\u5e26\u5bbd\u5206\u6790\u548cGPU\u4e92\u8fde<\/strong><\/h2>\n

\u5728\u9999\u6e2f\u6570\u636e\u4e2d\u5fc3\u6784\u5efa\u591aGPU\u7cfb\u7edf\u65f6\uff0c\u7406\u89e3\u5e26\u5bbd\u5206\u914d\u81f3\u5173\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u4f7f\u7528\u53ccCPU\u670d\u52a1\u5668\u914d\u7f6e\u7684\u8be6\u7ec6\u5206\u6790\uff1a<\/p>\n

\r\n# \u5e26\u5bbd\u5206\u914d\u793a\u4f8b\uff08\u53ccIntel Xeon\u5e73\u53f0\uff09\r\nCPU1 \u2192 GPU1: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU1 \u2192 GPU2: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU2 \u2192 GPU3: PCIe Gen4 x16 (31.5 GB\/s)\r\nCPU2 \u2192 GPU4: PCIe Gen4 x16 (31.5 GB\/s)\r\n\r\nCPU\u95f4\u901a\u4fe1\uff1aUPI\u94fe\u8def\r\n3\u6761UPI\u94fe\u8def \u00d7 23.3 GB\/s = \u603b\u8ba169.9 GB\/s\r\n<\/code><\/pre>\n
\u9999\u6e2f\u7279\u5b9a\u914d\u7f6e\u8003\u8651\u56e0\u7d20<\/strong><\/h2>\n
\u9999\u6e2f\u7684\u6c14\u5019\u4e3aGPU\u670d\u52a1\u5668\u90e8\u7f72\u5e26\u6765\u72ec\u7279\u6311\u6218\u3002\u9ad8\u6e7f\u5ea6\u548c\u9ad8\u6e29\u9700\u8981\u7279\u5b9a\u7684PCIe\u62d3\u6251\u7ed3\u6784\u8003\u8651\uff1a<\/p>\n
\n
PCIe\u63d2\u69fd\u7684\u70ed\u8bbe\u8ba1\u529f\u8017\uff08TDP\uff09\u5206\u5e03<\/li>\n
\u901a\u8fc7\u6218\u7565\u6027GPU\u5e03\u7f6e\u4f18\u5316\u6c14\u6d41<\/li>\n
\u9ad8\u5bc6\u5ea6\u914d\u7f6e\u7684\u5197\u4f59\u6563\u70ed\u7cfb\u7edf<\/li>\n<\/ul>\n
\u4e3a\u5728\u9999\u6e2f\u73af\u5883\u4e2d\u83b7\u5f97\u6700\u4f73\u6027\u80fd\uff0c\u5efa\u8bae\u91c7\u7528\u4ee5\u4e0bPCIe\u63d2\u69fd\u914d\u7f6e\uff1a<\/p>\n
\r\n# \u63a8\u8350\u7684PCIe\u63d2\u69fd\u914d\u7f6e\r\n\u63d2\u69fd1\uff1aGPU1\uff08\u4e3b\u8981\uff09 - PCIe Gen4 x16\r\n\u63d2\u69fd3\uff1aGPU2 - PCIe Gen4 x16\r\n\u63d2\u69fd5\uff1aGPU3 - PCIe Gen4 x16\r\n\u63d2\u69fd7\uff1aGPU4 - PCIe Gen4 x16\r\n\r\n# \u6ce8\u610f\uff1a\u4fdd\u6301\u6700\u5c112\u4e2a\u63d2\u69fd\u95f4\u8ddd\u4ee5\u8fdb\u884c\u6563\u70ed\u7ba1\u7406\r\n<\/code><\/pre>\n
\u6027\u80fd\u4f18\u5316\u6280\u672f<\/strong><\/h2>\n
\u8981\u5728\u9999\u6e2f\u670d\u52a1\u5668\u79df\u7528\u73af\u5883\u4e2d\u6700\u5927\u5316GPU\u670d\u52a1\u5668\u6027\u80fd\uff0c\u8bf7\u5b9e\u65bd\u8fd9\u4e9bPCIe\u62d3\u6251\u4f18\u5316\uff1a<\/p>\n
\n
