109 lines
2.8 KiB
Bash
109 lines
2.8 KiB
Bash
#!/bin/bash
|
||
set -eo pipefail # 脚本遇到错误时立即退出，确保执行安全性
|
||
|
||
|
||
# 定义用法说明函数
|
||
usage() {
|
||
cat << EOF
|
||
Usage: $(basename "$0") --host=<host_list>
|
||
|
||
Description:
|
||
自动执行MPI all_reduce性能测试，支持指定主机列表，自动计算进程数(np = 服务器数量 × 8)
|
||
|
||
Required Options:
|
||
--host=<host_list> 指定主机列表，格式为"ip1[:count],ip2[:count],..."（例如：10.0.21.131:8,10.0.21.132:8）
|
||
注：每个主机默认对应8个进程，count字段不影响np计算，仅用于MPI主机分配
|
||
|
||
Example:
|
||
bash $(basename "$0") --host=10.0.21.131:8,10.0.21.132:8
|
||
EOF
|
||
exit 1
|
||
}
|
||
|
||
|
||
# 解析命令行参数
|
||
HOST_LIST=""
|
||
for arg in "$@"; do
|
||
case "$arg" in
|
||
--host=*)
|
||
HOST_LIST="${arg#*=}" # 提取--host=后面的值
|
||
shift
|
||
;;
|
||
*)
|
||
echo "错误：未知参数 '$arg'"
|
||
usage
|
||
;;
|
||
esac
|
||
done
|
||
|
||
|
||
# 校验参数合法性
|
||
if [ -z "$HOST_LIST" ]; then
|
||
echo "错误：必须通过--host指定主机列表"
|
||
usage
|
||
fi
|
||
|
||
# 分割主机列表，统计服务器数量（按逗号分割）
|
||
IFS=',' read -ra HOST_ARRAY <<< "$HOST_LIST"
|
||
SERVER_COUNT=${#HOST_ARRAY[@]} # 服务器数量 = 主机列表条目数
|
||
|
||
if [ "$SERVER_COUNT" -eq 0 ]; then
|
||
echo "错误：主机列表为空，请检查--host参数格式"
|
||
usage
|
||
fi
|
||
|
||
|
||
# 计算np值（每个服务器8个进程）
|
||
NP=$((SERVER_COUNT * 8))
|
||
|
||
|
||
# 打印配置信息（便于核对）
|
||
echo "======================================"
|
||
echo "MPI运行配置信息"
|
||
echo "--------------------------------------"
|
||
echo "主机列表: $HOST_LIST"
|
||
echo "服务器数量: $SERVER_COUNT 台"
|
||
echo "总进程数(np): $NP (每台服务器8个进程)"
|
||
echo "======================================"
|
||
echo
|
||
|
||
|
||
# 执行mpirun命令
|
||
echo "开始执行MPI测试..."
|
||
echo "命令: mpirun ... (详细参数见下方)"
|
||
echo
|
||
|
||
mpirun \
|
||
-np "$NP" \
|
||
--bind-to numa \
|
||
--map-by ppr:8:node \
|
||
--host "$HOST_LIST" \
|
||
--allow-run-as-root \
|
||
-x LD_LIBRARY_PATH \
|
||
--mca pml ob1 \
|
||
--mca coll ^hcoll \
|
||
--mca btl tcp,self \
|
||
--mca btl_tcp_if_include bond0 \
|
||
--mca oob_tcp_if_include bond0 \
|
||
--bind-to none \
|
||
-x NCCL_IB_HCA==mlx5_bond_2,mlx5_bond_3,mlx5_bond_4,mlx5_bond_5,mlx5_bond_6,mlx5_bond_7,mlx5_bond_8,mlx5_bond_9 \
|
||
-x NCCL_IB_QPS_PER_CONNECTION=8 \
|
||
-x NCCL_MIN_NCHANNELS=32 \
|
||
-x NCCL_SOCKET_IFNAME=bond0 \
|
||
all_reduce_perf \
|
||
--minbytes=16G \
|
||
--maxbytes=32G \
|
||
--stepfactor=2 \
|
||
--datatype=float \
|
||
--iters=100
|
||
|
||
|
||
# 命令执行结果反馈
|
||
if [ $? -eq 0 ]; then
|
||
echo "======================================"
|
||
echo "MPI测试执行完成，结果已输出"
|
||
else
|
||
echo "======================================"
|
||
echo "错误：MPI测试执行失败，请检查日志" >&2
|
||
exit 1
|
||
fi |