Fix AdSim getting stuck instead of serving requests

packages/adsim/build-deps.sh

@@ -15,7 +15,7 @@ install_packages() {
         # Note: fast_float and boost are built from source since system packages are too old
         # On CentOS Stream 10+, gcc-toolset-14 doesn't exist (GCC 14 is default);
         # use libatomic-devel directly. On older systems, try gcc-toolset-14-libatomic-devel.
-        sudo dnf install -y clang jemalloc-devel xxhash-devel bzip2-devel libomp-devel gengetopt python3-devel gtest-devel \
+        sudo dnf install -y cmake bison flex openssl-devel numactl-devel clang jemalloc-devel xxhash-devel bzip2-devel libomp-devel gengetopt python3-devel gtest-devel \
             double-conversion double-conversion-devel libsodium-devel \
             gflags-devel glog-devel libunwind-devel libevent-devel lz4-devel libzstd-devel snappy-devel xz-devel binutils-devel
         # Install libatomic — try gcc-toolset first (RHEL 9), fall back to base package (CentOS Stream 10+)
@@ -24,7 +24,7 @@ install_packages() {
         # Ubuntu/Debian systems - map package names to Ubuntu equivalents
         # Note: fast_float and boost are built from source since system packages are too old
         sudo apt-get update
-        sudo apt-get install -y clang libjemalloc-dev libxxhash-dev libbz2-dev libomp-dev gengetopt libatomic1 python3-dev libgtest-dev \
+        sudo apt-get install -y cmake bison flex libssl-dev libnuma-dev clang libjemalloc-dev libxxhash-dev libbz2-dev libomp-dev gengetopt libatomic1 python3-dev libgtest-dev \
             libdouble-conversion-dev libsodium-dev \
             libgflags-dev libgoogle-glog-dev libunwind-dev libevent-dev liblz4-dev libzstd-dev libsnappy-dev liblzma-dev libiberty-dev
     else
@@ -54,12 +54,12 @@ export CXX="${ADSIM_CXX_COMPILER}"
 # Pin Facebook library versions for reproducible builds
 # Note: Facebook OSS libraries (folly, fizz, wangle, mvfst, fbthrift, fb303) are released
 # together with matching version tags to ensure compatibility
-FOLLY_VERSION=v2025.06.23.00
-FIZZ_VERSION=v2025.06.23.00
-WANGLE_VERSION=v2025.06.23.00
-MVFST_VERSION=v2025.06.23.00
-FBTHRIFT_VERSION=v2025.06.23.00
-FB303_VERSION=v2025.06.23.00
+FOLLY_VERSION=v2026.08.03.00
+FIZZ_VERSION=v2026.08.03.00
+WANGLE_VERSION=v2026.08.03.00
+MVFST_VERSION=v2026.08.03.00
+FBTHRIFT_VERSION=v2026.08.03.00
+FB303_VERSION=v2026.08.03.00
 
 # fast_float v8.0.0+ is required for folly's allow_leading_plus feature
 FAST_FLOAT_VERSION=v8.0.0

packages/adsim/install_adsim.sh

@@ -51,8 +51,15 @@ setup_directories() {
   # Create final benchmark output directory
   mkdir -p ${BENCHMARKS_DIR}
 
-  # Clean and recreate temporary build directory
-  rm -rf ${BUILD_DIR}
+  # Clean and recreate temporary build directory.
+  # SKIP_CLEAN=1 preserves an existing build dir so that deps/ and staging/ can
+  # be reused across iterations; otherwise every install rebuilds the whole
+  # dependency stack from scratch.
+  if [ "${SKIP_CLEAN}" = "1" ]; then
+    echo "[SETUP] SKIP_CLEAN=1 - preserving existing ${BUILD_DIR}"
+  else
+    rm -rf ${BUILD_DIR}
+  fi
   mkdir -p ${BUILD_DIR}
 
