\u5b58\u50a8\u67b6\u6784\u8bbe\u8ba1<\/strong><\/h2>\n

AI\u5de5\u4f5c\u8d1f\u8f7d\u9700\u8981\u7cbe\u5fc3\u89c4\u5212\u7684\u5b58\u50a8\u5c42\u6b21\u7ed3\u6784\u3002\u4ee5\u4e0b\u662f\u57fa\u4e8e\u751f\u4ea7\u90e8\u7f72\u7684\u63a8\u8350\u8bbe\u7f6e\uff1a<\/p>\n

\r\n# \u5b58\u50a8\u5c42\u914d\u7f6e\r\n\/data\r\n\u251c\u2500\u2500 hot_tier\/      # NVMe\u56fa\u6001\u786c\u76d8\uff1a2GB\/s\u4ee5\u4e0a\u8bfb\u5199\r\n\u2502   \u251c\u2500\u2500 active_models\/\r\n\u2502   \u2514\u2500\u2500 current_datasets\/\r\n\u251c\u2500\u2500 warm_tier\/     # SATA\u56fa\u6001\u786c\u76d8\uff1a\u7ea6500MB\/s\r\n\u2502   \u251c\u2500\u2500 model_checkpoints\/\r\n\u2502   \u2514\u2500\u2500 preprocessed_data\/\r\n\u2514\u2500\u2500 cold_tier\/     # HDD\u9635\u5217\uff1a\u5f52\u6863\u5b58\u50a8\r\n    \u251c\u2500\u2500 historical_models\/\r\n    \u2514\u2500\u2500 raw_datasets\/\r\n<\/code><\/pre>\n
\u6210\u672c\u4f18\u5316\u7b56\u7565<\/strong><\/h2>\n
\u6d1b\u6749\u77f6\u670d\u52a1\u5668\u6258\u7ba1\u8bbe\u65bd\u4e3aAI\u57fa\u7840\u8bbe\u65bd\u63d0\u4f9b\u6218\u7565\u6210\u672c\u4f18\u52bf\u3002\u5f71\u54cd\u603b\u62e5\u6709\u6210\u672c\uff08TCO\uff09\u7684\u5173\u952e\u56e0\u7d20\u5305\u62ec\uff1a<\/p>\n
\n
\u786c\u4ef6\u914d\u7f6e\u53ef\u6269\u5c55\u6027<\/li>\n
\u7535\u6e90\u4f7f\u7528\u6548\u7387\uff08PUE\uff09<\/li>\n
\u7f51\u7edc\u5e26\u5bbd\u5206\u914d<\/li>\n
\u652f\u6301\u670d\u52a1\u7ea7\u522b<\/li>\n
\u5236\u51b7\u57fa\u7840\u8bbe\u65bd\u6548\u7387<\/li>\n<\/ul>\n
\u6027\u80fd\u4f18\u5316\u6280\u672f<\/strong><\/h2>\n
\u6700\u5927\u5316AI\u670d\u52a1\u5668\u6027\u80fd\u9700\u8981\u7cfb\u7edf\u7ea7\u4f18\u5316\u3002\u4ee5\u4e0b\u662fGPU\u670d\u52a1\u5668\u4f18\u5316\u7684\u5b9e\u9645\u793a\u4f8b\uff1a<\/p>\n
\r\n# \/etc\/sysctl.conf \u4f18\u5316\r\nvm.swappiness=10\r\nvm.dirty_background_ratio=5\r\nvm.dirty_ratio=10\r\nnet.core.rmem_max=16777216\r\nnet.core.wmem_max=16777216\r\n<\/code><\/pre>\n
\u90e8\u7f72\u67b6\u6784\u6a21\u5f0f<\/strong><\/h2>\n
\u5bf9\u4e8e\u751f\u4ea7AI\u90e8\u7f72\uff0c\u8bf7\u8003\u8651\u8fd9\u79cd\u7ecf\u8fc7\u5b9e\u6218\u68c0\u9a8c\u7684\u67b6\u6784\uff1a<\/p>\n
\n
\u8d1f\u8f7d\u5747\u8861\u5c42\uff1a\u5e26\u6709\u81ea\u5b9a\u4e49\u5065\u5eb7\u68c0\u67e5\u7684HAProxy<\/li>\n
\u63a8\u7406\u670d\u52a1\u5668\uff1a\u6c34\u5e73\u6269\u5c55\u7684GPU\u8282\u70b9<\/li>\n
\u8bad\u7ec3\u96c6\u7fa4\uff1a\u4e13\u7528\u9ad8\u5185\u5b58GPU\u670d\u52a1\u5668<\/li>\n
\u5b58\u50a8\u5c42\uff1a\u5206\u5e03\u5f0fNVMe\u9635\u5217<\/li>\n<\/ul>\n
