谷咕云计算

# 弹性伸缩组配置模板
scaling_group:
  name: deepseek-cluster
  region: ap-southeast-1
  vswitch_ids: [vsw-xxx1, vsw-xxx2]
  multi_az_policy: COST_OPTIMIZED  # 成本优先策略
  spot_strategy:
    allocation_strategy: diversified  # 分散实例类型
    instance_types:
      - ecs.g7.2xlarge.spot  # 计算优化型
      - ecs.c7.4xlarge.spot  # 通用型
      - ecs.r7.2xlarge.spot  # 内存优化型
  spot_price_limit: 
    ecs.g7.2xlarge: 0.25  # 设置最高出价（按需价格的30%）

import aliyunsdkcore
from aliyunsdkecs.request.v20140526 import DescribeSpotPriceHistoryRequest

def handle_interruption(instance_id):
    # 查询2分钟内最低价实例类型
    req = DescribeSpotPriceHistoryRequest.DescribeSpotPriceHistoryRequest()
    req.set_ZoneId(zone)
    req.set_NetworkType('vpc')
    req.set_InstanceTypeFamily('ecs.g7')
    prices = client.do_action_with_exception(req)
    
    # 自动切换到最低价可用实例
    lowest_price_type = min(prices, key=lambda x:x['Price'])
    migrate_instance(instance_id, lowest_price_type)

# 基于CPU/GPU负载自动扩缩
aliyun ess CreateScalingRule \
  --ScalingGroupId sg-xxx \
  --AdjustmentType TotalCapacity \
  --AdjustmentValue 1 \
  --Cooldown 300 \
  --ScaleInEvaluationCount 3 \
  --ScaleOutEvaluationCount 1 \
  --MetricType GPU_MEM_USAGE \
  --Threshold 70

# 安装抢占式实例工具包
wget https://spot-tools.oss-ap-southeast-1.aliyuncs.com/spotctl-linux-amd64.tgz
tar -zxvf spotctl-linux-amd64.tgz && mv spotctl /usr/local/bin/

# 一键部署DeepSeek
spotctl deploy deepseek \
  --model-version 7b-chat \
  --spot-strategy diversified \
  --min-size 2 \
  --max-size 10 \
  --vpc vpc-xxx \
  --security-group sg-xxx

# /etc/spotctl/config.ini
[autoscale]
check_interval = 60
scale_up_threshold = 70  # GPU利用率阈值
scale_down_threshold = 30
scale_step = 2

[spot]
max_interruption_rate = 0.2  # 允许20%实例中断
backup_ondemand_ratio = 0.3  # 30%按需实例保底

# 每小时成本查询
SELECT 
    instance_type,
    SUM(spot_cost) AS spot_cost,
    SUM(ondemand_cost) AS ondemand_cost,
    (SUM(ondemand_cost) - SUM(spot_cost)) / SUM(ondemand_cost) AS saving_rate
FROM 
    billing_data
WHERE 
    service = 'DeepSeek'
GROUP BY 
    instance_type, hour

# 每2小时快照备份
aliyun ecs CreateSnapshot \
  --DiskId d-xxx \
  --RetentionDays 3 \
  --SnapshotName "DeepSeek_$(date +%Y%m%d%H)"

# 自动恢复脚本
spotctl restore --cluster deepseek --snapshot latest --force

# 启用动态显存分配
from deepseek import OptimizeConfig

config = OptimizeConfig(
    memory_map_optimization=True,
    enable_cuda_graph=True, 
    max_workspace_size=4096  # MB
)
model.optimize(config)

# 启用ENA和GPU Direct RDMA
modprobe ena
nvidia-smi topo -m
echo "options nvidia NVreg_EnablePCIERelaxedOrderingMode=1" > /etc/modprobe.d/nvidia.conf

# 实时监控Spot实例中断率
spotctl monitor --metric SpotInterruptionRate --threshold 0.2

# 自动平衡实例分布
spotctl rebalance --strategy cost --max-disruption 15%