更新 scripts/deploy_gpu_monitor.sh
This commit is contained in:
parent
2e0567988c
commit
5c1dd6f80c
|
|
@ -4,6 +4,13 @@
|
||||||
set -euo pipefail
|
set -euo pipefail
|
||||||
IFS=$'\n\t'
|
IFS=$'\n\t'
|
||||||
|
|
||||||
|
# 颜色定义
|
||||||
|
RED='\033[0;31m'
|
||||||
|
GREEN='\033[0;32m'
|
||||||
|
YELLOW='\033[0;33m'
|
||||||
|
BLUE='\033[0;34m'
|
||||||
|
NC='\033[0m' # 恢复默认颜色
|
||||||
|
|
||||||
# 配置信息
|
# 配置信息
|
||||||
SCRIPT_URL="http://116.205.97.109/scripts/gpu_metrics.sh" # 脚本下载地址
|
SCRIPT_URL="http://116.205.97.109/scripts/gpu_metrics.sh" # 脚本下载地址
|
||||||
TARGET_DIR="/opt/scripts" # 脚本存放目录
|
TARGET_DIR="/opt/scripts" # 脚本存放目录
|
||||||
|
|
@ -17,20 +24,20 @@ NODE_EXPORTER_SERVICE="node_exporter.service" # node_exporter服务名
|
||||||
|
|
||||||
# 帮助信息
|
# 帮助信息
|
||||||
show_help() {
|
show_help() {
|
||||||
echo "用法: $0 [选项]"
|
echo -e "${BLUE}用法: $0 [选项]${NC}"
|
||||||
echo "GPU监控脚本部署工具(必须指定--install或--uninstall)"
|
echo -e "${BLUE}GPU监控脚本部署工具(必须指定--install或--uninstall)${NC}"
|
||||||
echo ""
|
echo ""
|
||||||
echo "选项:"
|
echo -e "${GREEN}选项:${NC}"
|
||||||
echo " --install 安装并配置GPU监控脚本,重启node_exporter"
|
echo -e " ${YELLOW}--install${NC} 安装并配置GPU监控脚本,重启node_exporter"
|
||||||
echo " --uninstall 卸载GPU监控脚本及配置,删除指标文件并重启node_exporter"
|
echo -e " ${YELLOW}--uninstall${NC} 卸载GPU监控脚本及配置,删除指标文件并重启node_exporter"
|
||||||
echo " --help 显示此帮助信息"
|
echo -e " ${YELLOW}--help${NC} 显示此帮助信息"
|
||||||
echo ""
|
echo ""
|
||||||
}
|
}
|
||||||
|
|
||||||
# 检查是否以root权限运行
|
# 检查是否以root权限运行
|
||||||
check_root() {
|
check_root() {
|
||||||
if [ "$(id -u)" -ne 0 ]; then
|
if [ "$(id -u)" -ne 0 ]; then
|
||||||
echo "错误: 此脚本需要root权限运行,请使用sudo" >&2
|
echo -e "${RED}错误: 此脚本需要root权限运行,请使用sudo${NC}" >&2
|
||||||
exit 1
|
exit 1
|
||||||
fi
|
fi
|
||||||
}
|
}
|
||||||
|
|
@ -42,19 +49,19 @@ service_exists() {
|
||||||
|
|
||||||
# 重启node_exporter并检查状态
|
# 重启node_exporter并检查状态
|
||||||
restart_node_exporter() {
|
restart_node_exporter() {
|
||||||
echo "检查node_exporter服务状态..."
|
echo -e "${BLUE}检查node_exporter服务状态...${NC}"
|
||||||
if service_exists "${NODE_EXPORTER_SERVICE}"; then
|
if service_exists "${NODE_EXPORTER_SERVICE}"; then
|
||||||
echo "正在重启node_exporter服务..."
