Add DynamoRIO tracing support to DCPerf FeedSim

packages/feedsim/install_feedsim.sh

@@ -107,6 +107,22 @@ rsync -a --exclude=src \
     "${FEEDSIM_THIRD_PARTY_SRC}/"
 cd "${FEEDSIM_THIRD_PARTY_SRC}"
 
+# Optionally build DynamoRIO for tracing support.
+DR_TRACE_FLAGS=()
+if [ "${ENABLE_DR_TRACE:-0}" = "1" ]; then
+  msg "[DR_TRACE] Setting up DynamoRIO tracing support..."
+  BUILD_DIR="${FEEDSIM_THIRD_PARTY_SRC}"
+  export BUILD_DIR
+  # shellcheck disable=SC1091
+  source "${BENCHPRESS_ROOT}/packages/common/dr_trace/install_dynamorio.sh"
+  DR_TRACE_FLAGS=(
+    -DENABLE_DR_TRACE=ON
+    -DDR_INSTALL="${DR_INSTALL}"
+    -DDR_TRACE_DIR="${BENCHPRESS_ROOT}/packages/common/dr_trace"
+  )
+  msg "[DR_TRACE] DR_INSTALL=${DR_INSTALL}"
+fi
+
 # Installing cmake-4.0.3
 
 if ! [ -d "cmake-4.0.3" ]; then
@@ -380,6 +396,7 @@ cmake -G Ninja \
     -DFEEDSIM_USE_DLRM=ON \
     -DTorch_DIR="${FEEDSIM_THIRD_PARTY_SRC}/libtorch/share/cmake/Torch" \
     -DCMAKE_PREFIX_PATH="${FEEDSIM_THIRD_PARTY_SRC}/libtorch" \
+    "${DR_TRACE_FLAGS[@]}" \
     ../
 
 # Dependencies (fmt, folly, fbthrift, etc.) are already installed above via

packages/feedsim/install_feedsim_aarch64.sh

@@ -90,6 +90,22 @@ cp -f "${BENCHPRESS_ROOT}/packages/feedsim/third_party/src/third_party/CMakeList
     "${FEEDSIM_ROOT_SRC}/src/third_party/CMakeLists.txt"
 cd "${FEEDSIM_THIRD_PARTY_SRC}"
 
+# Optionally build DynamoRIO for tracing support.
+DR_TRACE_FLAGS=()
+if [ "${ENABLE_DR_TRACE:-0}" = "1" ]; then
+  msg "[DR_TRACE] Setting up DynamoRIO tracing support..."
+  BUILD_DIR="${FEEDSIM_THIRD_PARTY_SRC}"
+  export BUILD_DIR
+  # shellcheck disable=SC1091
+  source "${BENCHPRESS_ROOT}/packages/common/dr_trace/install_dynamorio.sh"
+  DR_TRACE_FLAGS=(
+    -DENABLE_DR_TRACE=ON
+    -DDR_INSTALL="${DR_INSTALL}"
+    -DDR_TRACE_DIR="${BENCHPRESS_ROOT}/packages/common/dr_trace"
+  )
+  msg "[DR_TRACE] DR_INSTALL=${DR_INSTALL}"
+fi
+
 DEP_CMAKE_VERSION="4.0.3"
 # Installing cmake
 if ! [ -d "cmake-${DEP_CMAKE_VERSION}-linux-aarch64" ]; then
@@ -415,6 +431,7 @@ cmake -G Ninja \
     -DCMAKE_EXE_LINKER_FLAGS_RELEASE="$FS_LDFLAGS" \
     -DFEEDSIM_USE_DLRM=ON \
     -DTorch_DIR="${FEEDSIM_THIRD_PARTY_SRC}/libtorch/share/cmake/Torch" \
+    "${DR_TRACE_FLAGS[@]}" \
     ../
 
 # Third-party deps (fmt, folly, fizz, wangle, mvfst, fbthrift) are built by this

packages/feedsim/third_party/src/workloads/ranking/CMakeLists.txt

@@ -404,6 +404,27 @@ if(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
     target_link_options(LeafNodeRank PRIVATE -Wl,--no-relax)
 endif()
 
+# DynamoRIO tracing support
+option(ENABLE_DR_TRACE "Enable DynamoRIO tracing hooks" OFF)
+if(ENABLE_DR_TRACE)
+    if(NOT DR_TRACE_DIR)
+        message(FATAL_ERROR "ENABLE_DR_TRACE=ON requires -DDR_TRACE_DIR=<path to packages/common/dr_trace>")
+    endif()
+    add_subdirectory(${DR_TRACE_DIR} ${CMAKE_BINARY_DIR}/dr_trace)
+    target_link_libraries(LeafNodeRank PRIVATE dr_trace)
+    if(NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        # Without this the linker emits one RWE PT_LOAD, DR records LeafNodeRank
+        # as a single 465MB module, and raw2trace segfaults decoding the gap
+        # between segments. Scoped to the traced build so the default binary is
+        # unchanged.
+        target_link_options(LeafNodeRank PRIVATE -Wl,-z,separate-code)
+    endif()
+    # drsyms stubs must be compiled into the binary rather than libdr_trace.a so
+    # they resolve references from --whole-archive drmemtrace_drstatic before the
+    # linker finishes.
+    target_sources(LeafNodeRank PRIVATE ${DR_TRACE_DIR}/drsyms_stubs.c)
+endif()
+
 # Re-link libraries at the end to resolve circular static library dependencies
 # (folly/glog use gflags DEFINE_* macros, fbthrift has internal cross-lib refs)
 target_link_libraries(LeafNodeRank PRIVATE

packages/feedsim/third_party/src/workloads/ranking/LeafNodeRank.cc

@@ -73,6 +73,42 @@
 #include "dwarfs/dlrm.h"
 #endif
 
+#ifdef DR_TRACE_INCLUDED
+#include <cstdlib>
+#include <cstring>
+#include "dr_trace.h"
+static constexpr uint32_t kDrTraceDelaySeconds = 15;
+static constexpr uint32_t kDrTraceDurationSeconds = 10;
+
+static constexpr int kFeedsimBasePort = 21212;
+
+static bool dr_trace_enabled_for_port(int port) {
+  const char* only_port = std::getenv("DR_TRACE_PORT");
+  if (only_port != nullptr) {
+    return std::strcmp(only_port, "all") == 0 || std::atoi(only_port) == port;
+  }
+  return std::getenv("IS_AUTOSCALE_RUN") == nullptr || port == kFeedsimBasePort;
+}
+
+// Force -raw_compress none for LeafNodeRank only.
+static void dr_trace_disable_raw_compression() {
+  if (std::getenv("DYNAMORIO_OPTIONS") != nullptr) {
+    return;
+  }
+  const DrTraceConfig defaults;
+  const char* outdir = std::getenv("DR_TRACE_OUTDIR");
+  const char* verbose = std::getenv("DR_TRACE_VERBOSE");
+  const std::string opts = std::string(
+                               "-rstats_to_stderr -no_hook_vsyscall"
+                               " -disable_traces -no_enable_reset"
+                               " -client_lib ';;-offline -raw_compress none"
+                               " -verbose ") +
+      (verbose != nullptr ? verbose : std::to_string(defaults.verbose)) +
+      " -outdir " + (outdir != nullptr ? outdir : defaults.outdir) + "'";
+  setenv("DYNAMORIO_OPTIONS", opts.c_str(), 1);
+}
+#endif
+
 #include "if/gen-cpp2/ranking_types.h"
 
 #include "../search/ICacheBuster.h"
@@ -2958,6 +2994,16 @@ int main(int argc, char** argv) {
   });
   debug_dump_thread.detach();
 
+#ifdef DR_TRACE_INCLUDED
+  if (dr_trace_enabled_for_port(args.port_arg)) {
+    dr_trace_disable_raw_compression();
+    trace_configure_env();
+    std::thread(
+        trace_execution_delay<kDrTraceDelaySeconds, kDrTraceDurationSeconds>)
+        .detach();
+  }
+#endif
+
   server.run();
 
   // One last dump after the server returns so the final state hits the