reanimator/bee
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

Add NVIDIA stress loader selection and DCGM 4 support

Browse Source
This commit is contained in:
Mikhail Chusavitin
2026-03-31 11:15:15 +03:00
parent 20f834aa96
commit 6dee8f3509
31 changed files with 789 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
iso/builder/Dockerfile
View File

@@ -23,6 +23,16 @@ RUN apt-get update -qq && apt-get install -y \
gcc \
make \
perl \
pkg-config \
yasm \
libssl-dev \
zlib1g-dev \
libbz2-dev \
libgmp-dev \
libpcap-dev \
libsqlite3-dev \
libcurl4-openssl-dev \
ocl-icd-opencl-dev \
linux-headers-amd64 \
&& rm -rf /var/lib/apt/lists/*

3
iso/builder/VERSIONS
View File

@@ -8,7 +8,8 @@ NCCL_TESTS_VERSION=2.13.10
NVCC_VERSION=12.8
CUBLAS_VERSION=13.0.2.14-1
CUDA_USERSPACE_VERSION=13.0.96-1
DCGM_VERSION=3.3.9
DCGM_VERSION=4.5.2-1
JOHN_JUMBO_COMMIT=67fcf9fe5a
ROCM_VERSION=6.3.4
ROCM_SMI_VERSION=7.4.0.60304-76~22.04
ROCM_BANDWIDTH_TEST_VERSION=1.4.0.60304-76~22.04

