Add networkbench webserver benchmark

.dockerignore

@@ -0,0 +1,48 @@
+# VCS and local editor state
+.git
+.gitignore
+.idea/
+.vscode/
+.DS_Store
+
+# Docker build files are not part of the source tree copied into the image.
+Dockerfile
+.dockerignore
+
+# Python caches and local environments
+__pycache__/
+*.py[cod]
+*.pyo
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+*venv*/
+
+# Native and package build outputs
+scratch/
+build/
+dist/
+*.egg-info/
+cmake-build-*/
+CMakeFiles/
+CMakeCache.txt
+compile_commands.json
+*.o
+*.a
+*.so
+*.dylib
+*.dSYM/
+
+# Profiling, benchmark, and local log artifacts
+*.profraw
+*.profdata
+*.profclangd
+*.fdata
+*.bolt
+perf.data*
+*.log
+
+# Misc
+assets/

Dockerfile

@@ -0,0 +1,151 @@
+FROM ubuntu:24.04
+
+SHELL ["/bin/bash", "-exo", "pipefail", "-c"]
+
+ARG DEBIAN_FRONTEND=noninteractive
+
+WORKDIR /root
+
+ENV CINDER_ROOT=/root/cinder
+ENV CINDERX_ROOT=/root/cinderx
+ENV CINDERX_VENV=/root/cinderx/.venv-mp312
+ENV CINDERX_BUNDLE_ARCHIVE=/root/cinderx/scratch/cinderx-bundled/libcinderx-bundled.a
+ENV NETWORKBENCH_JITLIST=/root/cinderx/cinderx/benchmarks/networkbench/networkbench.jitlist.txt
+ENV LLVM_VERSION=21
+ENV LLVM_ROOT=/usr/lib/llvm-$LLVM_VERSION
+ENV PATH=$LLVM_ROOT/bin:$PATH
+ENV PROFILE_TASK="$CINDERX_ROOT/cinderx/benchmarks/networkbench/run_server_client.py 1000"
+
+# Install the packages that used to be baked into ubuntu/cpython-build-benchmark.
+RUN apt-get update; \
+    apt-get install -yq --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        cmake \
+        gdb \
+        git \
+        gnupg \
+        linux-tools-generic \
+        libbz2-dev \
+        libffi-dev \
+        libgdbm-compat-dev \
+        libgdbm-dev \
+        liblzma-dev \
+        libncurses5-dev \
+        libreadline-dev \
+        libsqlite3-dev \
+        libssl-dev \
+        libzstd-dev \
+        lsb-release \
+        pkg-config \
+        python3 \
+        python3-dev \
+        python3-full \
+        python3-venv \
+        software-properties-common \
+        tk-dev \
+        uuid-dev \
+        wget \
+        xz-utils \
+        zlib1g-dev; \
+    wget -O /tmp/llvm.sh https://apt.llvm.org/llvm.sh; \
+    chmod +x /tmp/llvm.sh; \
+    /tmp/llvm.sh "$LLVM_VERSION" all; \
+    apt-get install -yq --no-install-recommends \
+        "bolt-$LLVM_VERSION" \
+        "libbolt-$LLVM_VERSION-dev"; \
+    LLVM_BOLT="$(command -v llvm-bolt || command -v "llvm-bolt-$LLVM_VERSION")"; \
+    MERGE_FDATA="$(command -v merge-fdata || command -v "merge-fdata-$LLVM_VERSION")"; \
+    LLVM_PROFDATA="$(command -v llvm-profdata || command -v "llvm-profdata-$LLVM_VERSION")"; \
+    PERF="$(command -v perf || find /usr/lib/linux-tools-* -type f -name perf -print -quit)"; \
+    test -n "$PERF"; \
+    ln -sf "$LLVM_BOLT" /usr/local/bin/llvm-bolt; \
+    ln -sf "$MERGE_FDATA" /usr/local/bin/merge-fdata; \
+    ln -sf "$LLVM_PROFDATA" /usr/local/bin/llvm-profdata; \
+    ln -sf "$PERF" /usr/local/bin/perf; \
+    command -v cmake; \
+    command -v perf; \
+    command -v llvm-bolt; \
+    command -v merge-fdata; \
+    command -v llvm-profdata; \
+    command -v llvm-ar; \
+    command -v llvm-ranlib; \
+    llvm-bolt --version; \
+    llvm-profdata --version; \
+    rm -rf /var/lib/apt/lists/* /tmp/llvm.sh
+
+RUN git clone -b meta/3.12 https://github.com/facebookincubator/cinder.git
+
+WORKDIR $CINDER_ROOT
+
+# Build a bootstrap interpreter first. CinderX needs a working meta/3.12 Python
+# to run its own PGO build before we can link those optimized objects into the
+# final CPython executable.
+RUN CC=clang CXX=clang++ ./configure --prefix="$CINDER_ROOT/mp312"
+RUN make -j
+RUN make install
+
+WORKDIR /root
+
+WORKDIR $CINDERX_ROOT
+
+COPY . "$CINDERX_ROOT/"
+
+RUN test -f "$NETWORKBENCH_JITLIST"
+
+RUN "$CINDER_ROOT/mp312/bin/python3" -m venv "$CINDERX_VENV"
+
+RUN "$CINDERX_VENV/bin/python" -m pip install --upgrade pip setuptools
+
+# Install CinderX with its PGO flow enabled. The local checkout supplies the
+# profile-task override, so the PGO workload is networkbench instead of the
+# default CPython test-suite task.
+RUN PYTHONJITLISTFILE="$NETWORKBENCH_JITLIST" \
+    CINDERX_ENABLE_PGO=1 \
+    CINDERX_ENABLE_LTO=1 \
+    CINDERX_PGO_PROFILE_TASK="$PROFILE_TASK" \
+    CC=clang \
+    CXX=clang++ \
+    "$CINDERX_VENV/bin/pip" -vvv install --ignore-requires-python --no-build-isolation --no-clean .
+
+# CMake's setuptools build produces a shared _cinderx extension. For the final
+# benchmark image, fold the PGO/LTO object files from that build into one static
+# archive and link it as a CPython built-in module in the final optimized
+# interpreter executable.
+RUN CINDERX_BUILD_DIR="$(find "$CINDERX_ROOT/scratch" -maxdepth 1 -type d -name 'temp.*' -print -quit)"; \
+    test -n "$CINDERX_BUILD_DIR"; \
+    mkdir -p "$(dirname "$CINDERX_BUNDLE_ARCHIVE")"; \
+    find "$CINDERX_BUILD_DIR" -path '*/CMakeFiles/*.dir/*' -name '*.o' -print0 \
+      | sort -z \
+      | xargs -0 llvm-ar rcs "$CINDERX_BUNDLE_ARCHIVE"; \
+    llvm-ranlib "$CINDERX_BUNDLE_ARCHIVE"; \
+    llvm-ar t "$CINDERX_BUNDLE_ARCHIVE" | grep '_cinderx.cpp.o' > /dev/null
+
+WORKDIR $CINDER_ROOT
+
+RUN make distclean
+
+RUN printf '%s\n' \
+      '*static*' \
+      "_cinderx -Wl,--whole-archive $CINDERX_BUNDLE_ARCHIVE -Wl,--no-whole-archive -lstdc++ -lz -ldl -lpthread -lm" \
+      > Modules/Setup.local; \
+    cat Modules/Setup.local
+
+# The final binary includes CinderX, so keep BOLT's core layout passes while
+# avoiding LLVM 21 AArch64 apply passes that crash or exceed the VM memory limit.
+RUN CC=clang \
+    CXX=clang++ \
+    LDFLAGS="-fuse-ld=lld -flto" \
+    BOLT_APPLY_FLAGS="-skip-funcs=_PyEval_EvalFrameDefault,sre_ucs1_match/1,sre_ucs2_match/1,sre_ucs4_match/1 -reorder-blocks=ext-tsp -reorder-functions=cdsort -split-functions -reorder-functions-use-hot-size -peepholes=none -use-gnu-stack" \
+    ./configure --prefix="$CINDER_ROOT/mp312" --enable-optimizations --enable-bolt
+RUN PYTHONPATH="$CINDERX_ROOT/cinderx/PythonLib" \
+    PYTHONJITLISTFILE="$NETWORKBENCH_JITLIST" \
+    make -j
+RUN make install
+
+WORKDIR $CINDERX_ROOT
+
+RUN find "$CINDERX_VENV" -name '_cinderx*.so' -delete; \
+    "$CINDERX_VENV/bin/python" -c 'import importlib.util; spec = importlib.util.find_spec("_cinderx"); assert spec is not None and spec.origin == "built-in", spec; import cinderx; assert cinderx.is_initialized(), cinderx.get_import_error()'
+
+ENTRYPOINT ["bash", "-c", "source $CINDERX_VENV/bin/activate && python cinderx/benchmarks/networkbench/run_bench.py -n 1"]

