Remove palloc

CMakeLists.txt

@@ -7,10 +7,6 @@ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 include_directories(include third_party)
 
-set(PA_OVERRIDE ON)
-set(PA_BUILD_TESTS OFF CACHE BOOL "Disable palloc tests for embedded pomai_cache" FORCE)
-add_subdirectory(third_party/palloc)
-
 add_library(pomaicache SHARED
   src/engine/engine.cpp
   src/engine/ssd_store.cpp
@@ -28,8 +24,6 @@ add_library(pomaicache SHARED
   src/bindings/c_api.cc
 )
 
-target_link_libraries(pomaicache PUBLIC palloc)
-
 include(CTest)
 if(BUILD_TESTING)
   add_library(mini_catch_main third_party/catch2/catch_main.cpp)
@@ -40,8 +34,12 @@ if(BUILD_TESTING)
   add_executable(test_ai_cache tests/test_ai_cache.cpp)
   target_link_libraries(test_ai_cache PRIVATE pomaicache mini_catch_main)
 
+  add_executable(test_prompt_cache tests/test_prompt_cache.cpp)
+  target_link_libraries(test_prompt_cache PRIVATE pomaicache mini_catch_main)
+
   add_test(NAME test_engine COMMAND test_engine)
   add_test(NAME test_ai_cache COMMAND test_ai_cache)
+  add_test(NAME test_prompt_cache COMMAND test_prompt_cache)
 
   if(NOT WIN32)
     add_executable(test_integration tests/test_integration.cpp)

ai_bench_summary.json

@@ -0,0 +1,11 @@
+{
+  "workloads": [
+    {"name":"embedding_zipf","ops_s":459.52,"p50_us":130.28,"p95_us":11175.92,"p99_us":12889.39,"p999_us":14281.39,"hit_rate":0.82},
+    {"name":"prompt_response_churn","ops_s":155.11,"p50_us":842.28,"p95_us":21994.37,"p99_us":25153.57,"p999_us":45459.81,"hit_rate":0.73},
+    {"name":"rerank_ttl_storm","ops_s":323.85,"p50_us":549.67,"p95_us":10760.80,"p99_us":11009.46,"p999_us":11728.75,"hit_rate":0.75},
+    {"name":"mixed_rag_pipeline","ops_s":51.31,"p50_us":8994.29,"p95_us":61316.55,"p99_us":69901.89,"p999_us":113447.43,"hit_rate":0.74}
+  ],
+  "ssd_mb_s": 0.0,
+  "engine_ctor_ms": 0.18,
+  "dedup_ratio": 0.0
+}

bench/prompt_cache_bench.cpp

@@ -213,9 +213,15 @@ run_prompt_workload_embedded(const std::string &name,
 }
 
 int main(int argc, char **argv) {
+  bool quick = false;
   std::string json_out = "prompt_cache_bench.json";
-  if (argc > 1)
-    json_out = argv[1];
+  for (int i = 1; i < argc; ++i) {
+    if (std::string(argv[i]) == "--quick")
+      quick = true;
+    else
+      json_out = argv[i];
+  }
+  const int ops_per_workload = quick ? 500 : 5000;
 
   EngineConfig cfg;
   cfg.memory_limit_bytes = 64 * 1024 * 1024;
@@ -234,19 +240,22 @@ int main(int argc, char **argv) {
   PromptCacheManager pcm(engine, ai, pcfg);
 
   std::cout << std::fixed << std::setprecision(2);
-  std::cout << "Embedded token/prompt cache benchmark (in-process, no network)\n";
+  std::cout << "Embedded token/prompt cache benchmark (in-process, no network)";
+  if (quick)
+    std::cout << " [--quick]";
+  std::cout << "\n";
 
   std::vector<PromptBenchResult> results;
   results.push_back(run_prompt_workload("chatty_short_sessions",
                                         pcm,
-                                        5000,
+                                        ops_per_workload,
                                         200,   // hot prefixes
                                         128,   // max tokens per prompt
                                         0.30   // writes
                                         ));
   results.push_back(run_prompt_workload("long_lived_system_prompts",
                                         pcm,
-                                        5000,
+                                        ops_per_workload,
                                         50,    // fewer prefixes, more reuse
                                         256,   // max tokens per prompt
                                         0.10   // mostly reads
@@ -259,7 +268,7 @@ int main(int argc, char **argv) {
   pomaicache::PomaiCache cache(embedded_cfg);
   results.push_back(run_prompt_workload_embedded("embedded_api_hot_prompts",
                                                  cache,
-                                                 5000,
+                                                 ops_per_workload,
                                                  100,
                                                  256,
                                                  0.25));

prompt_cache_bench.json

@@ -0,0 +1,7 @@
+{
+  "prompt_cache_workloads": [
+    {"name":"chatty_short_sessions","ops_s":142.875,"p50_us":5932.13,"p95_us":13378.5,"p99_us":20634.9,"hit_rate":0.143678,"avg_savings_ratio":0.00735079},
+    {"name":"long_lived_system_prompts","ops_s":67.667,"p50_us":14609.5,"p95_us":17571,"p99_us":24018.2,"hit_rate":0.569869,"avg_savings_ratio":0.0215384},
+    {"name":"embedded_api_hot_prompts","ops_s":137.045,"p50_us":8095.73,"p95_us":13802.2,"p99_us":14491.8,"hit_rate":0,"avg_savings_ratio":0}
+  ]
+}

