment ID). Useful for comparing profiles across different runs or configurations.

DetailedTiming

Gets or sets whether to include detailed timing breakdowns (default: false).

public bool DetailedTiming { get; set; }

Property Value

bool

Remarks

For Beginners: Track sub-operation timings within each operation. More detailed but adds overhead. Use for deep performance analysis.

Enabled

Gets or sets whether profiling is enabled (default: false).

public bool Enabled { get; set; }

Property Value

bool

Remarks

For Beginners: Set to true to collect performance data during training/inference. Profiling adds some overhead, so it's disabled by default. Enable when debugging performance.

MaxOperations

Gets or sets the maximum number of unique operations to track (default: 10000).

public int MaxOperations { get; set; }

Property Value

int

Remarks

For Beginners: Maximum distinct operation names the profiler will track. Prevents unbounded memory growth if your code generates many unique operation names. 10000 is typically more than enough for any ML workload.

MeasureMemory

Alias for TrackAllocations for more intuitive access.

public bool MeasureMemory { get; set; }

Property Value

bool

ReservoirSize

Gets or sets the reservoir size for percentile estimation (default: 1000).

public int ReservoirSize { get; set; }

Property Value

int

Remarks

For Beginners: How many samples to keep for calculating P50/P95/P99. Larger values give more accurate percentiles but use more memory. 1000 samples is usually sufficient for accurate estimates.

SamplingRate

Gets or sets the sampling rate for high-frequency operations (default: 1.0 = 100%).

public double SamplingRate { get; set; }

Property Value

double

Remarks

For Beginners: What percentage of operations to profile (0.0 to 1.0). 1.0 = profile everything, 0.1 = profile 10% of operations. Use lower values for very hot paths to reduce overhead.

TraceExecution

Alias for TrackCallHierarchy for more intuitive access.

public bool TraceExecution { get; set; }

Property Value

bool

TrackAllocations

Gets or sets whether to track memory allocations (default: true).

public bool TrackAllocations { get; set; }

Property Value

bool

Remarks

For Beginners: Records memory usage for each profiled operation. Helps identify memory-intensive operations during training. Combined with MemoryTracker for detailed analysis.

TrackCallHierarchy

Gets or sets whether to track parent-child call relationships (default: true).

public bool TrackCallHierarchy { get; set; }

Property Value

bool

Remarks

For Beginners: Tracks which operations call which other operations. Enables call tree analysis to see the hierarchy of operations. Adds slight overhead but provides valuable context.