NUMA\u8282\u70b9\u4f18\u5316\n
\n
\u5c06GPU\u7ed1\u5b9a\u5230\u672c\u5730NUMA\u8282\u70b9<\/li>\n
\u6700\u5c0f\u5316\u8de8NUMA\u901a\u4fe1<\/li>\n
\u4f18\u5316\u5185\u5b58\u5206\u914d\u6a21\u5f0f<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u4ee5\u4e0b\u662fLinux\u4e2dNUMA\u7ed1\u5b9a\u7684\u5b9e\u9645\u793a\u4f8b\uff1a<\/p>\n
\r\n# NUMA\u7ed1\u5b9a\u793a\u4f8b\r\nnumactl --cpunodebind=0 --membind=0 .\/gpu_application  # \u7528\u4e8eGPU0\/1\r\nnumactl --cpunodebind=1 --membind=1 .\/gpu_application  # \u7528\u4e8eGPU2\/3\r\n\r\n# \u68c0\u67e5NUMA\u62d3\u6251\r\nnvidia-smi topo -m\r\n<\/code><\/pre>\n
PCIe\u62d3\u6251\u7ed3\u6784\u57fa\u51c6\u6d4b\u8bd5<\/strong><\/h2>\n
\u6027\u80fd\u9a8c\u8bc1\u5bf9\u9999\u6e2fGPU\u670d\u52a1\u5668\u79df\u7528\u73af\u5883\u81f3\u5173\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u62d3\u6251\u6d4b\u8bd5\u7684\u7cfb\u7edf\u65b9\u6cd5\uff1a<\/p>\n
\r\n# \u5e26\u5bbd\u6d4b\u8bd5\u811a\u672c\r\n#!\/bin\/bash\r\nfor i in {0..3}; do\r\n    for j in {0..3}; do\r\n        if [ $i -ne $j ]; then\r\n            nvidia-smi topo -p2p r -i $i -j $j\r\n            .\/bandwidth_test --src $i --dst $j\r\n        fi\r\n    done\r\ndone\r\n<\/code><\/pre>\n
\u5e38\u89c1\u95ee\u9898\u6545\u969c\u6392\u9664<\/strong><\/h2>\n
\u5728\u9999\u6e2f\u670d\u52a1\u5668\u6258\u7ba1\u8bbe\u65bd\u90e8\u7f72GPU\u670d\u52a1\u5668\u65f6\uff0c\u9700\u8981\u6ce8\u610f\u8fd9\u4e9bPCIe\u62d3\u6251\u76f8\u5173\u95ee\u9898\uff1a<\/p>\n
\n
PCIe\u94fe\u8def\u8bad\u7ec3\u5931\u8d25\n
\n
\u68c0\u67e5\u7269\u7406\u8fde\u63a5\u5b8c\u6574\u6027<\/li>\n
\u9a8c\u8bc1BIOS PCIe\u4ee3\u6570\u8bbe\u7f6e<\/li>\n
\u76d1\u63a7\u7cfb\u7edf\u4e8b\u4ef6\u65e5\u5fd7<\/li>\n<\/ul>\n<\/li>\n
\u5e26\u5bbd\u964d\u7ea7\n
\n
\u76d1\u63a7PCIe\u94fe\u8def\u5bbd\u5ea6\u534f\u5546<\/li>\n
\u9a8c\u8bc1\u6563\u70ed\u6027\u80fd<\/li>\n
\u68c0\u67e5\u4f9b\u7535\u7a33\u5b9a\u6027<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
GPU\u57fa\u7840\u8bbe\u65bd\u7684\u672a\u6765\u5c55\u671b<\/strong><\/h2>\n
\u5c55\u671b\u9999\u6e2fGPU\u670d\u52a1\u5668\u79df\u7528\u9886\u57df\u7684\u672a\u6765\uff0c\u8bf7\u8003\u8651\u8fd9\u4e9b\u65b0\u5174\u6280\u672f\uff1a<\/p>\n
\n
PCIe Gen5\u5b9e\u73b0\n