73
iso/builder/bee-gpu-stress.c
View File

@@ -29,6 +29,7 @@ typedef void *CUfunction;
typedef void *CUstream;
#define CU_SUCCESS 0
#define CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT 16
#define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR 75
#define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR 76
@@ -97,6 +98,9 @@ typedef CUresult (*cuLaunchKernel_fn)(CUfunction,
CUstream,
void **,
void **);
typedef CUresult (*cuMemGetInfo_fn)(size_t *, size_t *);
typedef CUresult (*cuStreamCreate_fn)(CUstream *, unsigned int);
typedef CUresult (*cuStreamDestroy_fn)(CUstream);
typedef CUresult (*cuGetErrorName_fn)(CUresult, const char **);
typedef CUresult (*cuGetErrorString_fn)(CUresult, const char **);
@@ -118,6 +122,9 @@ struct cuda_api {
cuModuleLoadDataEx_fn cuModuleLoadDataEx;
cuModuleGetFunction_fn cuModuleGetFunction;
cuLaunchKernel_fn cuLaunchKernel;
cuMemGetInfo_fn cuMemGetInfo;
cuStreamCreate_fn cuStreamCreate;
cuStreamDestroy_fn cuStreamDestroy;
cuGetErrorName_fn cuGetErrorName;
cuGetErrorString_fn cuGetErrorString;
};
@@ -128,9 +135,10 @@ struct stress_report {
int cc_major;
int cc_minor;
int buffer_mb;
int stream_count;
unsigned long iterations;
uint64_t checksum;
char details[1024];
char details[16384];
};
static int load_symbol(void *lib, const char *name, void **out) {
@@ -144,7 +152,7 @@ static int load_cuda(struct cuda_api *api) {
if (!api->lib) {
return 0;
}
return
if (!(
load_symbol(api->lib, "cuInit", (void **)&api->cuInit) &&
load_symbol(api->lib, "cuDeviceGetCount", (void **)&api->cuDeviceGetCount) &&
load_symbol(api->lib, "cuDeviceGet", (void **)&api->cuDeviceGet) &&
@@ -160,7 +168,17 @@ static int load_cuda(struct cuda_api *api) {
load_symbol(api->lib, "cuMemcpyDtoH_v2", (void **)&api->cuMemcpyDtoH) &&
load_symbol(api->lib, "cuModuleLoadDataEx", (void **)&api->cuModuleLoadDataEx) &&
load_symbol(api->lib, "cuModuleGetFunction", (void **)&api->cuModuleGetFunction) &&
load_symbol(api->lib, "cuLaunchKernel", (void **)&api->cuLaunchKernel);
load_symbol(api->lib, "cuLaunchKernel", (void **)&api->cuLaunchKernel))) {
dlclose(api->lib);
memset(api, 0, sizeof(*api));
return 0;
}
load_symbol(api->lib, "cuMemGetInfo_v2", (void **)&api->cuMemGetInfo);
load_symbol(api->lib, "cuStreamCreate", (void **)&api->cuStreamCreate);
if (!load_symbol(api->lib, "cuStreamDestroy_v2", (void **)&api->cuStreamDestroy)) {
load_symbol(api->lib, "cuStreamDestroy", (void **)&api->cuStreamDestroy);
}
return 1;
}
static const char *cu_error_name(struct cuda_api *api, CUresult rc) {
@@ -220,6 +238,39 @@ static int query_compute_capability(struct cuda_api *api, CUdevice dev, int *maj
return 1;
}
static int query_multiprocessor_count(struct cuda_api *api, CUdevice dev, int *count) {
int mp_count = 0;
if (!check_rc(api,
"cuDeviceGetAttribute(multiprocessors)",
api->cuDeviceGetAttribute(&mp_count, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev))) {
return 0;
}
*count = mp_count;
return 1;
}
static size_t clamp_budget_to_free_memory(struct cuda_api *api, size_t requested_bytes) {
size_t free_bytes = 0;
size_t total_bytes = 0;
size_t max_bytes = requested_bytes;
if (!api->cuMemGetInfo) {
return requested_bytes;
}
if (api->cuMemGetInfo(&free_bytes, &total_bytes) != CU_SUCCESS || free_bytes == 0) {
return requested_bytes;
}
max_bytes = (free_bytes * 9u) / 10u;
if (max_bytes < (size_t)4u * 1024u * 1024u) {
max_bytes = (size_t)4u * 1024u * 1024u;
}
if (requested_bytes > max_bytes) {
return max_bytes;
}
return requested_bytes;
}
#if HAVE_CUBLASLT_HEADERS
static void append_detail(char *buf, size_t cap, const char *fmt, ...) {
size_t len = strlen(buf);
@@ -1095,13 +1146,16 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
int main(int argc, char **argv) {
int seconds = 5;
int size_mb = 64;
int device_index = 0;
for (int i = 1; i < argc; i++) {
if ((strcmp(argv[i], "--seconds") == 0 || strcmp(argv[i], "-t") == 0) && i + 1 < argc) {
seconds = atoi(argv[++i]);
} else if ((strcmp(argv[i], "--size-mb") == 0 || strcmp(argv[i], "-m") == 0) && i + 1 < argc) {
size_mb = atoi(argv[++i]);
} else if ((strcmp(argv[i], "--device") == 0 || strcmp(argv[i], "-d") == 0) && i + 1 < argc) {
device_index = atoi(argv[++i]);
} else {
fprintf(stderr, "usage: %s [--seconds N] [--size-mb N]\n", argv[0]);
fprintf(stderr, "usage: %s [--seconds N] [--size-mb N] [--device N]\n", argv[0]);
return 2;
}
}
@@ -1111,6 +1165,9 @@ int main(int argc, char **argv) {
if (size_mb <= 0) {
size_mb = 64;
}
if (device_index < 0) {
device_index = 0;
}
struct cuda_api cuda;
if (!load_cuda(&cuda)) {
@@ -1133,8 +1190,13 @@ int main(int argc, char **argv) {
return 1;
}
if (device_index >= count) {
fprintf(stderr, "device index %d out of range (found %d CUDA device(s))\n", device_index, count);
return 1;
}
CUdevice dev = 0;
if (!check_rc(&cuda, "cuDeviceGet", cuda.cuDeviceGet(&dev, 0))) {
if (!check_rc(&cuda, "cuDeviceGet", cuda.cuDeviceGet(&dev, device_index))) {
return 1;
}
@@ -1162,6 +1224,7 @@ int main(int argc, char **argv) {
}
printf("device=%s\n", report.device);
printf("device_index=%d\n", device_index);
printf("compute_capability=%d.%d\n", report.cc_major, report.cc_minor);
printf("backend=%s\n", report.backend);
printf("duration_s=%d\n", seconds);

