ReadyToRun y compilación por niveles

JIT, R2R y compilación por niveles

Cómo funciona el JIT

// What happens at runtime:
// 1. App starts — methods are stubs pointing to the JIT
// 2. First call to MyMethod() → JIT compiles IL → native x64 code
// 3. Subsequent calls → native code runs directly (no JIT overhead)

// Startup cost: each method's first call takes microseconds for JIT
// Peak performance: excellent — JIT knows the actual CPU features

// .NET 9 JIT optimizations applied automatically:
// - Inlining
// - Loop unrolling
// - Escape analysis
// - SIMD vectorization
// - Profile-Guided Optimization (PGO)

Compilación ReadyToRun (R2R)

// Enable R2R in .csproj:
<PropertyGroup>
  <PublishReadyToRun>true</PublishReadyToRun>

  <!-- Optional: composite R2R — whole app in one native image -->
  <PublishReadyToRunComposite>true</PublishReadyToRunComposite>
</PropertyGroup>

// Publish:
dotnet publish -c Release -r linux-x64 --self-contained

// Benefits:
// - Startup time improvement: ~20-30% faster than JIT-only
// - No runtime required (self-contained)

// Trade-off:
// - Larger binary (IL + R2R code both stored)
// - Publish takes longer (cross-gen2 runs at publish time)

R2R frente a Native AOT

// ReadyToRun:
// ✓ Still ships IL (fallback JIT for dynamically-generated code)
// ✓ Supports full reflection
// ✓ .NET runtime required on target
// ✓ Startup: ~30% faster than JIT
// Size: medium (IL + native)

// Native AOT:
// ✓ No IL, no runtime needed on target
// ✓ Startup: ~36x faster than JIT
// ✗ No reflection (must use source generators)
// ✗ Trimming limitations
// Size: smallest

// JIT only:
// ✓ Full reflection, no limitations
// ✓ Best peak throughput via PGO
// ✗ Slowest startup

Compilación por niveles

// Tiered compilation lifecycle:
// Tier 0: fast JIT with minimal optimization
//   → method is callable in ~100 microseconds
// After ~30 calls (configurable): marked as "hot"
// Tier 1: full JIT with all optimizations + PGO
//   → background thread re-compiles the method
//   → future calls use optimized native code

// Usually transparent — you don't need to configure it
// Disable (rarely needed, only for debugging JIT issues):
// DOTNET_TieredCompilation=0

// Disable tiered PGO if benchmarking:
// DOTNET_TieredPGO=0

Optimización guiada por perfiles (PGO)

// PGO automatically specializes code like:

// Before PGO (generic virtual dispatch):
void Process(IAnimal animal) => animal.Speak();
// → runtime checks type on every call

// After PGO (type-specialized inline):
void Process(IAnimal animal)
{
    if (animal is Dog)     { /* inlined Dog.Speak() */ }
    else if (animal is Cat){ /* inlined Cat.Speak() */ }
    else animal.Speak();   // slow path
}

// Enabled by default in .NET 8+
// Verify:
// DOTNET_ReadyToRun=0
// DOTNET_TieredCompilation=1  (default)
// DOTNET_TieredPGO=1          (default .NET 8+)

Hooks de inicio

// Create a hook class:
public class StartupHook
{
    public static void Initialize()
    {
        Console.WriteLine("[StartupHook] Initializing before Main");
        // e.g., register telemetry, inject config overrides
        Environment.SetEnvironmentVariable("FEATURE_FLAG", "1");
    }
}

// Compile to a DLL and set the env var:
export DOTNET_STARTUP_HOOKS=/path/to/MyHook.dll

// Then run your app normally:
dotnet run
// [StartupHook] Initializing before Main
// [Your App Output]

Opciones de configuración del runtime

// runtimeconfig.template.json (baked into app at publish):
{
  "configProperties": {
    "System.GC.Server": true,           // server GC (multi-core)
    "System.GC.Concurrent": true,       // background GC
    "System.GC.HeapHardLimit": 536870912, // 512 MB hard limit
    "System.Threading.ThreadPool.MinThreads": 20,
    "System.Threading.ThreadPool.MaxThreads": 200,
    "System.Runtime.TieredCompilation": true
  }
}

// Environment variable overrides (useful in containers):
// DOTNET_GCHeapHardLimit=536870912
// DOTNET_ThreadPool_UnfairSemaphoreSpinLimit=70
// DOTNET_GCConserve=1  // reduce memory at cost of throughput

Crossgen2 e imágenes compuestas

// Composite R2R: compile app + framework into one native image
<PropertyGroup>
  <PublishReadyToRun>true</PublishReadyToRun>
  <PublishReadyToRunComposite>true</PublishReadyToRunComposite>
</PropertyGroup>

// Normal R2R includes multiple images per assembly
// Composite merges them → fewer file reads at startup

// Manual crossgen2 invocation:
dotnet tool install -g dotnet-crossgen2
crossgen2 \
  --composite \
  --out composite.dll \
  MyApp.dll framework/System.Runtime.dll

// Inspect R2R content:
dotnet-ildasm --methodheader MyApp.dll | grep --color R2R

Caso real: ajustar el inicio de ASP.NET Core

// .csproj:
<PublishReadyToRun>true</PublishReadyToRun>  // pre-JIT at publish
<PublishReadyToRunComposite>true</PublishReadyToRunComposite>
<TieredCompilation>true</TieredCompilation>   // default

// Program.cs — avoid startup work:
// ✓ Use AddHealthChecks() NOT AddDbContextCheck() for faster probe
// ✓ Lazy-initialize expensive singletons
// ✓ Don't await long tasks in Program.cs before app.Run()

// Container settings:
// DOTNET_GCServer=1             → server GC for multi-core pods
// DOTNET_GCHeapHardLimit=536870912 → prevent OOM in constrained pods
// DOTNET_ThreadPool_MinThreads=20  → reduce thread starvation under burst load

// Measure startup:
// app.MapGet("/startup-time",
//   () => $"Ready in: {Stopwatch.GetElapsedTime(startTimestamp).TotalMs}ms");

Comprobación rápida

Recapitulación: ReadyToRun y compilación por niveles