\u4ee5\u4e0b\u662f\u793a\u4f8b\u90e8\u7f72\u914d\u7f6e\uff1a<\/p>\n
\r\nversion: '3.8'\r\nservices:\r\n  inference:\r\n    deploy:\r\n      replicas: 3\r\n      resources:\r\n        reservations:\r\n          devices:\r\n            - driver: nvidia\r\n              capabilities: [gpu]\r\n    volumes:\r\n      - model_storage:\/models\r\n      - cache:\/cache\r\n<\/code><\/pre>\n
\u76d1\u63a7\u548c\u7ef4\u62a4\u6700\u4f73\u5b9e\u8df5<\/strong><\/h2>\n
\u4f7f\u7528\u6b64\u5806\u6808\u4e3aAI\u57fa\u7840\u8bbe\u65bd\u5b9e\u65bd\u5168\u9762\u76d1\u63a7\uff1a<\/p>\n
\r\n# \u76d1\u63a7\u5806\u6808\u7ec4\u4ef6\r\n\u6307\u6807\u6536\u96c6\uff1aPrometheus\r\n\u53ef\u89c6\u5316\uff1aGrafana\r\n\u65e5\u5fd7\u7ba1\u7406\uff1aELK Stack\r\nGPU\u6307\u6807\uff1aDCGM-Exporter\r\n\u8b66\u62a5\u7ba1\u7406\uff1aAlertManager\r\n<\/code><\/pre>\n
\u9700\u8981\u76d1\u63a7\u7684\u5173\u952e\u6307\u6807\uff1a<\/p>\n
\n
GPU\u5229\u7528\u7387\u548c\u5185\u5b58\u4f7f\u7528\u60c5\u51b5<\/li>\n
CUDA\u5185\u5b58\u5206\u914d\u6a21\u5f0f<\/li>\n
PCIe\u5e26\u5bbd\u5229\u7528\u7387<\/li>\n
\u5b58\u50a8I\/O\u6a21\u5f0f<\/li>\n
\u6bcf\u4e2a\u6a21\u578b\u7684\u7f51\u7edc\u541e\u5410\u91cf<\/li>\n<\/ul>\n
\u6269\u5c55\u8003\u8651\u56e0\u7d20<\/strong><\/h2>\n
\u5728\u6d1b\u6749\u77f6\u6570\u636e\u4e2d\u5fc3\u6269\u5c55AI\u57fa\u7840\u8bbe\u65bd\u65f6\uff0c\u8bf7\u8003\u8651\u4ee5\u4e0b\u67b6\u6784\u6a21\u5f0f\uff1a<\/p>\n
\r\n# \u6269\u5c55\u6a21\u5f0f\u793a\u4f8b\r\n\u6c34\u5e73\u6269\u5c55\uff1a\r\n- \u5411\u63a8\u7406\u96c6\u7fa4\u6dfb\u52a0GPU\u8282\u70b9\r\n- \u5728\u8282\u70b9\u95f4\u5206\u914d\u6a21\u578b\u5206\u7247\r\n- \u5b9e\u65bd\u57fa\u4e8e\u8d1f\u8f7d\u7684\u81ea\u52a8\u6269\u5c55\r\n\r\n\u5782\u76f4\u6269\u5c55\uff1a\r\n- \u5347\u7ea7\u5230\u66f4\u9ad8VRAM\u7684GPU\r\n- \u589e\u52a0CPU\u6838\u5fc3\u6570\r\n- \u6269\u5c55NVMe\u5b58\u50a8\u5bb9\u91cf\r\n<\/code><\/pre>\n
\u5b89\u5168\u5b9e\u65bd<\/strong><\/h2>\n
\u4f7f\u7528\u4ee5\u4e0b\u57fa\u672c\u63aa\u65bd\u4fdd\u62a4\u60a8\u7684AI\u57fa\u7840\u8bbe\u65bd\uff1a<\/p>\n
\n
\u901a\u8fc7VLAN\u5b9e\u73b0\u7f51\u7edc\u9694\u79bb<\/li>\n
GPU\u7279\u5b9a\u8bbf\u95ee\u63a7\u5236<\/li>\n
\u6a21\u578b\u6784\u4ef6\u52a0\u5bc6<\/li>\n
API\u8eab\u4efd\u9a8c\u8bc1\u5c42<\/li>\n<\/ul>\n
\u9762\u5411\u672a\u6765\u7684\u57fa\u7840\u8bbe\u65bd<\/strong><\/h2>\n
\u5728\u89c4\u5212AI\u670d\u52a1\u5668\u79df\u7528\u57fa\u7840\u8bbe\u65bd\u65f6\uff0c\u8bf7\u8003\u8651\u4ee5\u4e0b\u65b0\u5174\u8d8b\u52bf\uff1a<\/p>\n