4
iso/builder/build-cublas.sh
View File

@@ -1,9 +1,9 @@
#!/bin/sh
# build-cublas.sh — download cuBLASLt/cuBLAS/cudart runtime + headers for bee-gpu-stress.
# build-cublas.sh — download cuBLASLt/cuBLAS/cudart runtime + headers for bee-gpu-burn worker.
#
# Downloads .deb packages from NVIDIA's CUDA apt repository (Debian 12, x86_64),
# verifies them against Packages.gz, and extracts the small subset we need:
# - headers for compiling bee-gpu-stress against cuBLASLt
# - headers for compiling bee-gpu-burn worker against cuBLASLt
# - runtime libs for libcublas, libcublasLt, libcudart inside the ISO
set -e

55
iso/builder/build-john.sh Normal file
View File

@@ -0,0 +1,55 @@
#!/bin/sh
# build-john.sh — build John the Ripper jumbo with OpenCL support for the LiveCD.
#
# Downloads a pinned source snapshot from the official openwall/john repository,
# builds it inside the builder container, and caches the resulting run/ tree.
set -e
JOHN_COMMIT="$1"
DIST_DIR="$2"
[ -n "$JOHN_COMMIT" ] || { echo "usage: $0 <john-commit> <dist-dir>"; exit 1; }
[ -n "$DIST_DIR" ] || { echo "usage: $0 <john-commit> <dist-dir>"; exit 1; }
echo "=== John the Ripper jumbo ${JOHN_COMMIT} ==="
CACHE_DIR="${DIST_DIR}/john-${JOHN_COMMIT}"
CACHE_ROOT="${BEE_CACHE_DIR:-${DIST_DIR}/cache}"
DOWNLOAD_CACHE_DIR="${CACHE_ROOT}/john-downloads"
SRC_TAR="${DOWNLOAD_CACHE_DIR}/john-${JOHN_COMMIT}.tar.gz"
SRC_URL="https://github.com/openwall/john/archive/${JOHN_COMMIT}.tar.gz"
if [ -x "${CACHE_DIR}/run/john" ] && [ -f "${CACHE_DIR}/run/john.conf" ]; then
echo "=== john cached, skipping build ==="
echo "run dir: ${CACHE_DIR}/run"
exit 0
fi
mkdir -p "${DOWNLOAD_CACHE_DIR}"
if [ ! -f "${SRC_TAR}" ]; then
echo "=== downloading john source snapshot ==="
wget --show-progress -O "${SRC_TAR}" "${SRC_URL}"
fi
BUILD_TMP=$(mktemp -d)
trap 'rm -rf "${BUILD_TMP}"' EXIT INT TERM
cd "${BUILD_TMP}"
tar xf "${SRC_TAR}"
SRC_DIR=$(find . -maxdepth 1 -type d -name 'john-*' | head -1)
[ -n "${SRC_DIR}" ] || { echo "ERROR: john source directory not found"; exit 1; }
cd "${SRC_DIR}/src"
echo "=== configuring john ==="
./configure
echo "=== building john ==="
make clean >/dev/null 2>&1 || true
make -j"$(nproc)"
mkdir -p "${CACHE_DIR}"
cp -a "../run" "${CACHE_DIR}/run"
chmod +x "${CACHE_DIR}/run/john"
echo "=== john build complete ==="
echo "run dir: ${CACHE_DIR}/run"

20
iso/builder/build-nvidia-module.sh
View File

@@ -10,7 +10,7 @@
# Output layout:
# $CACHE_DIR/modules/ — nvidia*.ko files
# $CACHE_DIR/bin/ — nvidia-smi, nvidia-debugdump
# $CACHE_DIR/lib/ — libnvidia-ml.so*, libcuda.so* (for nvidia-smi)
# $CACHE_DIR/lib/ — libnvidia-ml.so*, libcuda.so*, OpenCL-related libs
set -e
@@ -133,7 +133,14 @@ fi
# Copy ALL userspace library files.
# libnvidia-ptxjitcompiler is required by libcuda for PTX JIT compilation
# (cuModuleLoadDataEx with PTX source) — without it CUDA_ERROR_JIT_COMPILER_NOT_FOUND.
for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
for lib in \
libnvidia-ml \
libcuda \
libnvidia-ptxjitcompiler \
libnvidia-opencl \
libnvidia-compiler \
libnvidia-nvvm \
libnvidia-fatbinaryloader; do
count=0
for f in $(find "$EXTRACT_DIR" -maxdepth 1 -name "${lib}.so.*" 2>/dev/null); do
cp "$f" "$CACHE_DIR/lib/" && count=$((count+1))
@@ -150,7 +157,14 @@ ko_count=$(ls "$CACHE_DIR/modules/"*.ko 2>/dev/null | wc -l)
[ "$ko_count" -gt 0 ] || { echo "ERROR: no .ko files built in $CACHE_DIR/modules/"; exit 1; }
# Create soname symlinks: use [0-9][0-9]* to avoid circular symlink (.so.1 has single digit)
for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
for lib in \
libnvidia-ml \
libcuda \
libnvidia-ptxjitcompiler \
libnvidia-opencl \
libnvidia-compiler \
libnvidia-nvvm \
libnvidia-fatbinaryloader; do
versioned=$(ls "$CACHE_DIR/lib/${lib}.so."[0-9][0-9]* 2>/dev/null | head -1)
[ -n "$versioned" ] || continue
base=$(basename "$versioned")

