Add staged NVIDIA burn ramp-up mode

This commit is contained in:
Mikhail Chusavitin
2026-04-09 15:21:14 +03:00
parent a78fdadd88
commit 9481ca2805
11 changed files with 226 additions and 47 deletions

View File

@@ -0,0 +1,110 @@
#!/bin/sh
set -eu
SECONDS=300
STAGGER_SECONDS=180
DEVICES=""
EXCLUDE=""
usage() {
echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3]" >&2
exit 2
}
normalize_list() {
echo "${1:-}" | tr ',' '\n' | sed 's/[[:space:]]//g' | awk 'NF' | sort -n | uniq | paste -sd, -
}
contains_csv() {
needle="$1"
haystack="${2:-}"
echo ",${haystack}," | grep -q ",${needle},"
}
resolve_dcgmproftester() {
for candidate in dcgmproftester dcgmproftester13 dcgmproftester12 dcgmproftester11; do
if command -v "${candidate}" >/dev/null 2>&1; then
command -v "${candidate}"
return 0
fi
done
return 1
}
while [ "$#" -gt 0 ]; do
case "$1" in
--seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
--stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
--devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
--exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
*) usage ;;
esac
done
PROF=$(resolve_dcgmproftester) || { echo "dcgmproftester not found in PATH" >&2; exit 1; }
ALL_DEVICES=$(nvidia-smi --query-gpu=index --format=csv,noheader,nounits 2>/dev/null | sed 's/[[:space:]]//g' | awk 'NF' | paste -sd, -)
[ -n "${ALL_DEVICES}" ] || { echo "nvidia-smi found no NVIDIA GPUs" >&2; exit 1; }
DEVICES=$(normalize_list "${DEVICES}")
EXCLUDE=$(normalize_list "${EXCLUDE}")
SELECTED="${DEVICES}"
if [ -z "${SELECTED}" ]; then
SELECTED="${ALL_DEVICES}"
fi
FINAL=""
for id in $(echo "${SELECTED}" | tr ',' ' '); do
[ -n "${id}" ] || continue
if contains_csv "${id}" "${EXCLUDE}"; then
continue
fi
if [ -z "${FINAL}" ]; then
FINAL="${id}"
else
FINAL="${FINAL},${id}"
fi
done
[ -n "${FINAL}" ] || { echo "no NVIDIA GPUs selected after filters" >&2; exit 1; }
echo "loader=dcgmproftester-staggered"
echo "selected_gpus=${FINAL}"
echo "stagger_seconds=${STAGGER_SECONDS}"
TMP_DIR=$(mktemp -d)
trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
gpu_pos=0
WORKERS=""
for id in $(echo "${FINAL}" | tr ',' ' '); do
gpu_pos=$((gpu_pos + 1))
log="${TMP_DIR}/gpu-${id}.log"
extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
gpu_seconds=$(( SECONDS + extra_sec ))
echo "starting gpu ${id} seconds=${gpu_seconds}"
CUDA_VISIBLE_DEVICES="${id}" "${PROF}" --no-dcgm-validation -t 1004 -d "${gpu_seconds}" >"${log}" 2>&1 &
pid=$!
WORKERS="${WORKERS} ${pid}:${id}:${log}"
if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
sleep "${STAGGER_SECONDS}"
fi
done
status=0
for spec in ${WORKERS}; do
pid=${spec%%:*}
rest=${spec#*:}
id=${rest%%:*}
log=${rest#*:}
if wait "${pid}"; then
echo "gpu ${id} finished: OK"
else
rc=$?
echo "gpu ${id} finished: FAILED rc=${rc}"
status=1
fi
sed "s/^/[gpu ${id}] /" "${log}" || true
done
exit "${status}"

17
iso/overlay/usr/local/bin/bee-gpu-burn Normal file → Executable file
View File