\n
\u6bcf\u4e2ax16\u63d2\u69fd63 GB\/s\u7406\u8bba\u5e26\u5bbd<\/li>\n
\u589e\u5f3a\u7684\u9519\u8bef\u68c0\u6d4b\u548c\u7ea0\u6b63<\/li>\n
\u6539\u8fdb\u7684\u7535\u6e90\u7ba1\u7406\u529f\u80fd<\/li>\n<\/ul>\n<\/li>\n
\u8ba1\u7b97\u5feb\u901f\u94fe\u63a5\uff08CXL\uff09\u96c6\u6210\n
\n
\u7f13\u5b58\u4e00\u81f4\u6027\u652f\u6301<\/li>\n
\u5185\u5b58\u6c60\u5316\u529f\u80fd<\/li>\n
\u964d\u4f4eGPU-CPU\u901a\u4fe1\u5ef6\u8fdf<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
\u7ed3\u8bba<\/strong><\/h2>\n
\u5728\u9999\u6e2f\u7684\u670d\u52a1\u5668\u79df\u7528\u73af\u5883\u4e2d\uff0c\u4f18\u5316GPU\u670d\u52a1\u5668\u4e2d\u7684PCIe\u62d3\u6251\u7ed3\u6784\u4ecd\u7136\u662f\u9ad8\u6027\u80fd\u8ba1\u7b97\u7684\u57fa\u7840\u3002\u7406\u89e3PCIe\u901a\u9053\u3001\u5e26\u5bbd\u5206\u914d\u548c\u6563\u70ed\u8003\u8651\u56e0\u7d20\u4e4b\u95f4\u7684\u590d\u6742\u5173\u7cfb\uff0c\u80fd\u591f\u5b9e\u73b0\u6700\u4f73GPU\u670d\u52a1\u5668\u914d\u7f6e\u3002\u968f\u7740\u9999\u6e2f\u7ee7\u7eed\u53d1\u5c55\u6210\u4e3a\u4e3b\u8981\u6570\u636e\u4e2d\u5fc3\u67a2\u7ebd\uff0c\u5b9e\u65bd\u8fd9\u4e9bPCIe\u62d3\u6251\u6700\u4f73\u5b9e\u8df5\u5c06\u786e\u4fddGPU\u670d\u52a1\u5668\u79df\u7528\u90e8\u7f72\u7684\u6700\u5927\u6027\u80fd\u548c\u53ef\u9760\u6027\u3002<\/p>\n
\u5982\u9700\u8fdb\u4e00\u6b65\u534f\u52a9\u4f18\u5316\u9999\u6e2f\u6570\u636e\u4e2d\u5fc3\u7684GPU\u670d\u52a1\u5668PCIe\u62d3\u6251\u7ed3\u6784\uff0c\u8bf7\u8054\u7cfb\u6211\u4eec\u7684\u6280\u672f\u56e2\u961f\u8ba8\u8bba\u60a8\u7684\u5177\u4f53\u670d\u52a1\u5668\u79df\u7528\u9700\u6c42\u3002<\/p>\n
[\/vc_column_text][\/vc_column][\/vc_row]<\/p>\n<\/div>","protected":false},"excerpt":{"rendered":"
[vc_row][vc_column][vc_column_text css=””] \u5728\u9999\u6e2f\u5feb\u901f\u53d1\u5c55\u7684\u6570\u636e\u4e2d\u5fc3\u73af\u5883\u4e2d\uff0c\u7406\u89e3GPU\u670d\u52a1\u5668\u4e2d\u7684PCIe\u62d3\u6251\u7ed3\u6784\u5bf9\u673a\u5668\u5b66\u4e60\u5de5\u7a0b\u5e08\u548c\u7cfb\u7edf\u67b6\u6784\u5e08\u6765\u8bf4\u53d8\u5f97\u81f3\u5173\u91cd\u8981\u3002\u672c\u6280\u672f\u6df1\u5ea6\u6307\u5357\u63a2\u8ba8PCIe\u67b6\u6784\uff0c\u91cd\u70b9\u5173\u6ce8\u9999\u6e2f\u670d\u52a1\u5668\u79df\u7528\u8bbe\u65bd\u4e2dGPU\u52a0\u901f\u8ba1\u7b97\u73af\u5883\u7684\u4f18\u5316\u6280\u672f\u3002 