42
iso/builder/build.sh
View File

@@ -183,7 +183,7 @@ else
fi
# --- NVIDIA-only build steps ---
GPU_STRESS_BIN="${DIST_DIR}/bee-gpu-stress-linux-amd64"
GPU_BURN_WORKER_BIN="${DIST_DIR}/bee-gpu-burn-worker-linux-amd64"
if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
echo ""
echo "=== downloading cuBLAS/cuBLASLt/cudart ${NCCL_CUDA_VERSION} userspace ==="
@@ -196,20 +196,20 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
GPU_STRESS_NEED_BUILD=1
if [ -f "$GPU_STRESS_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_STRESS_BIN" ]; then
if [ -f "$GPU_BURN_WORKER_BIN" ] && [ "${BUILDER_DIR}/bee-gpu-stress.c" -ot "$GPU_BURN_WORKER_BIN" ]; then
GPU_STRESS_NEED_BUILD=0
fi
if [ "$GPU_STRESS_NEED_BUILD" = "1" ]; then
echo "=== building bee-gpu-stress ==="
echo "=== building bee-gpu-burn worker ==="
gcc -O2 -s -Wall -Wextra \
-I"${CUBLAS_CACHE}/include" \
-o "$GPU_STRESS_BIN" \
-o "$GPU_BURN_WORKER_BIN" \
"${BUILDER_DIR}/bee-gpu-stress.c" \
-ldl -lm
echo "binary: $GPU_STRESS_BIN"
echo "binary: $GPU_BURN_WORKER_BIN"
else
echo "=== bee-gpu-stress up to date, skipping build ==="
echo "=== bee-gpu-burn worker up to date, skipping build ==="
fi
fi
@@ -246,6 +246,9 @@ rm -f \
"${OVERLAY_STAGE_DIR}/root/.ssh/authorized_keys" \
"${OVERLAY_STAGE_DIR}/usr/local/bin/bee" \
"${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress" \
"${OVERLAY_STAGE_DIR}/usr/local/bin/john" \
"${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker" \
"${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john" \
"${OVERLAY_STAGE_DIR}/usr/local/bin/bee-smoketest" \
"${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
@@ -293,9 +296,13 @@ mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/bin"
cp "${DIST_DIR}/bee-linux-amd64" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee"
if [ "$BEE_GPU_VENDOR" = "nvidia" ] && [ -f "$GPU_STRESS_BIN" ]; then
cp "${GPU_STRESS_BIN}" "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress"
if [ "$BEE_GPU_VENDOR" = "nvidia" ] && [ -f "$GPU_BURN_WORKER_BIN" ]; then
mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee" "${OVERLAY_STAGE_DIR}/usr/local/bin"
cp "${GPU_BURN_WORKER_BIN}" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/bee-gpu-burn-worker"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-burn" 2>/dev/null || true
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-john-gpu-stress" 2>/dev/null || true
ln -sfn bee-gpu-burn "${OVERLAY_STAGE_DIR}/usr/local/bin/bee-gpu-stress"
fi
# --- inject smoketest into overlay so it runs directly on the live CD ---
@@ -334,6 +341,8 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
cp "${NVIDIA_CACHE}/bin/nvidia-bug-report.sh" "${OVERLAY_STAGE_DIR}/usr/local/bin/" 2>/dev/null || true
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/nvidia-bug-report.sh" 2>/dev/null || true
cp "${NVIDIA_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/" 2>/dev/null || true
mkdir -p "${OVERLAY_STAGE_DIR}/etc/OpenCL/vendors"
printf 'libnvidia-opencl.so.1\n' > "${OVERLAY_STAGE_DIR}/etc/OpenCL/vendors/nvidia.icd"
# Inject GSP firmware into /lib/firmware/nvidia/<version>/
if [ -d "${NVIDIA_CACHE}/firmware" ] && [ "$(ls -A "${NVIDIA_CACHE}/firmware" 2>/dev/null)" ]; then
@@ -353,7 +362,7 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
cp "${NCCL_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
echo "=== NCCL: $(ls "${NCCL_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
# Inject cuBLAS/cuBLASLt/cudart runtime libs used by bee-gpu-stress tensor-core GEMM path
# Inject cuBLAS/cuBLASLt/cudart runtime libs used by the bee-gpu-burn worker tensor-core GEMM path
cp "${CUBLAS_CACHE}/lib/"* "${OVERLAY_STAGE_DIR}/usr/lib/"
echo "=== cuBLAS: $(ls "${CUBLAS_CACHE}/lib/" | wc -l) files injected into /usr/lib/ ==="
@@ -372,6 +381,16 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
cp "${NCCL_TESTS_CACHE}/bin/all_reduce_perf" "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/bin/all_reduce_perf"
echo "=== all_reduce_perf injected ==="
echo ""
echo "=== building john jumbo ${JOHN_JUMBO_COMMIT} ==="
sh "${BUILDER_DIR}/build-john.sh" "${JOHN_JUMBO_COMMIT}" "${DIST_DIR}"
JOHN_CACHE="${DIST_DIR}/john-${JOHN_JUMBO_COMMIT}"
mkdir -p "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john"
rsync -a --delete "${JOHN_CACHE}/run/" "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john/run/"
ln -sfn ../lib/bee/john/run/john "${OVERLAY_STAGE_DIR}/usr/local/bin/john"
chmod +x "${OVERLAY_STAGE_DIR}/usr/local/lib/bee/john/run/john"
echo "=== john injected ==="
fi
# --- embed build metadata ---
@@ -385,7 +404,8 @@ NCCL_VERSION=${NCCL_VERSION}
NCCL_CUDA_VERSION=${NCCL_CUDA_VERSION}
CUBLAS_VERSION=${CUBLAS_VERSION}
CUDA_USERSPACE_VERSION=${CUDA_USERSPACE_VERSION}
NCCL_TESTS_VERSION=${NCCL_TESTS_VERSION}"
NCCL_TESTS_VERSION=${NCCL_TESTS_VERSION}
JOHN_JUMBO_COMMIT=${JOHN_JUMBO_COMMIT}"
GPU_BUILD_INFO="nvidia:${NVIDIA_DRIVER_VERSION}"
elif [ "$BEE_GPU_VENDOR" = "amd" ]; then
GPU_VERSION_LINE="ROCM_VERSION=${ROCM_VERSION}"