@@ -2,13 +2,14 @@
set -eu
SECONDS=5
STAGGER_SECONDS=0
SIZE_MB=0
DEVICES=""
EXCLUDE=""
WORKER="/usr/local/lib/bee/bee-gpu-burn-worker"
usage() {
echo "usage: $0 [--seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
echo "usage: $0 [--seconds N] [--stagger-seconds N] [--size-mb N] [--devices 0,1] [--exclude 2,3]" >&2
exit 2
}
@@ -25,6 +26,7 @@ contains_csv() {
while [ "$#" -gt 0 ]; do
case "$1" in
--seconds|-t) [ "$#" -ge 2 ] || usage; SECONDS="$2"; shift 2 ;;
--stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
--size-mb|-m) [ "$#" -ge 2 ] || usage; SIZE_MB="$2"; shift 2 ;;
--devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
--exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
@@ -61,14 +63,18 @@ done
echo "loader=bee-gpu-burn"
echo "selected_gpus=${FINAL}"
echo "stagger_seconds=${STAGGER_SECONDS}"
export CUDA_DEVICE_ORDER="PCI_BUS_ID"
TMP_DIR=$(mktemp -d)
trap 'rm -rf "${TMP_DIR}"' EXIT INT TERM
GPU_COUNT=$(echo "${FINAL}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
gpu_pos=0
WORKERS=""
for id in $(echo "${FINAL}" | tr ',' ' '); do
gpu_pos=$((gpu_pos + 1))
log="${TMP_DIR}/gpu-${id}.log"
gpu_size_mb="${SIZE_MB}"
if [ "${gpu_size_mb}" -le 0 ] 2>/dev/null; then
@@ -79,11 +85,16 @@ for id in $(echo "${FINAL}" | tr ',' ' '); do
gpu_size_mb=512
fi
fi
echo "starting gpu ${id} size=${gpu_size_mb}MB"
extra_sec=$(( STAGGER_SECONDS * (GPU_COUNT - gpu_pos) ))
gpu_seconds=$(( SECONDS + extra_sec ))
echo "starting gpu ${id} size=${gpu_size_mb}MB seconds=${gpu_seconds}"
CUDA_VISIBLE_DEVICES="${id}" \
"${WORKER}" --device 0 --seconds "${SECONDS}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
"${WORKER}" --device 0 --seconds "${gpu_seconds}" --size-mb "${gpu_size_mb}" >"${log}" 2>&1 &
pid=$!
WORKERS="${WORKERS} ${pid}:${id}:${log}"
if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${gpu_pos}" -lt "${GPU_COUNT}" ]; then
sleep "${STAGGER_SECONDS}"
fi
done
status=0

16
iso/overlay/usr/local/bin/bee-john-gpu-stress Normal file → Executable file
View File

@@ -2,6 +2,7 @@
set -eu
DURATION_SEC=300
STAGGER_SECONDS=0
DEVICES=""
EXCLUDE=""
FORMAT=""
@@ -12,7 +13,7 @@ export OCL_ICD_VENDORS="/etc/OpenCL/vendors"
export LD_LIBRARY_PATH="/usr/lib:/usr/local/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
usage() {
echo "usage: $0 [--seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
echo "usage: $0 [--seconds N] [--stagger-seconds N] [--devices 0,1] [--exclude 2,3] [--format name]" >&2
exit 2
}
@@ -118,6 +119,7 @@ ensure_opencl_ready() {
while [ "$#" -gt 0 ]; do
case "$1" in
--seconds|-t) [ "$#" -ge 2 ] || usage; DURATION_SEC="$2"; shift 2 ;;
--stagger-seconds) [ "$#" -ge 2 ] || usage; STAGGER_SECONDS="$2"; shift 2 ;;
--devices) [ "$#" -ge 2 ] || usage; DEVICES="$2"; shift 2 ;;
--exclude) [ "$#" -ge 2 ] || usage; EXCLUDE="$2"; shift 2 ;;
--format) [ "$#" -ge 2 ] || usage; FORMAT="$2"; shift 2 ;;
@@ -170,6 +172,7 @@ done
echo "loader=john"
echo "selected_gpus=${FINAL}"
echo "john_devices=${JOHN_DEVICES}"
echo "stagger_seconds=${STAGGER_SECONDS}"
cd "${JOHN_DIR}"
@@ -232,14 +235,21 @@ trap cleanup EXIT INT TERM
echo "format=${CHOSEN_FORMAT}"
echo "target_seconds=${DURATION_SEC}"
echo "slice_seconds=${TEST_SLICE_SECONDS}"
DEADLINE=$(( $(date +%s) + DURATION_SEC ))
TOTAL_DEVICES=$(echo "${JOHN_DEVICES}" | tr ',' '\n' | awk 'NF' | wc -l | tr -d '[:space:]')
_first=1
pos=0
for opencl_id in $(echo "${JOHN_DEVICES}" | tr ',' ' '); do
pos=$((pos + 1))
[ "${_first}" = "1" ] || sleep 3
_first=0
run_john_loop "${opencl_id}" "${DEADLINE}" &
extra_sec=$(( STAGGER_SECONDS * (TOTAL_DEVICES - pos) ))
deadline=$(( $(date +%s) + DURATION_SEC + extra_sec ))
run_john_loop "${opencl_id}" "${deadline}" &
pid=$!
PIDS="${PIDS} ${pid}"
if [ "${STAGGER_SECONDS}" -gt 0 ] && [ "${pos}" -lt "${TOTAL_DEVICES}" ]; then
sleep "${STAGGER_SECONDS}"
fi
done
FAIL=0
for pid in ${PIDS}; do