\n
\u9ad8\u5bc6\u5ea6\u673a\u67b6\u7684\u6db2\u51b7\u89e3\u51b3\u65b9\u6848<\/li>\n
PCIe Gen 5\u517c\u5bb9\u6027<\/li>\n
CXL\u5185\u5b58\u6269\u5c55\u652f\u6301<\/li>\n
\u91cf\u5b50\u5c31\u7eea\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd<\/li>\n<\/ul>\n
\u7ed3\u8bba<\/strong><\/h2>\n
\u9009\u62e9\u6b63\u786e\u7684AI\u670d\u52a1\u5668\u79df\u7528\u89e3\u51b3\u65b9\u6848\u9700\u8981\u5e73\u8861\u8ba1\u7b97\u80fd\u529b\u3001\u53ef\u6269\u5c55\u6027\u548c\u6210\u672c\u6548\u76ca\u3002\u6d1b\u6749\u77f6\u6570\u636e\u4e2d\u5fc3\u901a\u8fc7\u5c06\u5148\u8fdb\u7684GPU\u670d\u52a1\u5668\u6258\u7ba1\u670d\u52a1\u4e0e\u6700\u4f73\u7f51\u7edc\u8fde\u63a5\u76f8\u7ed3\u5408\uff0c\u4e3aAI\u6a21\u578b\u90e8\u7f72\u63d0\u4f9b\u6218\u7565\u4f18\u52bf\u3002\u65e0\u8bba\u60a8\u662f\u90e8\u7f72\u5927\u578b\u8bed\u8a00\u6a21\u578b\u8fd8\u662f\u8fd0\u884c\u4e13\u95e8\u7684\u673a\u5668\u5b66\u4e60\u5de5\u4f5c\u8d1f\u8f7d\uff0c\u5173\u952e\u662f\u8981\u5c06\u57fa\u7840\u8bbe\u65bd\u80fd\u529b\u4e0e\u60a8\u7684\u7279\u5b9aAI\u8ba1\u7b97\u9700\u6c42\u76f8\u5339\u914d\u3002<\/p>\n
\u5bf9\u4e8e\u6b63\u5728\u63a2\u7d22AI\u57fa\u7840\u8bbe\u65bd\u9009\u9879\u7684\u6280\u672f\u56e2\u961f\uff0c\u5efa\u8bae\u4ece\u8f83\u5c0f\u7684\u90e8\u7f72\u5f00\u59cb\uff0c\u9a8c\u8bc1\u6027\u80fd\u6307\u6807\u540e\u518d\u8fdb\u884c\u6269\u5c55\u3002\u8bf7\u8054\u7cfb\u6211\u4eec\u7684\u5de5\u7a0b\u56e2\u961f\uff0c\u83b7\u53d6\u8be6\u7ec6\u89c4\u683c\u548c\u5b9a\u5236\u7684AI\u670d\u52a1\u5668\u79df\u7528\u914d\u7f6e\uff0c\u4ee5\u6ee1\u8db3\u60a8\u7684\u673a\u5668\u5b66\u4e60\u9700\u6c42\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e3aAI\u6a21\u578b\u90e8\u7f72\u9009\u62e9\u5408\u9002\u7684\u670d\u52a1\u5668\u57fa\u7840\u8bbe\u65bd\u9700\u8981\u8003\u8651\u8d85\u8d8a\u4f20\u7edf\u670d\u52a1\u5668\u79df\u7528\u8981\u6c42\u7684\u590d\u6742\u6280\u672f\u56e0\u7d20\u3002\u65e0\u8bba\u60a8\u662f\u90e8\u7f72transformer\u6a21\u578b\u8fd8\u662f\u8fd0\u884c\u5bc6\u96c6\u578b\u795e\u7ecf\u7f51\u7edc\u8ba1\u7b97\uff0c\u60a8\u7684AI\u670d\u52a1\u5668\u79df\u7528\u8bbe\u7f6e\u90fd\u5c06\u51b3\u5b9a\u9879\u76ee\u7684\u6210\u8d25\u3002 AI\u5de5\u4f5c\u8d1f\u8f7d\u7684\u786c\u4ef6\u9700\u6c42\u5206\u6790 \u73b0\u4ee3AI\u5de5\u4f5c\u8d1f\u8f7d\u9700\u8981\u4e13\u95e8\u7684\u786c\u4ef6\u914d\u7f6e\u3002\u8ba9\u6211\u4eec\u901a\u8fc7\u5b9e\u9645\u57fa\u51c6\u6765\u5206\u6790\u57fa\u672c\u7ec4\u4ef6\uff1a # \u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u7684GPU\u5185\u5b58\u4f7f\u7528\u793a\u4f8b \u6a21\u578b\u5927\u5c0f \u6240\u9700VRAM \u63a8\u8350GPU 3B\u53c2\u6570 24GB NVIDIA A5000 7B\u53c2\u6570 40GB NVIDIA A6000 13B\u53c2\u6570 80GB NVIDIA A100 70B\u53c2\u6570 140GB+ \u591a\u4e2aA100 GPU\u67b6\u6784\u8003\u8651\u56e0\u7d20 \u5728\u9009\u62e9AI\u5de5\u4f5c\u8d1f\u8f7d\u7684GPU\u670d\u52a1\u5668\u65f6\uff0c\u67b6\u6784\u517c\u5bb9\u6027\u53d8\u5f97\u81f3\u5173\u91cd\u8981\u3002\u6700\u65b0\u7684NVIDIA Ampere\u548cHopper\u67b6\u6784\u63d0\u4f9b\u663e\u8457\u4f18\u52bf\uff1a \u5f20\u91cf\u6838\u5fc3\uff1a\u5bf9\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\u81f3\u5173\u91cd\u8981 NVLink\u8fde\u63a5\uff1a\u5b9e\u73b0\u591aGPU\u6269\u5c55 PCIe Gen 4\u652f\u6301\uff1a\u51cf\u5c11\u6570\u636e\u4f20\u8f93\u74f6\u9888 \u4ee5\u4e0b\u662fGPU\u4f7f\u7528\u7387\u76d1\u63a7\u7684\u5b9e\u9645\u793a\u4f8b\uff1a #!\/bin\/bash # GPU\u76d1\u63a7\u811a\u672c nvidia-smi –query-gpu=timestamp,name,utilization.gpu,memory.used,memory.total –format=csv -l 1 \u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\u9700\u6c42 AI\u6a21\u578b\u90e8\u7f72\u9700\u8981\u5f3a\u5927\u7684\u7f51\u7edc\u57fa\u7840\u8bbe\u65bd\u3002\u6d1b\u6749\u77f6\u6570\u636e\u4e2d\u5fc3\u901a\u8fc7\u4e0e\u4e3b\u8981\u4e91\u670d\u52a1\u63d0\u4f9b\u5546\u548c\u4e9a\u592a\u8def\u7ebf\u7684\u76f4\u63a5\u8fde\u63a5\u63d0\u4f9b\u6218\u7565\u4f18\u52bf\u3002\u8003\u8651\u4ee5\u4e0b\u7f51\u7edc\u89c4\u683c\uff1a \u6700\u4f4e10 Gbps\u4e13\u7528\u4e0a\u884c\u94fe\u8def \u4f4e\u5ef6\u8fdf\u8fde\u63a5\uff08\u5230\u4e3b\u8981\u4ea4\u6362\u70b9< 2ms\uff09 BGP\u8def\u7531\u4ee5\u5b9e\u73b0\u6700\u4f73\u8def\u5f84\u9009\u62e9 \u5b58\u50a8\u67b6\u6784\u8bbe\u8ba1 AI\u5de5\u4f5c\u8d1f\u8f7d\u9700\u8981\u7cbe\u5fc3\u89c4\u5212\u7684\u5b58\u50a8\u5c42\u6b21\u7ed3\u6784\u3002\u4ee5\u4e0b\u662f\u57fa\u4e8e\u751f\u4ea7\u90e8\u7f72\u7684\u63a8\u8350\u8bbe\u7f6e\uff1a # \u5b58\u50a8\u5c42\u914d\u7f6e \/data \u251c\u2500\u2500 hot_tier\/ # [...]\n\n
Read More…<\/a><\/p>\n","protected":false},"author":3,"featured_media":21770,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[4816],"tags":[6099,6101,6100,5108,5028],"class_list":["post-21773","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-america-dedicated-server-sc","tag-ai-server-hosting-sc","tag-ai-model-deployment-sc","tag-machine-learning-infrastructure-sc","tag-gpu-server-hosting-sc","tag-la-data-center-sc"],"acf":[],"yoast_head":"\n