|
echo -e "${BLUE}正在重启node_exporter服务...${NC}"
|
||||||
systemctl restart "${NODE_EXPORTER_SERVICE}" || true # 允许失败,不中断流程
|
systemctl restart "${NODE_EXPORTER_SERVICE}" || true # 允许失败,不中断流程
|
||||||
|
|
||||||
echo "检查node_exporter状态..."
|
echo -e "${BLUE}检查node_exporter状态...${NC}"
|
||||||
if systemctl is-active --quiet "${NODE_EXPORTER_SERVICE}"; then
|
if systemctl is-active --quiet "${NODE_EXPORTER_SERVICE}"; then
|
||||||
echo "node_exporter状态: 运行中"
|
echo -e "${GREEN}node_exporter状态: 运行中${NC}"
|
||||||
else
|
else
|
||||||
echo "警告: node_exporter未正常运行,请查看日志:journalctl -u ${NODE_EXPORTER_SERVICE}" >&2
|
echo -e "${YELLOW}警告: node_exporter未正常运行,请查看日志:journalctl -u ${NODE_EXPORTER_SERVICE}${NC}" >&2
|
||||||
fi
|
fi
|
||||||
else
|
else
|
||||||
echo "警告: node_exporter服务不存在,跳过重启" >&2
|
echo -e "${YELLOW}警告: node_exporter服务不存在,跳过重启${NC}" >&2
|
||||||
fi
|
fi
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
@ -64,7 +71,7 @@ add_cron_job() {
|
||||||
|
|
||||||
# 检查任务是否已存在(精确匹配)
|
# 检查任务是否已存在(精确匹配)
|
||||||
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
||||||
echo "定时任务已存在,跳过添加"
|
echo -e "${YELLOW}定时任务已存在,跳过添加${NC}"
|
||||||
return 0
|
return 0
|
||||||
fi
|
fi
|
||||||
|
|
||||||
|
|
@ -73,7 +80,7 @@ add_cron_job() {
|
||||||
crontab -l 2>/dev/null > "${backup_file}" || true
|
crontab -l 2>/dev/null > "${backup_file}" || true
|
||||||
|
|
||||||
# 添加新任务
|
# 添加新任务
|
||||||
echo "添加定时任务: ${cron_line}"
|
echo -e "${BLUE}添加定时任务: ${cron_line}${NC}"
|
||||||
(
|
(
|
||||||
crontab -l 2>/dev/null || true # 获取现有crontab(如果有)
|
crontab -l 2>/dev/null || true # 获取现有crontab(如果有)
|
||||||
echo "${cron_line}" # 添加新任务
|
echo "${cron_line}" # 添加新任务
|
||||||
|
|
@ -81,13 +88,13 @@ add_cron_job() {
|
||||||
|
|
||||||
# 验证添加结果
|
# 验证添加结果
|
||||||
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
||||||
echo "定时任务添加成功"
|
echo -e "${GREEN}定时任务添加成功${NC}"
|
||||||
else
|
else
|
||||||
echo "错误: 定时任务添加失败" >&2
|
echo -e "${RED}错误: 定时任务添加失败${NC}" >&2
|
||||||
# 恢复备份
|
# 恢复备份
|
||||||
if [ -f "${backup_file}" ]; then
|
if [ -f "${backup_file}" ]; then
|
||||||
crontab "${backup_file}"
|
crontab "${backup_file}"
|
||||||
echo "已恢复原crontab"
|
echo -e "${BLUE}已恢复原crontab${NC}"
|
||||||
fi
|
fi
|
||||||
return 1
|
return 1
|
||||||
fi
|
fi
|
||||||
|
|
@ -95,28 +102,28 @@ add_cron_job() {
|
||||||
|
|
||||||
# 安装函数
|
# 安装函数
|
||||||
install_script() {
|
install_script() {
|
||||||
echo "===== 开始安装GPU监控脚本 ====="
|
echo -e "${GREEN}===== 开始安装GPU监控脚本 =====${NC}"
|
||||||
|
|
||||||
# 1. 清理旧指标文件(避免残留数据)
|
# 1. 清理旧指标文件(避免残留数据)
|
||||||
echo "清理旧指标文件..."