   # Enter build directory for subsequent operations

packages/adsim/install_fbgemm.sh

@@ -16,10 +16,10 @@
 MINICONDA_PREFIX=${FBGEMM_STAGING_DIR}/miniconda
 
 # Version of FBGEMM to install
-FBGEMM_VERSION=v1.4.0
+FBGEMM_VERSION=v1.8.0
 
 # Version of PyTorch to install
-PYTORCH_VERSION=2.8.0
+PYTORCH_VERSION=2.13.0
 
 MINICONDA_VERSION="5.1-0"
 
@@ -809,7 +809,7 @@ setup_directories() {
   # Change the current directory to the build directory
   # pushd saves the current directory on a stack so we can return to it later with popd
   echo "[SETUP] Changing to build directory..."
-  pushd ${FBGEMM_STAGING_DIR} || exist
+  pushd ${FBGEMM_STAGING_DIR} || exit 1
 
   echo "[SETUP] Directory setup complete."
 }
@@ -906,7 +906,7 @@ clone_fbgemm_repo() {
   echo "#!/bin/bash" > fbgemm/.github/scripts/fbgemm_gpu_postbuild.bash
 
   # Change the current directory to the FBGEMM GPU directory
-  pushd fbgemm/fbgemm_gpu || exist
+  pushd fbgemm/fbgemm_gpu || exit 1
 
   echo "[SETUP] FBGEMM repository setup complete."
 }
@@ -962,13 +962,13 @@ install_fbgemm() {
   echo "[BUILD] Building and installing FBGEMM..."
   # shellcheck disable=SC2086
   # print_exec conda run -n $BUILD_ENV python setup.py ${run_multicore} install --package_variant=cpu
-  popd || exist
-  pushd fbgemm || exist
+  popd || exit 1
+  pushd fbgemm || exit 1
   mkdir build
   cp ${ADSIM_PROJ_ROOT}/buildfiles/fbgemm/cmake.txt build
-  pushd build || exist
+  pushd build || exit 1
   source cmake.txt
-  popd || exist
+  popd || exit 1
 
   # Test if the FBGEMM library can be imported in the Conda environment
   # Generate a standalone executable for the project
@@ -1006,7 +1006,7 @@ cleanup() {
 
   # This restores the directory we were in before entering the build directory
   echo "[CLEANUP] Returning to original directory..."
-  popd || exist
+  popd || exit 1
 
   # Remove the build directory to clean up after the build process
   # This saves disk space by removing intermediate build files

packages/adsim/patches/treadmill.patch

@@ -7416,7 +7416,7 @@ diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclud
 diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclude=gen-cpp2' '--exclude=build' '--exclude=third_party' treadmill-src/src/Treadmill.cpp treadmill/src/Treadmill.cpp
 --- treadmill-src/src/Treadmill.cpp	1969-12-31 16:00:00.000000000 -0800
 +++ treadmill/src/Treadmill.cpp	2025-08-04 15:44:21.787727598 -0700
-@@ -0,0 +1,129 @@
+@@ -0,0 +1,131 @@
 +/*
 + *  Copyright (c) 2014, Facebook, Inc.
 + *  All rights reserved.
@@ -7540,7 +7540,9 @@ diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclud
 +  // Set the usage information
 +  std::string usage("Treadmill loadtester");
 +  google::SetUsageMessage(usage);
-+  const folly::Init init(&argc, &argv);
++  // Must outlive init(): ~folly::Init calls SingletonVault::destroyInstances(),
++  // so a plain local would tear the vault down before run() uses any singleton.
++  static const folly::Init init(&argc, &argv);
 +}
 +
 +} // namespace treadmill

packages/adsim/qps_search.sh

@@ -99,7 +99,7 @@ function parse_opts()
 }
 
 function log() {
-  if [ true == $VERBOSE_LOGGING ]; then
+  if [ "true" = "$VERBOSE_LOGGING" ]; then
     echo "$1"
   fi
 }
@@ -119,7 +119,7 @@ function measure_latency()
     log "QPS is 0, returning large latency value to avoid blocking"
     LATENCY=999999999
     ACHIEVED_QPS=0
-    if [ true == $CSV_LOGGING ]; then
+    if [ "true" = "$CSV_LOGGING" ]; then
       requested_qps="$1"
       sequence="$2"
       runtime="$3"
@@ -161,7 +161,7 @@ function measure_latency()
   ACHIEVED_QPS=$(cat "$THROUGHPUT_LOG")
   log "Latency $TGT_QUANTILE = $LATENCY; Achieved QPS = $ACHIEVED_QPS"
 
-  if [ true == $CSV_LOGGING ]; then
+  if [ "true" = "$CSV_LOGGING" ]; then
     requested_qps="$1"
     sequence="$2"
     runtime="$3"
@@ -173,6 +173,7 @@ function measure_latency()
 
 function coarse_search()
 {
+  local start_qps=1
   log "Starting coarse search to find maximum QPS..."
 
   # Find largest power of 2 <= NWORKERS
@@ -241,7 +242,7 @@ function binary_search()
 
   while [ "$qps_l" -lt $(( qps_r - 1 )) ]; do
     qps_m=$(( (qps_l + qps_r) / 2 ))
-    measure_latency $"$qps_m" "$SEQ" "$RUNTIME"
+    measure_latency "$qps_m" "$SEQ" "$RUNTIME"
     SEQ=$((SEQ + 1))
     lat=${LATENCY%.*}
 
@@ -262,12 +263,12 @@ function binary_search()
     qps_opt=$qps_r
     lat_opt=$lat_r
   fi
-  if [ true == "$CSV_LOGGING" ]; then
+  if [ "true" = "$CSV_LOGGING" ]; then
     final_qps=$(cat "$THROUGHPUT_LOG")
     echo "${SEQ},${qps_opt},${final_qps},${TGT_QUANTILE},${TGT_LATENCY},${lat_opt}"
   fi
 
-  measure_latency $"$qps_opt" "$SEQ" "$RUNTIME"
+  measure_latency "$qps_opt" "$SEQ" "$RUNTIME"
 
   log "Optimal QPS = $qps_opt, achieving latency = $lat_opt"
   log "$TREADMILL_EXE --hostname $ADSIM_HOST --port $ADSIM_PORT " \
@@ -276,7 +277,7 @@ function binary_search()
 }
 
 parse_opts "$@"
-if [ true == "$CSV_LOGGING" ]; then
+if [ "true" = "$CSV_LOGGING" ]; then
     echo "seq,requested_qps,achieved_qps,target_latency_quantile,target_latency_usec,achieved_latency_usec"
 fi
 # If min and max QPS are not manually specified, run coarse search to find them

packages/adsim/run_adsim.py

@@ -230,7 +230,7 @@ def run_cmd(cmd: List[str], timeout=None, dryrun=False, verbose=False) -> str:
         print(" ".join(cmd))
     if dryrun:
         return None
-    if timeout <= 0:
+    if timeout is not None and timeout <= 0:
         timeout = None
 
     proc = subprocess.Popen(cmd)

packages/adsim/src/cpp2/server/dwarfs/CMakeLists.txt

@@ -48,14 +48,14 @@ add_dependencies(gemm fbgemm)
 add_library(embedding Embedding.cc Embedding.h)
 
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
-    target_compile_options(gemm PRIVATE
+    target_compile_options(embedding PRIVATE
         ${COROUTINES_FLAG}
         -m64
         -mavx2
         -mfma
         -masm=intel)
 elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
-    target_compile_options(gemm PRIVATE
+    target_compile_options(embedding PRIVATE
         ${COROUTINES_FLAG}
         -march=armv8.5-a+sve2
         -mcpu=native)