scripts/run_tests_and_benches.sh

@@ -0,0 +1,25 @@
+#!/usr/bin/env bash
+# Run all unit tests and embedded benchmarks (no server required).
+# Use prompt_cache_bench --quick so the full run completes in reasonable time.
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+BUILD_DIR="${BUILD_DIR:-${ROOT}/build}"
+mkdir -p "${BUILD_DIR}"
+
+cmake -S "${ROOT}" -B "${BUILD_DIR}" -DBUILD_TESTING=ON -DBUILD_PYTHON_BINDINGS=OFF
+cmake --build "${BUILD_DIR}" -j
+
+echo "========== Running tests =========="
+ctest --test-dir "${BUILD_DIR}" --output-on-failure
+
+echo ""
+echo "========== Running benchmarks =========="
+"${BUILD_DIR}/pomai_cache_bench"
+"${BUILD_DIR}/ai_artifact_bench"
+"${BUILD_DIR}/vector_cache_bench"
+"${BUILD_DIR}/prompt_cache_bench" --quick
+"${BUILD_DIR}/pomai_cache_crash_harness"
+
+echo ""
+echo "All tests and benchmarks completed successfully."

tests/crash_harness.cpp

@@ -2,6 +2,7 @@
 #include <csignal>
 #include <cstdlib>
 #include <cstring>
+#include <iostream>
 #include <netinet/in.h>
 #include <optional>
 #include <string>
@@ -86,6 +87,10 @@ pid_t spawn_server(int port, const std::string &dir, const std::string &fsync) {
 } // namespace
 
 int main(int argc, char **argv) {
+  if (access("./pomai_cache_server", X_OK) != 0) {
+    std::cout << "SKIP: pomai_cache_server not found (build the server to run crash harness)\n";
+    return 0;
+  }
   std::string fsync = "everysec";
   if (argc > 1)
     fsync = argv[1];

vector_bench_results.json

@@ -0,0 +1,18 @@
+{
+  "vector_benchmarks": [
+    {"label":"dim128-1K","dim":128,"dataset_size":1000,"insert_ops_s":51700.9,"search_ops_s":4701.12,"search_p50_us":204.181,"search_p99_us":318.114,"memory_mb":0.586435,"bytes_per_vector":614.922,"recall_at_k":0.346,"sim_hit_rate":0.346,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim128-10K","dim":128,"dataset_size":10000,"insert_ops_s":5217.21,"search_ops_s":410.475,"search_p50_us":2420.76,"search_p99_us":2793.61,"memory_mb":5.87361,"bytes_per_vector":615.892,"recall_at_k":0.352,"sim_hit_rate":0.352,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim384-1K","dim":384,"dataset_size":1000,"insert_ops_s":52397.8,"search_ops_s":1679.02,"search_p50_us":593.619,"search_p99_us":753.053,"memory_mb":1.563,"bytes_per_vector":1638.92,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim384-10K","dim":384,"dataset_size":10000,"insert_ops_s":5606.39,"search_ops_s":152.023,"search_p50_us":6556.22,"search_p99_us":7069.5,"memory_mb":15.6392,"bytes_per_vector":1639.89,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim768-1K","dim":768,"dataset_size":1000,"insert_ops_s":50942.3,"search_ops_s":841.376,"search_p50_us":1184.97,"search_p99_us":1373.65,"memory_mb":3.02784,"bytes_per_vector":3174.92,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim768-10K","dim":768,"dataset_size":10000,"insert_ops_s":5775.09,"search_ops_s":79.3694,"search_p50_us":12559.8,"search_p99_us":13241.6,"memory_mb":30.2877,"bytes_per_vector":3175.89,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim1536-1K","dim":1536,"dataset_size":1000,"insert_ops_s":43567.7,"search_ops_s":408.786,"search_p50_us":2442.17,"search_p99_us":2671.27,"memory_mb":5.95753,"bytes_per_vector":6246.92,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4},
+    {"label":"dim1536-10K","dim":1536,"dataset_size":10000,"insert_ops_s":5547.31,"search_ops_s":39.6758,"search_p50_us":24957.7,"search_p99_us":29559.9,"memory_mb":59.5845,"bytes_per_vector":6247.89,"recall_at_k":0.334,"sim_hit_rate":0.334,"quant_f16_ratio":2,"quant_i8_ratio":4}
+  ],
+  "e2e_comparisons": [
+    {"scenario":"identical_prompts","exact_hit_rate":0,"sim_hit_rate":1,"hit_rate_boost":1,"avg_latency_us":296.367,"dollar_saved":15,"tokens_saved":500000},
+    {"scenario":"slight_rephrase","exact_hit_rate":0,"sim_hit_rate":1,"hit_rate_boost":1,"avg_latency_us":356.408,"dollar_saved":15,"tokens_saved":500000},
+    {"scenario":"moderate_rephrase","exact_hit_rate":0,"sim_hit_rate":1,"hit_rate_boost":1,"avg_latency_us":302.969,"dollar_saved":15,"tokens_saved":500000},
+    {"scenario":"heavy_rephrase","exact_hit_rate":0,"sim_hit_rate":0.878,"hit_rate_boost":0.878,"avg_latency_us":271.672,"dollar_saved":13.17,"tokens_saved":439000}
+  ]
+}