PCIe\u6280\u672f\u57fa\u7840 PCIe\u67b6\u6784\u6784\u6210\u4e86\u73b0\u4ee3GPU\u670d\u52a1\u5668\u7684\u9aa8\u5e72\u3002\u6bcf\u6761PCIe\u901a\u9053\u5728Gen3\u4e0b\u8fd0\u884c\u901f\u5ea6\u4e3a8 GT\/s\uff0cGen4\u4e3a16 GT\/s\uff0cGen5\u4e3a32 GT\/s\uff0c\u7531\u4e8e\u7f16\u7801\u5f00\u9500\uff0c\u5b9e\u9645\u5e26\u5bbd\u7565\u4f4e\u3002\u4f8b\u5982\uff0cPCIe Gen4 x16\u94fe\u8def\u63d0\u4f9b\u7ea631.5 GB\/s\u7684\u7406\u8bba\u5e26\u5bbd\uff1a \u5e26\u5bbd = (\u901a\u9053\u6570 * \u4f20\u8f93\u901f\u7387 * \u7f16\u7801\u6548\u7387) \/ 8 Gen4 x16 = (16 * 16 GT\/s * 0.9878) \/ 8 \u2248 31.5 GB\/s GPU\u670d\u52a1\u5668PCIe\u62d3\u6251\u7ed3\u6784\u67b6\u6784 \u73b0\u4ee3GPU\u670d\u52a1\u5668\u5b9e\u73b0\u4e86\u5404\u79cdPCIe\u62d3\u6251\u8bbe\u8ba1\u3002\u4ee5\u4e0b\u662f\u5e38\u89c1\u67b6\u6784\u7684\u6280\u672f\u7ec6\u5206\uff1a \u76f4\u63a5CPU-GPU\u8fde\u63a5 \u6700\u4f4e\u5ef6\u8fdf\uff08\u4e9a\u5fae\u79d2\u7ea7\uff09 \u6bcf\u4e2aGPU\u90fd\u6709\u5b8c\u6574PCIe\u5e26\u5bbd \u53d7CPU PCIe\u901a\u9053\u6570\u91cf\u9650\u5236 PCIe\u4ea4\u6362\u673a\u5b9e\u73b0 \u63d0\u9ad8GPU\u5bc6\u5ea6 \u5e26\u5bbd\u5171\u4eab\u573a\u666f \u989d\u5916\u5ef6\u8fdf\uff08\u7ea6100ns\uff09 \u5e26\u5bbd\u5206\u6790\u548cGPU\u4e92\u8fde \u5728\u9999\u6e2f\u6570\u636e\u4e2d\u5fc3\u6784\u5efa\u591aGPU\u7cfb\u7edf\u65f6\uff0c\u7406\u89e3\u5e26\u5bbd\u5206\u914d\u81f3\u5173\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u4f7f\u7528\u53ccCPU\u670d\u52a1\u5668\u914d\u7f6e\u7684\u8be6\u7ec6\u5206\u6790\uff1a # \u5e26\u5bbd\u5206\u914d\u793a\u4f8b\uff08\u53ccIntel Xeon\u5e73\u53f0\uff09 CPU1 \u2192 GPU1: PCIe Gen4 […]<\/p>\n
Read More…<\/a><\/p>\n","protected":false},"author":2,"featured_media":21897,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[62],"tags":[6184,6182,6183,6185,4330],"class_list":["post-21903","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-hong-kong-dedicated-server-sc","tag-gpu-interconnect-architecture-sc","tag-gpu-server-pcie-topology-sc","tag-pcie-bandwidth-optimization-sc","tag-data-center-gpu-configuration-sc","tag-hong-kong-gpu-hosting-sc"],"acf":[],"yoast_head":"\n