Merge pull request #4 from ffengc/dev

Final paper deliverable
Restructure README, add multi-trial harness and final paper figures.

Author: ffengc

.gitignore

@@ -3,4 +3,5 @@ server/
 *.bak
 docs/
 logs/
-__pycache__/
+__pycache__/
+paper/

Predictor.hpp

@@ -13,6 +13,7 @@
 
 #include <iostream>
 #include <algorithm>
+#include <chrono>
 #include <cmath>
 #include "log.hpp"
 
@@ -89,6 +90,8 @@ class Predictor {
     // Core function: called once per second with the observed RPS.
     // Returns the target total number of warm workers to maintain.
     int UpdateAndPredict(int current_rps) {
+        auto t0 = std::chrono::high_resolution_clock::now();
+
         // Step 1: EWMA — track periodic baseline
         if (_ewma == 0.0) {
             _ewma = current_rps;
@@ -112,6 +115,14 @@ class Predictor {
         // Step 3: Little's Law  N = ceil(lambda * T) + M_safety
         int target = (int)std::ceil(predicted_lambda * _avg_service_time) + _safety_margin;
 
+        // Per-tick inference latency — emitted at NORMAL so it lands in
+        // server.log for paper Section 5.4 (Pareto/overhead) extraction.
+        // Grep handle: [PredLatency] mode=ewma us=...
+        long long us = std::chrono::duration_cast<std::chrono::microseconds>(
+            std::chrono::high_resolution_clock::now() - t0).count();
+        logMessage(NORMAL, "[PredLatency] mode=ewma us=%lld rps=%d target=%d",
+                   us, current_rps, target);
+
         logMessage(DEBUG, "[Predictor] RPS=%d  EWMA=%.2f  T=%.3fs  Target=%d",
                    current_rps, _ewma, _avg_service_time, target);
 
@@ -129,6 +140,8 @@ class Predictor {
     // low. Normalising by a running sigma removes the RPS-magnitude
     // dependence, so C2/C4 fire at the same Ramp window as C1/C3.
     int AdaptivePredict(int current_rps) {
+        auto t0 = std::chrono::high_resolution_clock::now();
+
         // Step 1: EWMA baseline (same as UpdateAndPredict).
         if (_ewma == 0.0) {
             _ewma = current_rps;
@@ -168,6 +181,11 @@ class Predictor {
 
         int target = (int)std::ceil(predicted_lambda * _avg_service_time) + _safety_margin;
 
+        long long us = std::chrono::duration_cast<std::chrono::microseconds>(
+            std::chrono::high_resolution_clock::now() - t0).count();
+        logMessage(NORMAL, "[PredLatency] mode=ewma_adaptive us=%lld rps=%d target=%d",
+                   us, current_rps, target);
+
         logMessage(DEBUG, "[Predictor/adaptive] RPS=%d  EWMA=%.2f  sigma=%.2f  z=%.2f  CUSUM_std=%.2f  T=%.3fs  Target=%d",
                    current_rps, _ewma, sigma_safe, z, _cusum_std, _avg_service_time, target);
 
@@ -179,7 +197,14 @@ class Predictor {
     // Target is computed directly from the current observed RPS via Little's Law.
     // Responds only to what just happened — never pre-warms ahead of demand.
     int ReactivePredict(int current_rps) {
+        auto t0 = std::chrono::high_resolution_clock::now();
         int target = (int)std::ceil(current_rps * _avg_service_time) + _safety_margin;
+
+        long long us = std::chrono::duration_cast<std::chrono::microseconds>(
+            std::chrono::high_resolution_clock::now() - t0).count();
+        logMessage(NORMAL, "[PredLatency] mode=reactive us=%lld rps=%d target=%d",
+                   us, current_rps, target);
+
         logMessage(DEBUG, "[Predictor/Reactive] RPS=%d  T=%.3fs  Target=%d",
                    current_rps, _avg_service_time, target);
         return target;