|
echo -e "${BLUE}清理旧指标文件...${NC}"
|
||||||
if [ -f "${METRICS_FILE}" ]; then
|
if [ -f "${METRICS_FILE}" ]; then
|
||||||
rm -f "${METRICS_FILE}"
|
rm -f "${METRICS_FILE}"
|
||||||
echo "已删除旧指标文件: ${METRICS_FILE}"
|
echo -e "${GREEN}已删除旧指标文件: ${METRICS_FILE}${NC}"
|
||||||
fi
|
fi
|
||||||
|
|
||||||
# 2. 创建目标目录并下载脚本
|
# 2. 创建目标目录并下载脚本
|
||||||
mkdir -p "${TARGET_DIR}"
|
mkdir -p "${TARGET_DIR}"
|
||||||
cd "${TARGET_DIR}"
|
cd "${TARGET_DIR}"
|
||||||
|
|
||||||
echo "正在下载GPU监控脚本..."
|
echo -e "${BLUE}正在下载GPU监控脚本...${NC}"
|
||||||
wget -q -O "${TARGET_SCRIPT}" "${SCRIPT_URL}" || {
|
wget -q -O "${TARGET_SCRIPT}" "${SCRIPT_URL}" || {
|
||||||
echo "错误: 无法从 ${SCRIPT_URL} 下载脚本" >&2
|
echo -e "${RED}错误: 无法从 ${SCRIPT_URL} 下载脚本${NC}" >&2
|
||||||
exit 1
|
exit 1
|
||||||
}
|
}
|
||||||
chmod +x "${TARGET_SCRIPT}"
|
chmod +x "${TARGET_SCRIPT}"
|
||||||
|
|
||||||
# 3. 脚本配置(确保GPU总数为8)
|
# 3. 脚本配置(确保GPU总数为8)
|
||||||
echo "配置脚本参数..."
|
echo -e "${BLUE}配置脚本参数...${NC}"
|
||||||
sed -i 's/EXPECTED_GPUS=".*"/EXPECTED_GPUS="0 1 2 3 4 5 6 7"/' "${TARGET_SCRIPT}"
|
sed -i 's/EXPECTED_GPUS=".*"/EXPECTED_GPUS="0 1 2 3 4 5 6 7"/' "${TARGET_SCRIPT}"
|
||||||
|
|
||||||
# 4. 准备metrics目录
|
# 4. 准备metrics目录
|
||||||
|
|
@ -124,13 +131,13 @@ install_script() {
|
||||||
chmod 755 "${METRICS_DIR}"
|
chmod 755 "${METRICS_DIR}"
|
||||||
|
|
||||||
# 5. 配置定时任务(安全添加,避免重复)
|
# 5. 配置定时任务(安全添加,避免重复)
|
||||||
echo "配置定时任务..."
|
echo -e "${BLUE}配置定时任务...${NC}"
|
||||||
add_cron_job || exit 1
|
add_cron_job || exit 1
|
||||||
|
|
||||||
# 6. 首次执行脚本生成指标
|
# 6. 首次执行脚本生成指标
|
||||||
echo "测试执行脚本..."