2
iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot
View File

@@ -60,6 +60,8 @@ chmod +x /usr/local/bin/bee 2>/dev/null || true
chmod +x /usr/local/bin/bee-log-run 2>/dev/null || true
if [ "$GPU_VENDOR" = "nvidia" ]; then
chmod +x /usr/local/bin/bee-nvidia-load 2>/dev/null || true
chmod +x /usr/local/bin/bee-gpu-burn 2>/dev/null || true
chmod +x /usr/local/bin/bee-john-gpu-stress 2>/dev/null || true
fi
# Reload udev rules

10
iso/builder/config/package-lists/bee-nvidia.list.chroot
View File

@@ -1,2 +1,8 @@
# NVIDIA DCGM (Data Center GPU Manager) — dcgmi diag for acceptance testing
datacenter-gpu-manager=1:%%DCGM_VERSION%%
# NVIDIA DCGM (Data Center GPU Manager) — dcgmi diag for acceptance testing.
# DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with CUDA 13 userspace,
# so install the CUDA 13 build plus proprietary diagnostic components explicitly.
datacenter-gpu-manager-4-cuda13=1:%%DCGM_VERSION%%
datacenter-gpu-manager-4-proprietary=1:%%DCGM_VERSION%%
datacenter-gpu-manager-4-proprietary-cuda13=1:%%DCGM_VERSION%%
ocl-icd-libopencl1
clinfo