cinderx/benchmarks/networkbench/README.md

@@ -0,0 +1,62 @@
+# networkbench
+
+`networkbench` is a small benchmark meant to reproduce the shape of a Python
+webserver. It exercises request parsing, route dispatch, middleware, async I/O,
+filesystem-backed state, response encoding, and CPU work inside request handlers.
+(File I/O through async thread executors substitutes for database query I/O.)
+
+The benchmark runs a local HTTP server on `localhost:8080` and drives it with an
+async client. The client uploads network matrices, then sends concurrent requests
+against two main endpoints:
+
+- `GET /network` reads a stored matrix and returns it.
+- `GET /reachable` decodes a graph payload and computes whether two nodes are
+  reachable.
+
+
+Performance is measured in the client as number of requests/second, after the
+upload phase.
+
+## Running
+
+Run one server/client benchmark pass:
+
+```bash
+cd cinderx/benchmarks/networkbench
+python run_server_client.py 10000
+```
+
+The numeric argument is the number of client requests. The client prints
+`Average requests per second` after the run.
+
+Run the comparison harness:
+
+```bash
+python run_bench.py 10000 -n 5
+```
+
+`run_bench.py` compares:
+
+- `cinderx_jitlist`: runs with `networkbench.jitlist.txt`.
+- `cinderx_disable`: runs with `CINDERX_DISABLE=1`.
+
+## Useful Knobs
+
+These environment variables tune the workload:
+
+- `SERVER_PROCESS_COUNT`: number of worker processes, default `8`.
+- `CLIENT_MAX_INFLIGHT_REQUESTS`: client concurrency limit, default `16`.
+- `NETWORK_MATRIX_COUNT`: matrices uploaded before the timed run, default `16`.
+- `NETWORK_GET_PERCENT`: percentage of timed requests sent to `/network`,
+  default `90`.
+- `NETWORK_STORAGE_DIR`: directory used for stored matrix files.
+
+## JIT List
+
+Regenerate the JIT list from a debug run with:
+
+```bash
+python generate_networkbench_jitlist.py 10000
+```
+
+This rewrites `networkbench.jitlist.txt` by default.