|
echo -e "${BLUE}测试执行脚本...${NC}"
|
||||||
"${TARGET_SCRIPT}" || {
|
"${TARGET_SCRIPT}" || {
|
||||||
echo "错误: 脚本执行失败,请检查日志: ${LOG_FILE}" >&2
|
echo -e "${RED}错误: 脚本执行失败,请检查日志: ${LOG_FILE}${NC}" >&2
|
||||||
exit 1
|
exit 1
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
@ -138,20 +145,20 @@ install_script() {
|
||||||
restart_node_exporter
|
restart_node_exporter
|
||||||
|
|
||||||
# 8. 验证结果
|
# 8. 验证结果
|
||||||
echo "===== 安装完成 ====="
|
echo -e "${GREEN}===== 安装完成 =====${NC}"
|
||||||
echo "脚本位置: ${TARGET_SCRIPT}"
|
echo -e "${BLUE}脚本位置:${NC} ${TARGET_SCRIPT}"
|
||||||
echo "定时任务: $(crontab -l | grep "${TARGET_SCRIPT}")"
|
echo -e "${BLUE}定时任务:${NC} $(crontab -l | grep "${TARGET_SCRIPT}")"
|
||||||
echo "日志文件: ${LOG_FILE}"
|
echo -e "${BLUE}日志文件:${NC} ${LOG_FILE}"
|
||||||
echo "指标文件: ${METRICS_FILE}(已生成)"
|
echo -e "${BLUE}指标文件:${NC} ${METRICS_FILE}${GREEN}(已生成)${NC}"
|
||||||
echo "node_exporter状态: $(systemctl is-active "${NODE_EXPORTER_SERVICE}" 2>/dev/null || echo "未安装")"
|
echo -e "${BLUE}node_exporter状态:${NC} $(systemctl is-active "${NODE_EXPORTER_SERVICE}" 2>/dev/null || echo -e "${YELLOW}未安装${NC}")"
|
||||||
}
|
}
|
||||||
|
|
||||||
# 卸载函数
|
# 卸载函数
|
||||||
uninstall_script() {
|
uninstall_script() {
|
||||||
echo "===== 开始卸载GPU监控脚本 ====="
|
echo -e "${YELLOW}===== 开始卸载GPU监控脚本 =====${NC}"
|
||||||
|
|
||||||
# 1. 移除定时任务
|
# 1. 移除定时任务
|
||||||
echo "删除定时任务..."
|
echo -e "${BLUE}删除定时任务...${NC}"
|
||||||
local cron_line="${CRON_SCHEDULE} ${TARGET_SCRIPT} >> ${LOG_FILE} 2>&1"
|
local cron_line="${CRON_SCHEDULE} ${TARGET_SCRIPT} >> ${LOG_FILE} 2>&1"
|
||||||
|
|
||||||
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
if crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
||||||
|
|
@ -164,51 +171,51 @@ uninstall_script() {
|
||||||
|
|
||||||
# 验证移除结果
|
# 验证移除结果
|
||||||
if ! crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
if ! crontab -l 2>/dev/null | grep -qxF "${cron_line}"; then
|
||||||
echo "已成功删除定时任务"
|
echo -e "${GREEN}已成功删除定时任务${NC}"
|
||||||
else
|
else
|
||||||
echo "错误: 定时任务删除失败" >&2
|
echo -e "${RED}错误: 定时任务删除失败${NC}" >&2
|
||||||
# 恢复备份
|
# 恢复备份
|
||||||
crontab "${backup_file}"
|
crontab "${backup_file}"
|
||||||
echo "已恢复原crontab"
|
echo -e "${BLUE}已恢复原crontab${NC}"
|
||||||
exit 1
|
exit 1
|
||||||
fi
|
fi
|
||||||
else
|
else
|
||||||
echo "未找到相关定时任务,跳过"
|
echo -e "${YELLOW}未找到相关定时任务,跳过${NC}"
|
||||||
fi
|
fi
|
||||||
|
|
||||||
# 2. 删除脚本文件
|
# 2. 删除脚本文件
|
||||||
echo "删除监控脚本..."
|
echo -e "${BLUE}删除监控脚本...${NC}"
|
||||||
if [ -f "${TARGET_SCRIPT}" ]; then
|
if [ -f "${TARGET_SCRIPT}" ]; then
|
||||||
rm -f "${TARGET_SCRIPT}"
|
rm -f "${TARGET_SCRIPT}"
|
||||||
echo "已删除脚本: ${TARGET_SCRIPT}"
|
echo -e "${GREEN}已删除脚本: ${TARGET_SCRIPT}${NC}"
|
||||||
else
|
else
|
||||||
echo "脚本文件不存在,跳过"
|
echo -e "${YELLOW}脚本文件不存在,跳过${NC}"
|
||||||
fi
|
fi
|
||||||
|
|
||||||
# 3. 删除指标文件
|
# 3. 删除指标文件
|
||||||
echo "删除指标文件..."
|
echo -e "${BLUE}删除指标文件...${NC}"
|
||||||
if [ -f "${METRICS_FILE}" ]; then
|
if [ -f "${METRICS_FILE}" ]; then
|
||||||
rm -f "${METRICS_FILE}"
|
rm -f "${METRICS_FILE}"
|
||||||
echo "已删除指标文件: ${METRICS_FILE}"
|
echo -e "${GREEN}已删除指标文件: ${METRICS_FILE}${NC}"
|
||||||
else
|
else
|
||||||
echo "指标文件不存在,跳过"
|
echo -e "${YELLOW}指标文件不存在,跳过${NC}"
|
||||||
fi
|
fi
|
||||||
|
|
||||||
# 4. 删除日志文件
|
# 4. 删除日志文件
|
||||||
echo "删除日志文件..."
|
echo -e "${BLUE}删除日志文件...${NC}"
|
||||||
if [ -f "${LOG_FILE}" ]; then
|
if [ -f "${LOG_FILE}" ]; then
|
||||||
rm -f "${LOG_FILE}"
|
rm -f "${LOG_FILE}"
|
||||||
echo "已删除日志文件: ${LOG_FILE}"
|
echo -e "${GREEN}已删除日志文件: ${LOG_FILE}${NC}"
|
||||||
else
|
else
|
||||||
echo "日志文件不存在,跳过"
|
echo -e "${YELLOW}日志文件不存在,跳过${NC}"
|
||||||
fi
|
fi
|
||||||
|
|
||||||
# 5. 重启node_exporter
|
# 5. 重启node_exporter
|
||||||
restart_node_exporter
|
restart_node_exporter
|
||||||
|
|
||||||
# 6. 验证结果
|
# 6. 验证结果
|
||||||
echo "===== 卸载完成 ====="
|
echo -e "${YELLOW}===== 卸载完成 =====${NC}"
|
||||||
echo "node_exporter状态: $(systemctl is-active "${NODE_EXPORTER_SERVICE}" 2>/dev/null || echo "未安装")"
|
echo -e "${BLUE}node_exporter状态:${NC} $(systemctl is-active "${NODE_EXPORTER_SERVICE}" 2>/dev/null || echo -e "${YELLOW}未安装${NC}")"
|
||||||
}
|
}
|
||||||
|
|
||||||
# 主逻辑
|
# 主逻辑
|
||||||
|
|
@ -216,7 +223,7 @@ check_root
|
||||||
|
|
||||||
# 参数检查
|
# 参数检查
|
||||||
if [ $# -ne 1 ]; then
|
if [ $# -ne 1 ]; then
|
||||||
echo "错误: 必须指定--install或--uninstall" >&2
|
echo -e "${RED}错误: 必须指定--install或--uninstall${NC}" >&2
|
||||||
show_help
|
show_help
|
||||||
exit 1
|
exit 1
|
||||||
fi
|
fi
|
||||||
|
|
@ -234,8 +241,8 @@ case "$1" in
|
||||||
exit 0
|
exit 0
|
||||||
;;
|
;;
|
||||||
*)
|
*)
|
||||||
echo "错误: 未知参数 '$1'" >&2
|
echo -e "${RED}错误: 未知参数 '$1'${NC}" >&2
|
||||||
show_help
|
show_help
|
||||||
exit 1
|
exit 1
|
||||||
;;
|
;;
|
||||||
esac
|
esac
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue