{"id":174,"date":"2026-09-03T00:13:22","date_gmt":"2026-09-02T16:13:22","guid":{"rendered":"https:\/\/www.xiaoxindong.cloud\/?p=174"},"modified":"2026-09-06T19:50:43","modified_gmt":"2026-09-06T11:50:43","slug":"gpu-%e5%8e%8b%e6%b5%8b%e4%b8%8e%e6%99%ba%e7%ae%97%e7%bd%91%e7%bb%9c","status":"publish","type":"post","link":"https:\/\/www.xiaoxindong.cloud\/?p=174","title":{"rendered":"GPU \u538b\u6d4b\u4e0e\u667a\u7b97\u7f51\u7edc"},"content":{"rendered":"\n

\u672c\u7bc7\u8986\u76d6\u786c\u4ef6\u5168\u94fe\u8def\u538b\u6d4b<\/strong>\u3001\u6a21\u578b\u670d\u52a1\u538b\u6d4b<\/strong>\u3001\u667a\u7b97\u7f51\u7edc\u4e0e\u591a\u673a\u901a\u4fe1\u6d4b\u8bd5<\/strong>\u4e09\u5927\u677f\u5757\uff0c\u68b3\u7406\u6838\u5fc3\u5de5\u5177\u3001\u5173\u952e\u6307\u6807\u4e0e\u6392\u969c\u601d\u8def<\/p>\n\n\n\n

\n\n\n\n

\u4e00\u3001\u538b\u6d4b\u6982\u8ff0<\/h2>\n\n\n\n

1. \u538b\u6d4b\u7684\u6838\u5fc3\u4ef7\u503c<\/h3>\n\n\n\n
    \n
  • \u7a33\u5b9a\u6027\u9a8c\u8bc1<\/strong>\uff1a\u68c0\u6d4b\u9ad8\u8d1f\u8f7d\u4e0b\u786c\u4ef6\u3001\u9a71\u52a8\u3001\u8f6f\u4ef6\u662f\u5426\u51fa\u73b0\u5d29\u6e83\u3001\u964d\u9891\u3001\u6570\u636e\u9519\u8bef<\/li>\n\n\n\n
  • \u9690\u60a3\u63d0\u524d\u53d1\u73b0<\/strong>\uff1a\u63d0\u524d\u66b4\u9732\u6563\u70ed\u3001\u4f9b\u7535\u3001\u663e\u5b58\u3001\u94fe\u8def\u7b49\u9690\u6027\u786c\u4ef6\u95ee\u9898<\/li>\n\n\n\n
  • \u6027\u80fd\u57fa\u51c6\u8bc4\u4f30<\/strong>\uff1a\u83b7\u53d6\u6781\u9650\u6027\u80fd\u6307\u6807\uff0c\u652f\u6491\u4e0d\u540c\u914d\u7f6e\u7684\u6a2a\u5411\u5bf9\u6bd4<\/li>\n\n\n\n
  • \u5bb9\u91cf\u89c4\u5212\u4f9d\u636e<\/strong>\uff1a\u4e3a\u4e1a\u52a1\u4e0a\u7ebf\u3001\u8d44\u6e90\u8c03\u5ea6\u3001\u53c2\u6570\u4f18\u5316\u63d0\u4f9b\u6570\u636e\u652f\u6491<\/li>\n<\/ul>\n\n\n\n

    2. \u4e03\u5c42\u538b\u6d4b\u4f53\u7cfb<\/h3>\n\n\n\n

    \u4ece\u5e95\u5c42\u786c\u4ef6\u5230\u4e0a\u5c42\u670d\u52a1\uff0c\u4ece\u5355\u673a\u5230\u96c6\u7fa4\uff0c\u5f62\u6210\u5b8c\u6574\u9a8c\u8bc1\u95ed\u73af\uff1a<\/p>\n\n\n\n

    \u538b\u6d4b\u7c7b\u522b<\/th>\u9a8c\u8bc1\u76ee\u6807<\/th>\u6838\u5fc3\u5173\u6ce8\u6307\u6807<\/th>\u5e38\u7528\u5de5\u5177<\/th><\/tr><\/thead>
    CPU \u538b\u6d4b<\/td>\u8ba1\u7b97\u80fd\u529b\u3001\u591a\u6838\u8c03\u5ea6\u3001\u6563\u70ed\u7a33\u5b9a\u6027<\/td>\u5229\u7528\u7387\u3001load\u3001\u6e29\u5ea6\u3001bogo ops<\/td>stress-ng<\/td><\/tr>
    \u5185\u5b58\u538b\u6d4b<\/td>\u5bb9\u91cf\u3001\u5e26\u5bbd\u3001ECC \u7a33\u5b9a\u6027\u3001\u9ad8\u8d1f\u8f7d\u53ef\u9760\u6027<\/td>\u5185\u5b58\u5360\u7528\u3001\u5e26\u5bbd\u3001ECC \u9519\u8bef\u3001OOM<\/td>stress-ng \/ memtester<\/td><\/tr>
    \u78c1\u76d8\u538b\u6d4b<\/td>\u987a\u5e8f \/ \u968f\u673a\u8bfb\u5199\u6027\u80fd\u3001IO \u7a33\u5b9a\u6027<\/td>IOPS\u3001\u541e\u5410\u3001\u65f6\u5ef6\u3001\u961f\u5217\u957f\u5ea6<\/td>fio \/ iostat<\/td><\/tr>
    \u7f51\u7edc\u538b\u6d4b<\/td>\u5e26\u5bbd\u3001\u65f6\u5ef6\u3001\u4e22\u5305\u3001\u6296\u52a8<\/td>\u5e26\u5bbd\u3001\u5ef6\u8fdf\u3001\u4e22\u5305\u7387<\/td>iperf3 \/ ping<\/td><\/tr>
    GPU \u538b\u6d4b<\/td>\u7b97\u529b\u3001\u663e\u5b58\u3001\u6e29\u5ea6\u3001\u529f\u8017\u3001\u6ee1\u8f7d\u7a33\u5b9a\u6027<\/td>\u5229\u7528\u7387\u3001\u663e\u5b58\u3001\u6e29\u5ea6\u3001ECC\/XID \u9519\u8bef<\/td>gpu-burn \/ DCGM<\/td><\/tr>
    \u6a21\u578b\u538b\u6d4b<\/td>\u670d\u52a1\u5e76\u53d1\u80fd\u529b\u3001\u54cd\u5e94\u65f6\u5ef6\u3001Token \u541e\u5410<\/td>QPS\u3001TPS\u3001TTFT\u3001P99 \u5ef6\u8fdf<\/td>vLLM benchmark \/ wrk<\/td><\/tr>
    \u591a\u673a\u901a\u4fe1\u538b\u6d4b<\/td>\u591a\u5361 \/ \u591a\u673a\u901a\u4fe1\u5e26\u5bbd\u4e0e\u65f6\u5ef6<\/td>AllReduce \u5e26\u5bbd\u3001\u901a\u4fe1\u5ef6\u8fdf<\/td>nccl-tests \/ ib_write_bw<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

    \u6838\u5fc3\u601d\u8def<\/strong>\uff1a\u5148\u6d4b\u786c\u4ef6\u8d44\u6e90\uff0c\u518d\u6d4b GPU \u4e0e\u6a21\u578b\uff0c\u6700\u540e\u9a8c\u8bc1\u591a\u673a\u901a\u4fe1\u3002<\/p>\n\n\n\n

    \n\n\n\n

    \u4e8c\u3001\u57fa\u7840\u786c\u4ef6\u538b\u6d4b\uff08\u719f\u7ec3\u638c\u63e1\uff09<\/h2>\n\n\n\n

    1. CPU \u538b\u6d4b<\/h3>\n\n\n\n

    \u5de5\u5177\uff1astress-ng<\/h4>\n\n\n\n
    # Ubuntu\u5b89\u88c5\napt install -y stress-ng\n<\/code><\/pre>\n\n\n\n
    \u5e38\u7528\u547d\u4ee4<\/h4>\n\n\n\n
    # \u5168\u6838\u5fc3\u538b\u6d4b30\u79d2\uff0c\u8f93\u51fa\u7b80\u8981\u7edf\u8ba1\nstress-ng --cpu 0 --timeout 30s --metrics-brief\n<\/code><\/pre>\n\n\n\n
      \n
    • --cpu 0<\/code>\uff1a\u81ea\u52a8\u4f7f\u7528\u6240\u6709\u903b\u8f91 CPU \u6838\u5fc3<\/li>\n\n\n\n
    • --timeout 30s<\/code>\uff1a\u538b\u6d4b\u6301\u7eed\u65f6\u957f<\/li>\n\n\n\n
    • --metrics-brief<\/code>\uff1a\u8f93\u51fa\u7cbe\u7b80\u6027\u80fd\u7edf\u8ba1<\/li>\n<\/ul>\n\n\n\n
      \u7ed3\u679c\u5173\u952e\u9879<\/h4>\n\n\n\n
        \n
      • bogo ops\/s<\/code>\uff1a\u6bcf\u79d2\u5b8c\u6210\u7684\u8ba1\u7b97\u64cd\u4f5c\u6570\uff0c\u8861\u91cf CPU \u8ba1\u7b97\u6027\u80fd<\/li>\n\n\n\n
      • usr time<\/code> \/ sys time<\/code>\uff1a\u7528\u6237\u6001 \/ \u5185\u6838\u6001\u8017\u65f6<\/li>\n\n\n\n
      • successful run completed<\/code>\uff1a\u65e0\u5f02\u5e38\u9000\u51fa\u5219\u6d4b\u8bd5\u901a\u8fc7<\/li>\n<\/ul>\n\n\n\n
        2. \u5185\u5b58\u538b\u6d4b<\/h3>\n\n\n\n
        \u6838\u5fc3\u76ee\u6807<\/h4>\n\n\n\n
        \u4e0d\u53ea\u662f\u5360\u6ee1\u5185\u5b58\uff0c\u91cd\u70b9\u9a8c\u8bc1\u9ad8\u5360\u7528\u4e0b\u7684\u7a33\u5b9a\u6027<\/strong>\uff0c\u6392\u67e5 ECC \u9519\u8bef\u3001OOM\u3001\u5f02\u5e38\u91cd\u542f\u3001Swap \u6296\u52a8\u3002<\/p>\n\n\n\n
        \u5e38\u7528\u547d\u4ee4<\/h4>\n\n\n\n
        # 8\u4e2a\u7ebf\u7a0b\uff0c\u6bcf\u7ebf\u7a0b\u7533\u8bf74G\u5185\u5b58\nstress-ng --vm 8 --vm-bytes 4G --timeout 30s\n\n# \u5360\u7528\u7cfb\u7edf80%\u5185\u5b58\nstress-ng --vm 4 --vm-bytes 80% --timeout 30s\n\n# \u968f\u673a\u5185\u5b58\u8bbf\u95ee\u6a21\u5f0f\uff0c\u66f4\u6613\u6392\u67e5\u7a33\u5b9a\u6027\u95ee\u9898\nstress-ng --vm 4 --vm-bytes 80% --vm-method all --timeout 30s\n<\/code><\/pre>\n\n\n\n
        \u5173\u6ce8\u91cd\u70b9<\/h4>\n\n\n\n
        \u5185\u5b58\u4f7f\u7528\u7387\u3001Swap \u4f7f\u7528\u91cf\u3001ECC \u9519\u8bef\u8ba1\u6570\u3001\u7cfb\u7edf\u65e5\u5fd7\u3001\u662f\u5426\u51fa\u73b0\u5361\u987f\u6216\u5d29\u6e83\u3002<\/p>\n\n\n\n
        3. \u78c1\u76d8\u538b\u6d4b<\/h3>\n\n\n\n
        \u5de5\u5177\uff1afio<\/h4>\n\n\n\n
        apt install fio -y\n<\/code><\/pre>\n\n\n\n
        \u987a\u5e8f\u5199\u6d4b\u8bd5\u793a\u4f8b<\/h4>\n\n\n\n
        fio --name=seq_write \\\n  --filename=\/data\/testfile \\\n  --size=10G \\\n  --bs=1M \\\n  --rw=write \\\n  --direct=1 \\\n  --numjobs=1 \\\n  --runtime=60 \\\n  --time_based\n<\/code><\/pre>\n\n\n\n
        \u6838\u5fc3\u5173\u6ce8\u6307\u6807<\/h4>\n\n\n\n
          \n
        • IOPS<\/strong>\uff1a\u6bcf\u79d2\u8bfb\u5199\u64cd\u4f5c\u6b21\u6570<\/li>\n\n\n\n
        • BW\uff08\u5e26\u5bbd\uff09<\/strong>\uff1a\u6bcf\u79d2\u6570\u636e\u541e\u5410\u91cf<\/li>\n\n\n\n
        • lat\uff08\u65f6\u5ef6\uff09<\/strong>\uff1a\u5e73\u5747 \/ \u6700\u5927 IO \u5ef6\u8fdf<\/li>\n\n\n\n
        • \u767e\u5206\u4f4d\u5ef6\u8fdf<\/strong>\uff1aP95\/P99 \u5ef6\u8fdf\uff0c\u53cd\u6620\u6781\u7aef\u4f53\u9a8c<\/li>\n<\/ul>\n\n\n\n
          \u5b9e\u65f6\u72b6\u6001\u67e5\u770b\uff1aiostat<\/h4>\n\n\n\n
          iostat -x 1\n<\/code><\/pre>\n\n\n\n
          \u91cd\u70b9\u5173\u6ce8\uff1a<\/p>\n\n\n\n
            \n
          • %util<\/code>\uff1a\u78c1\u76d8\u7e41\u5fd9\u7a0b\u5ea6<\/li>\n\n\n\n
          • r_await<\/code> \/ w_await<\/code>\uff1a\u8bfb \/ \u5199\u8bf7\u6c42\u5e73\u5747\u7b49\u5f85\u65f6\u95f4<\/li>\n\n\n\n
          • r\/s<\/code>\u3001w\/s<\/code>\uff1a\u6bcf\u79d2\u8bfb \/ \u5199\u6b21\u6570<\/li>\n\n\n\n
          • rkB\/s<\/code>\u3001wkB\/s<\/code>\uff1a\u6bcf\u79d2\u8bfb \/ \u5199\u6570\u636e\u91cf<\/li>\n<\/ul>\n\n\n\n
            \n\n\n\n
            \u4e09\u3001GPU \u786c\u4ef6\u538b\u6d4b<\/h2>\n\n\n\n
            1. gpu-burn \u538b\u529b\u6d4b\u8bd5\u5de5\u5177<\/h3>\n\n\n\n
            \u5de5\u4f5c\u539f\u7406<\/h4>\n\n\n\n
            \u57fa\u4e8e CUDA \u5f00\u53d1\u7684 GPU \u538b\u529b\u6d4b\u8bd5\u5de5\u5177\uff0c\u901a\u8fc7\u8c03\u7528 CUDA Core \u6267\u884c\u5927\u91cf\u77e9\u9635\u4e58\u6cd5\u4e0e\u6d6e\u70b9\u8fd0\u7b97<\/strong>\uff0c\u8ba9 GPU \u63a5\u8fd1\u6ee1\u8d1f\u8f7d\u72b6\u6001\uff0c\u540c\u65f6\u6301\u7eed\u6821\u9a8c\u8ba1\u7b97\u7ed3\u679c\u3002\u82e5\u51fa\u73b0\u8ba1\u7b97\u9519\u8bef\uff08errors>0\uff09\uff0c\u53ef\u5b9a\u4f4d\u663e\u5b58\u9519\u8bef\u3001GPU \u6838\u5fc3\u5f02\u5e38\u3001\u4f9b\u7535\u4e0d\u8db3\u3001\u6563\u70ed\u4e0d\u826f\u7b49\u786c\u4ef6\u7a33\u5b9a\u6027\u95ee\u9898\u3002<\/p>\n\n\n\n
            \u5b89\u88c5\u7f16\u8bd1<\/h4>\n\n\n\n
            # \u89e3\u538b\u6e90\u7801\u5305\nunzip gpu-burn-master.zip\n# \u5b89\u88c5\u7f16\u8bd1\u4f9d\u8d56\napt install make\n# \u7f16\u8bd1\uff08\u6307\u5b9aGPU\u7b97\u529b\u67b6\u6784\uff0cV100\u5bf9\u5e94compute_70\uff09\nmake clean && make COMPUTE=70\n<\/code><\/pre>\n\n\n\n
            \u5e38\u7528\u7528\u6cd5<\/h4>\n\n\n\n
            # \u5217\u51fa\u7cfb\u7edf\u6240\u6709GPU\u8bbe\u5907\n.\/gpu_burn -l\n\n# \u5168\u5361\u53cc\u7cbe\u5ea6\u6d4b\u8bd51\u5c0f\u65f6\n.\/gpu_burn -d 3600\n\n# \u6307\u5b9aGPU 0\uff0c\u4f7f\u752880%\u663e\u5b58\uff0c\u6d4b\u8bd5300\u79d2\n.\/gpu_burn -i 0 -d -m 80% 300\n<\/code><\/pre>\n\n\n\n
            \u7ed3\u679c\u5206\u6790<\/h4>\n\n\n\n
            \u751f\u4ea7\u9a8c\u6536\u901a\u5e38\u538b\u6d4b 4~6 \u5c0f\u65f6\uff0c\u91cd\u70b9\u5173\u6ce8\uff1a<\/p>\n\n\n\n
              \n
            1. errors<\/strong>\uff1a\u5fc5\u987b\u4e3a 0\uff1b\u51fa\u73b0\u9519\u8bef\u8bf4\u660e\u663e\u5b58\u3001\u6838\u5fc3\u6216\u4f9b\u7535\u5b58\u5728\u786c\u4ef6\u95ee\u9898<\/li>\n\n\n\n
            2. Gflop\/s<\/strong>\uff1a\u6bcf\u79d2\u6d6e\u70b9\u8fd0\u7b97\u901f\u5ea6\uff0c\u9700\u7a33\u5b9a\u65e0\u5927\u5e45\u6ce2\u52a8<\/li>\n\n\n\n
            3. temps<\/strong>\uff1aGPU \u6838\u5fc3\u6e29\u5ea6\uff0c\u89c2\u5bdf\u662f\u5426\u8fc7\u70ed\u89e6\u53d1\u964d\u9891<\/li>\n\n\n\n
            4. \u663e\u5b58\u5360\u7528\u3001\u529f\u8017\u3001\u98ce\u6247\u72b6\u6001\u662f\u5426\u6b63\u5e38<\/li>\n<\/ol>\n\n\n\n
              2. CUDA Samples \u9a8c\u8bc1\u5de5\u5177<\/h3>\n\n\n\n
              NVIDIA \u5b98\u65b9\u63d0\u4f9b\u7684 CUDA \u793a\u4f8b\u7a0b\u5e8f\u96c6\uff0c\u7528\u4e8e\u73af\u5883\u9a8c\u8bc1\u3001\u57fa\u7840\u529f\u80fd\u4e0e\u5e26\u5bbd\u6d4b\u8bd5\u3002<\/p>\n\n\n\n
              2.1 deviceQuery<\/h4>\n\n\n\n
                \n
              • \u4f5c\u7528<\/strong>\uff1a\u679a\u4e3e\u5e76\u663e\u793a GPU \u786c\u4ef6\u5c5e\u6027\uff0c\u9a8c\u8bc1\u9a71\u52a8\u4e0e CUDA \u73af\u5883\u662f\u5426\u6b63\u5e38<\/li>\n\n\n\n
              • \u5173\u952e\u8f93\u51fa<\/strong>\uff1aGPU \u578b\u53f7\u3001CUDA \u7b97\u529b\u7248\u672c\u3001SM \u6570\u91cf\u3001\u663e\u5b58\u5bb9\u91cf\u3001Warp \u5927\u5c0f\u3001ECC \u652f\u6301\u3001PCIe \u4fe1\u606f<\/li>\n\n\n\n
              • \u573a\u666f<\/strong>\uff1a\u9a71\u52a8 \/ CUDA \u5b89\u88c5\u540e\u7684\u7b2c\u4e00\u9879\u9a8c\u8bc1\uff0cGPU \u7248 \u201c\u786c\u4ef6\u81ea\u68c0\u5de5\u5177\u201d<\/li>\n<\/ul>\n\n\n\n
                2.2 bandwidthTest\u3010\u91cd\u70b9\u3011<\/h4>\n\n\n\n
                  \n
                • \u4f5c\u7528<\/strong>\uff1a\u6d4b\u8bd5\u4e09\u7ea7\u6570\u636e\u4f20\u8f93\u5e26\u5bbd\n
                    \n
                  1. Host\u2192Device<\/strong>\uff1a\u4e3b\u673a\u5185\u5b58\u2192GPU \u663e\u5b58\uff08PCIe \u5e26\u5bbd\uff09<\/li>\n\n\n\n
                  2. Device\u2192Host<\/strong>\uff1aGPU \u663e\u5b58\u2192\u4e3b\u673a\u5185\u5b58<\/li>\n\n\n\n
                  3. Device\u2192Device<\/strong>\uff1aGPU \u663e\u5b58\u5185\u90e8\u62f7\u8d1d\uff08\u663e\u5b58\u5e26\u5bbd\uff09<\/li>\n<\/ol>\n<\/li>\n\n\n\n
                  4. \u5178\u578b\u53c2\u8003\uff08Tesla V100\uff09<\/strong>\uff1a\n
                      \n
                    • Host\u2194Device\uff1a\u7ea6 12~25 GB\/s\uff08PCIe Gen3 x16\uff09<\/li>\n\n\n\n
                    • Device\u2192Device\uff1a\u7ea6 700~900 GB\/s\uff08HBM2 \u663e\u5b58\uff09<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
                      2.3 p2pBandwidthLatencyTest<\/h4>\n\n\n\n
                        \n
                      • \u4f5c\u7528<\/strong>\uff1a\u6d4b\u8bd5\u591a\u5f20 GPU \u4e4b\u95f4 P2P\uff08Peer-to-Peer\uff09\u901a\u4fe1\u7684\u5e26\u5bbd\u4e0e\u5ef6\u8fdf<\/li>\n\n\n\n
                      • \u6838\u5fc3\u5224\u65ad<\/strong>\uff1a\n
                          \n
                        • NVLink \u76f4\u8fde\uff1a\u5e26\u5bbd\u6700\u9ad8\u3001\u5ef6\u8fdf\u6700\u4f4e<\/li>\n\n\n\n
                        • \u540c PCIe Switch\uff1a\u53ef P2P\uff0c\u6027\u80fd\u6b21\u4e4b<\/li>\n\n\n\n
                        • \u8de8 CPU\/PHB\uff1a\u5e26\u5bbd\u4f4e\u3001\u5ef6\u8fdf\u9ad8<\/li>\n<\/ul>\n<\/li>\n\n\n\n
                        • \u573a\u666f<\/strong>\uff1a\u591a\u5361\u670d\u52a1\u5668\u5fc5\u6d4b\uff0c\u9a8c\u8bc1 NVLink\/PCIe \u62d3\u6251\u4e92\u8054\u8d28\u91cf<\/li>\n<\/ul>\n\n\n\n
                          3. GPU \u670d\u52a1\u5668\u9a8c\u6536\u6807\u51c6\u6d41\u7a0b\u3010\u9762\u8bd5\u5e38\u95ee\u3011<\/h3>\n\n\n\n
                          \u786c\u4ef6\u4e0eBMC\u68c0\u67e5\n    \u2193\nnvidia-smi\u8bc6\u522b\u68c0\u67e5\n    \u2193\nPCIe\u94fe\u8def\u4e0e\u62d3\u6251\u68c0\u67e5\n    \u2193\nCUDA Samples\u57fa\u7840\u9a8c\u8bc1\n    \u2193\nNV\u5e26\u5bbd\u6d4b\u8bd5\n    \u2193\nP2P\u4e0eNVLink\u68c0\u67e5\n    \u2193\nNCCL\u96c6\u5408\u901a\u4fe1\u6d4b\u8bd5\n    \u2193\nDCGM\u5065\u5eb7\u8bca\u65ad\n    \u2193\nGPU\u957f\u65f6\u95f4\u538b\u529b\u6d4b\u8bd5\n    \u2193\n\u6a21\u578b\u6216\u4e1a\u52a1\u538b\u6d4b\n<\/code><\/pre>\n\n\n\n
                          \n\n\n\n
                          \u56db\u3001\u5927\u6a21\u578b\u670d\u52a1\u538b\u6d4b<\/h2>\n\n\n\n
                          1. \u4e3a\u4ec0\u4e48\u505a\u6a21\u578b\u538b\u6d4b<\/h3>\n\n\n\n
                          \u9a8c\u8bc1\u6a21\u578b\u670d\u52a1\u5728\u771f\u5b9e\u8bf7\u6c42\u4e0b\u7684\u5e76\u53d1\u627f\u8f7d\u80fd\u529b\u3001\u54cd\u5e94\u65f6\u5ef6\u3001Token \u541e\u5410\u4e0e\u7a33\u5b9a\u6027<\/strong>\uff0c\u63d0\u524d\u53d1\u73b0\u6392\u961f\u3001\u8d85\u65f6\u3001OOM\u3001\u541e\u5410\u74f6\u9888\uff0c\u4e3a\u4e0a\u7ebf\u5bb9\u91cf\u89c4\u5212\u4e0e\u53c2\u6570\u4f18\u5316\u63d0\u4f9b\u4f9d\u636e\u3002<\/p>\n\n\n\n
                          2. \u5de5\u5177\uff1avLLM benchmark<\/h3>\n\n\n\n
                          \u538b\u6d4b\u547d\u4ee4\u793a\u4f8b<\/h4>\n\n\n\n
                          vllm bench serve \\\n  --backend openai-chat \\\n  --host 127.0.0.1 \\\n  --port 8000 \\\n  --endpoint \/v1\/chat\/completions \\\n  --model Qwen\/Qwen2-1.5B-Instruct \\\n  --dataset-name random \\\n  --num-prompts 500 \\\n  --random-input-len 512 \\\n  --random-output-len 128 \\\n  --max-concurrency 1 \\\n  --temperature 0\n<\/code><\/pre>\n\n\n\n
                          \u6838\u5fc3\u53c2\u6570<\/h4>\n\n\n\n
                          \u53c2\u6570<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
                          --backend<\/code><\/td>\u63a5\u53e3\u7c7b\u578b\uff0c\u5982 openai-chat<\/td><\/tr>
                          --num-prompts<\/code><\/td>\u603b\u6d4b\u8bd5\u8bf7\u6c42\u6570\u91cf<\/td><\/tr>
                          --random-input-len<\/code><\/td>\u8f93\u5165 Prompt \u957f\u5ea6\uff08Token\uff09<\/td><\/tr>
                          --random-output-len<\/code><\/td>\u8f93\u51fa Token \u957f\u5ea6<\/td><\/tr>
                          --max-concurrency<\/code><\/td>\u6700\u5927\u5e76\u53d1\u8bf7\u6c42\u6570<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                          3. \u6838\u5fc3\u6307\u6807\u3010\u9762\u8bd5\u5e38\u95ee\u3011<\/h3>\n\n\n\n
                          \u6307\u6807<\/th>\u5168\u79f0<\/th>\u542b\u4e49<\/th><\/tr><\/thead>
                          Request Throughput (QPS)<\/td>\u6bcf\u79d2\u8bf7\u6c42\u6570<\/td>\u6bcf\u79d2\u5904\u7406\u7684\u7528\u6237\u8bf7\u6c42\u6570\u91cf<\/td><\/tr>
                          Output Token Throughput (TPS)<\/td>\u6bcf\u79d2 Token \u6570<\/td>\u6bcf\u79d2\u751f\u6210\u7684\u8f93\u51fa Token \u6570\u91cf<\/td><\/tr>
                          Mean TTFT<\/td>Time To First Token<\/td>\u9996 Token \u5ef6\u8fdf\uff0c\u4ece\u8bf7\u6c42\u53d1\u8d77\u5230\u7b2c\u4e00\u4e2a Token \u8fd4\u56de\u7684\u65f6\u95f4<\/td><\/tr>
                          Mean TPOT \/ ITL<\/td>Time Per Output Token<\/td>\u6bcf\u751f\u6210\u4e00\u4e2a Token \u7684\u5e73\u5747\u95f4\u9694\u65f6\u95f4<\/td><\/tr>
                          Failed Requests<\/td>\u5931\u8d25\u8bf7\u6c42\u6570<\/td>\u6b63\u5e38\u5fc5\u987b\u4e3a 0<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                          \n\n\n\n
                          \u4e94\u3001AI \u667a\u7b97\u7f51\u7edc\u4e0e\u591a\u673a\u901a\u4fe1<\/h2>\n\n\n\n
                          1. RDMA \u57fa\u7840<\/h3>\n\n\n\n
                          \u5b9a\u4e49<\/h4>\n\n\n\n
                          RDMA\uff08Remote Direct Memory Access\uff09\uff1a\u8ba9\u4e00\u53f0\u670d\u52a1\u5668\u7684\u7f51\u5361\u76f4\u63a5\u8bbf\u95ee\u53e6\u4e00\u53f0\u670d\u52a1\u5668\u7684\u5185\u5b58\uff0c\u7ed5\u8fc7\u64cd\u4f5c\u7cfb\u7edf\u5185\u6838\u4e0e CPU<\/strong>\uff0c\u5b9e\u73b0\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u5e26\u5bbd\u3001\u4f4e CPU \u5360\u7528\u7684\u6570\u636e\u4f20\u8f93\u3002<\/p>\n\n\n\n
                          \u4e0e\u4f20\u7edf TCP\/IP \u5bf9\u6bd4<\/h4>\n\n\n\n
                          \u7ef4\u5ea6<\/th>\u4f20\u7edf TCP\/IP<\/th>RDMA<\/th><\/tr><\/thead>
                          \u6570\u636e\u8def\u5f84<\/td>\u7ecf\u5185\u6838\u534f\u8bae\u6808\u3001\u591a\u6b21\u5185\u5b58\u62f7\u8d1d\u3001CPU \u53c2\u4e0e\u642c\u8fd0<\/td>\u5185\u6838\u65c1\u8def\u3001\u96f6\u62f7\u8d1d\u3001\u7f51\u5361\u786c\u4ef6\u76f4\u63a5\u642c\u8fd0<\/td><\/tr>
                          \u5ef6\u8fdf<\/td>\u8f83\u9ad8<\/td>\u6781\u4f4e\uff08\u5fae\u79d2\u7ea7\uff09<\/td><\/tr>
                          CPU \u5360\u7528<\/td>\u9ad8<\/td>\u6781\u4f4e<\/td><\/tr>
                          \u9002\u7528\u573a\u666f<\/td>\u901a\u7528\u4e1a\u52a1\u7f51\u7edc<\/td>\u9ad8\u6027\u80fd\u8ba1\u7b97\u3001AI \u5206\u5e03\u5f0f\u8bad\u7ec3<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                          GPU Direct RDMA<\/h4>\n\n\n\n
                          \u914d\u5408 GPU Direct \u6280\u672f\uff0c\u7f51\u5361\u53ef\u76f4\u63a5\u8bbf\u95ee GPU \u663e\u5b58\uff0c\u6570\u636e\u5728\u7f51\u5361\u4e0e GPU \u663e\u5b58\u95f4\u76f4\u63a5\u4f20\u8f93\uff0c\u5c3d\u91cf\u7ed5\u8fc7 CPU \u548c\u7cfb\u7edf\u5185\u5b58\uff0c\u5927\u5e45\u63d0\u5347\u591a\u673a GPU \u901a\u4fe1\u6548\u7387\u3002<\/p>\n\n\n\n
                          2. RDMA \u4e3b\u6d41\u5b9e\u73b0\u65b9\u5f0f\u5bf9\u6bd4<\/h3>\n\n\n\n
                          \u5bf9\u6bd4\u9879<\/th>InfiniBand (IB)<\/th>RoCE v2<\/th><\/tr><\/thead>
                          \u5e95\u5c42\u7f51\u7edc<\/td>IB \u4e13\u7528\u7f51\u7edc<\/td>\u6807\u51c6\u4ee5\u592a\u7f51<\/td><\/tr>
                          \u662f\u5426\u57fa\u4e8e IP<\/td>\u5426<\/td>\u662f\uff08UDP\/IP \u5c01\u88c5\uff09<\/td><\/tr>
                          \u5ef6\u8fdf<\/td>\u6781\u4f4e<\/td>\u8f83\u4f4e\uff08\u7565\u9ad8\u4e8e IB\uff09<\/td><\/tr>
                          \u541e\u5410\u91cf<\/td>\u6781\u9ad8<\/td>\u9ad8<\/td><\/tr>
                          \u65e0\u635f\u4fdd\u8bc1<\/td>\u539f\u751f\u65e0\u635f<\/td>\u9700\u8981 PFC+ECN \u7b49\u7f51\u7edc\u914d\u7f6e\u4fdd\u969c<\/td><\/tr>
                          \u6210\u672c<\/td>\u9ad8\uff08\u4e13\u7528\u7f51\u5361 + \u4ea4\u6362\u673a\uff09<\/td>\u4f4e\uff08\u590d\u7528\u73b0\u6709\u4ee5\u592a\u7f51\u8bbe\u5907\uff09<\/td><\/tr>
                          \u9002\u7528\u573a\u666f<\/td>\u8d85\u5927\u89c4\u6a21 AI \u96c6\u7fa4\u3001HPC\u3001\u6781\u81f4\u6027\u80fd<\/td>\u5927\u591a\u6570 AI \u8bad\u7ec3\u96c6\u7fa4\u3001\u4e91\u73af\u5883\u3001\u6210\u672c\u654f\u611f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                          3. InfiniBand \u5173\u952e\u7ec4\u4ef6\u4e0e\u8fd0\u7ef4<\/h3>\n\n\n\n
                          \u6838\u5fc3\u7ec4\u4ef6<\/h4>\n\n\n\n
                            \n
                          • HCA \u5361<\/strong>\uff1a\u4e3b\u673a\u4fa7 IB \u9002\u914d\u5668\uff08\u5982 NVIDIA ConnectX \u7cfb\u5217\uff09\uff0c\u8d1f\u8d23 RDMA \u901a\u4fe1<\/li>\n\n\n\n
                          • IB \u4ea4\u6362\u673a<\/strong>\uff1a\u4e13\u7528 IB \u4ea4\u6362\u8bbe\u5907<\/li>\n\n\n\n
                          • Subnet Manager (SM)<\/strong>\uff1a\u5b50\u7f51\u7ba1\u7406\u5668\uff0c\u8d1f\u8d23 IB \u7f51\u7edc\u7684\u8def\u5f84\u8ba1\u7b97\u3001LID \u5206\u914d\u4e0e\u62d3\u6251\u7ba1\u7406<\/li>\n<\/ul>\n\n\n\n
                            \u5e38\u7528\u547d\u4ee4<\/h4>\n\n\n\n
                            # \u5b89\u88c5\u5de5\u5177\u5305\napt install infiniband-diags opensm\n\n# \u52a0\u8f7d\u9a71\u52a8\u6a21\u5757\nmodprobe mlx4_ib\nmodprobe ib_umad\n\n# \u67e5\u770bIB\u7f51\u5361\u72b6\u6001\nibstat\nibstatus\n\n# \u67e5\u770bRDMA\u8bbe\u5907\u4fe1\u606f\nibv_devices\nibv_devinfo\n<\/code><\/pre>\n\n\n\n
                            \n
                            \u3010\u5e38\u89c1\u6392\u969c\u3011IB \u7269\u7406\u94fe\u8def LinkUp \u4f46\u72b6\u6001 Initializing\uff1a\u901a\u5e38\u662f\u7f3a\u5c11 Subnet Manager\uff0c\u542f\u52a8 opensm \u670d\u52a1\u5373\u53ef\u3002<\/p>\n<\/blockquote>\n\n\n\n
                            4. \u591a\u673a RDMA \u5e26\u5bbd\u6d4b\u8bd5\uff1aib_write_bw<\/h3>\n\n\n\n
                            # \u670d\u52a1\u7aef\u6267\u884c\nib_write_bw -d mlx4_0 -i 1 -D 60 -s 65536 --report_gbits\n\n# \u5ba2\u6237\u7aef\u6267\u884c\nib_write_bw -d mlx4_0 -i 1 -D 60 -s 65536 --report_gbits \u670d\u52a1\u7aefIP\n<\/code><\/pre>\n\n\n\n
                              \n
                            • \u6838\u5fc3\u6307\u6807\uff1aBW average<\/code>\uff0c\u5373\u5e73\u5747\u5199\u5165\u5e26\u5bbd\uff08Gb\/s\uff09<\/li>\n<\/ul>\n\n\n\n
                              5. \u591a\u673a\u591a\u5361\u96c6\u5408\u901a\u4fe1\u6d4b\u8bd5\uff1aNCCL<\/h3>\n\n\n\n
                              \u96c6\u5408\u901a\u4fe1\u6838\u5fc3\u64cd\u4f5c<\/h4>\n\n\n\n
                              \u64cd\u4f5c<\/th>\u8bf4\u660e<\/th>\u5178\u578b\u573a\u666f<\/th><\/tr><\/thead>
                              AllReduce<\/strong><\/td>\u6240\u6709 GPU \u6570\u636e\u6c47\u603b\u8ba1\u7b97\uff0c\u7ed3\u679c\u540c\u6b65\u56de\u6240\u6709 GPU<\/td>\u5206\u5e03\u5f0f\u8bad\u7ec3\u68af\u5ea6\u540c\u6b65\uff08\u6700\u6838\u5fc3\uff09<\/td><\/tr>
                              AllGather<\/td>\u6bcf\u5f20\u5361\u8d21\u732e\u90e8\u5206\u6570\u636e\uff0c\u6700\u7ec8\u6bcf\u5f20\u5361\u90fd\u5f97\u5230\u5b8c\u6574\u6570\u636e<\/td>\u5f20\u91cf\u5e76\u884c\u7ed3\u679c\u62fc\u63a5<\/td><\/tr>
                              Reduce<\/td>\u6240\u6709\u5361\u6570\u636e\u6c47\u603b\u5230\u5355\u5f20\u5361<\/td>\u7ed3\u679c\u6536\u96c6<\/td><\/tr>
                              Broadcast<\/td>\u5355\u5f20\u5361\u6570\u636e\u5e7f\u64ad\u5230\u6240\u6709\u5361<\/td>\u53c2\u6570\u521d\u59cb\u5316\u3001\u4e3b\u8282\u70b9\u540c\u6b65<\/td><\/tr>
                              ReduceScatter<\/td>\u5148\u6c47\u603b\u518d\u62c6\u5206\u5230\u5404\u5361<\/td>\u5927\u6a21\u578b\u8bad\u7ec3\uff0c\u964d\u4f4e\u663e\u5b58\u538b\u529b<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
                              nccl-tests \u5de5\u5177\u4f7f\u7528<\/h4>\n\n\n\n
                              \u5355\u673a\u6d4b\u8bd5<\/h5>\n\n\n\n
                              # \u7f16\u8bd1\nmake -j$(nproc) CUDA_HOME=\/usr\/local\/cuda NCCL_HOME=\/usr\n\n# \u6d4b\u8bd5AllReduce\u6027\u80fd\n.\/build\/all_reduce_perf -b 8M -e 1G -f 2 -g 2\n<\/code><\/pre>\n\n\n\n
                              \u591a\u673a\u6d4b\u8bd5<\/h5>\n\n\n\n
                              mpirun \\\n  --allow-run-as-root \\\n  -np 2 \\\n  -H 192.168.32.50:1,192.168.32.72:1 \\\n  -x CUDA_VISIBLE_DEVICES=0 \\\n  \/opt\/nccl-tests\/build\/all_reduce_perf \\\n  -b 8M -e 1G -f 2 -g 1\n<\/code><\/pre>\n\n\n\n
                              \u7ed3\u679c\u6838\u5fc3\u6307\u6807<\/h4>\n\n\n\n
                                \n
                              1. algbw<\/strong>\uff1a\u7b97\u6cd5\u5e26\u5bbd\uff0c\u53cd\u6620 AllReduce \u5b9e\u9645\u4e1a\u52a1\u6027\u80fd<\/li>\n\n\n\n
                              2. busbw<\/strong>\uff1a\u603b\u7ebf\u5e26\u5bbd\uff0c\u6298\u7b97\u540e\u7684\u7269\u7406\u901a\u4fe1\u5e26\u5bbd\uff0c\u7528\u4e8e\u6a2a\u5411\u5bf9\u6bd4<\/li>\n\n\n\n
                              3. #wrong<\/strong>\uff1a\u6570\u636e\u6821\u9a8c\u9519\u8bef\u6570\uff0c\u5fc5\u987b\u4e3a 0<\/li>\n<\/ol>\n\n\n\n
                                6. PCIe \u964d\u901f\u6392\u67e5<\/h3>\n\n\n\n
                                \u5e38\u89c1\u539f\u56e0<\/h4>\n\n\n\n
                                  \n
                                1. BIOS PCIe \u914d\u7f6e\u9519\u8bef\uff0c\u901f\u7387 \/ \u5bbd\u5ea6\u8bbe\u7f6e\u964d\u7ea7<\/li>\n\n\n\n
                                2. GPU \/ \u7f51\u5361\u63a5\u89e6\u4e0d\u826f\uff0c\u91d1\u624b\u6307\u6c27\u5316<\/li>\n\n\n\n
                                3. Riser \/ \u8f6c\u63a5\u677f\u6545\u969c\u6216\u8bbe\u8ba1\u964d\u901f<\/li>\n\n\n\n
                                4. \u4e3b\u677f PCIe Lane \u5206\u914d\u4e0d\u8db3\uff0cCPU \u63d2\u69fd\u901a\u9053\u95ee\u9898<\/li>\n\n\n\n
                                5. PCIe Switch \u6545\u969c<\/li>\n\n\n\n
                                6. \u8282\u80fd\u673a\u5236\u5bfc\u81f4\u7a7a\u95f2\u65f6\u964d\u901f\uff08\u8d1f\u8f7d\u540e\u6062\u590d\u5219\u6b63\u5e38\uff09<\/li>\n\n\n\n
                                7. GPU \u672c\u8eab PCIe \u63a5\u53e3\u786c\u4ef6\u6545\u969c<\/li>\n<\/ol>\n\n\n\n
                                  \u6392\u67e5\u547d\u4ee4<\/h4>\n\n\n\n
                                  lspci -s 04:00.0 -vv | grep -E 'LnkCap|LnkSta'\n<\/code><\/pre>\n\n\n\n
                                    \n
                                  • LnkCap<\/code>\uff1a\u7aef\u53e3\u652f\u6301\u7684\u6700\u5927\u901f\u7387 \/ \u5bbd\u5ea6<\/li>\n\n\n\n
                                  • LnkSta<\/code>\uff1a\u5f53\u524d\u5b9e\u9645\u8fd0\u884c\u7684\u901f\u7387 \/ \u5bbd\u5ea6<\/li>\n\n\n\n
                                  • \u82e5\u5f53\u524d\u5bbd\u5ea6 \/ \u901f\u7387\u4f4e\u4e8e\u652f\u6301\u7684\u6700\u5927\u503c\uff0c\u5373\u4e3a\u964d\u901f\uff1b\u518d\u901a\u8fc7\u4ea4\u53c9\u6362\u69fd\u3001\u6362\u5361\u5b9a\u4f4d\u6545\u969c\u70b9\u3002<\/li>\n<\/ul>\n\n\n\n
                                    \n\n\n\n
                                    \u516d\u3001\u6c47\u603b<\/h2>\n\n\n\n
                                      \n
                                    1. GPU \u670d\u52a1\u5668\u4e0a\u7ebf\u524d\u7684\u5b8c\u6574\u538b\u6d4b\u9a8c\u6536\u6d41\u7a0b\u662f\u4ec0\u4e48\uff1f<\/strong> \u7b54\uff1a\u5148\u505a\u57fa\u7840\u786c\u4ef6\u68c0\u67e5\uff08BMC\u3001nvidia-smi \u8bc6\u522b\u3001PCIe \u62d3\u6251\uff09\uff1b\u518d\u7528 CUDA Samples \u505a\u73af\u5883\u4e0e\u5e26\u5bbd\u9a8c\u8bc1\uff1b\u7136\u540e\u505a NCCL \u96c6\u5408\u901a\u4fe1\u6d4b\u8bd5\uff1b\u63a5\u7740\u7528 DCGM \u505a\u5168\u91cf\u5065\u5eb7\u8bca\u65ad\uff1b\u518d\u7528 gpu-burn \u505a 4~6 \u5c0f\u65f6\u957f\u65f6\u95f4\u6ee1\u8f7d\u538b\u529b\u6d4b\u8bd5\uff1b\u6700\u540e\u90e8\u7f72\u6a21\u578b\u505a\u4e1a\u52a1\u538b\u6d4b\uff0c\u9a8c\u8bc1\u5e76\u53d1\u4e0e\u65f6\u5ef6\u3002<\/li>\n\n\n\n
                                    2. gpu-burn \u7684\u5de5\u4f5c\u539f\u7406\u662f\u4ec0\u4e48\uff1f\u538b\u6d4b\u5173\u6ce8\u4ec0\u4e48\u6307\u6807\uff1f<\/strong> \u7b54\uff1agpu-burn \u57fa\u4e8e CUDA\uff0c\u901a\u8fc7\u5927\u91cf\u77e9\u9635\u4e58\u6cd5\u548c\u6d6e\u70b9\u8fd0\u7b97\u8ba9 GPU \u6ee1\u8d1f\u8f7d\uff0c\u540c\u65f6\u6821\u9a8c\u8ba1\u7b97\u7ed3\u679c\u3002\u538b\u6d4b\u91cd\u70b9\u5173\u6ce8\uff1a\u8ba1\u7b97\u9519\u8bef\u6570 errors\uff08\u5fc5\u987b\u4e3a 0\uff09\u3001Gflops \u7b97\u529b\u7a33\u5b9a\u6027\u3001GPU \u6e29\u5ea6\u4e0e\u529f\u8017\u3001\u663e\u5b58\u5360\u7528\u3002<\/li>\n\n\n\n
                                    3. AllReduce \u662f\u4ec0\u4e48\uff1f\u548c Reduce \u6709\u4ec0\u4e48\u533a\u522b\uff1f<\/strong> \u7b54\uff1aReduce \u662f\u5c06\u6240\u6709 GPU \u7684\u6570\u636e\u6c47\u603b\u8ba1\u7b97\uff0c\u7ed3\u679c\u53ea\u4fdd\u5b58\u5728\u4e00\u5f20 GPU \u4e0a\uff1bAllReduce \u662f\u5148\u505a Reduce \u6c47\u603b\uff0c\u518d\u628a\u7ed3\u679c\u5e7f\u64ad\u56de\u6240\u6709 GPU\uff0c\u6700\u7ec8\u6bcf\u5f20\u5361\u90fd\u6709\u5b8c\u6574\u7ed3\u679c\u3002AllReduce \u662f\u5206\u5e03\u5f0f\u8bad\u7ec3\u68af\u5ea6\u540c\u6b65\u6700\u6838\u5fc3\u7684\u64cd\u4f5c\u3002<\/li>\n\n\n\n
                                    4. RDMA \u76f8\u6bd4\u4f20\u7edf TCP \u6709\u4ec0\u4e48\u4f18\u52bf\uff1f<\/strong> \u7b54\uff1aRDMA \u7ed5\u8fc7\u64cd\u4f5c\u7cfb\u7edf\u5185\u6838\u534f\u8bae\u6808\uff0c\u5b9e\u73b0\u96f6\u62f7\u8d1d\u6570\u636e\u4f20\u8f93\uff0cCPU \u5360\u7528\u6781\u4f4e\u3001\u5ef6\u8fdf\u5fae\u79d2\u7ea7\u3001\u5e26\u5bbd\u66f4\u9ad8\u3002\u9002\u5408 AI \u5206\u5e03\u5f0f\u8bad\u7ec3\u3001\u9ad8\u6027\u80fd\u8ba1\u7b97\u7b49\u901a\u4fe1\u5bc6\u96c6\u578b\u573a\u666f\u3002<\/li>\n\n\n\n
                                    5. PCIe \u964d\u901f\u5e38\u89c1\u539f\u56e0\u6709\u54ea\u4e9b\uff0c\u600e\u4e48\u6392\u67e5\uff1f<\/strong> \u7b54\uff1a\u5e38\u89c1\u539f\u56e0\u5305\u62ec BIOS \u914d\u7f6e\u9519\u8bef\u3001\u786c\u4ef6\u63a5\u89e6\u4e0d\u826f\u3001\u8f6c\u63a5\u677f\u6545\u969c\u3001\u901a\u9053\u5206\u914d\u4e0d\u8db3\u3001\u4ea4\u6362\u673a\u95ee\u9898\u3001\u8282\u80fd\u673a\u5236\u7b49\u3002\u7528lspci -vv<\/code>\u5bf9\u6bd4 LnkCap\uff08\u80fd\u529b\uff09\u548c LnkSta\uff08\u5f53\u524d\uff09\u5224\u65ad\u662f\u5426\u964d\u901f\uff0c\u518d\u901a\u8fc7\u4ea4\u53c9\u6362\u69fd\u3001\u6362\u5361\u5b9a\u4f4d\u6545\u969c\u70b9\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"
                                      \u672c\u7bc7\u8986\u76d6\u786c\u4ef6\u5168\u94fe\u8def\u538b\u6d4b\u3001\u6a21\u578b\u670d\u52a1\u538b\u6d4b\u3001\u667a\u7b97\u7f51\u7edc\u4e0e\u591a\u673a\u901a\u4fe1\u6d4b\u8bd5\u4e09\u5927\u677f\u5757\uff0c\u68b3\u7406\u6838\u5fc3\u5de5\u5177\u3001\u5173\u952e\u6307\u6807\u4e0e\u6392\u969c\u601d\u8def \u4e00\u3001\u538b […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-174","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/174","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=174"}],"version-history":[{"count":2,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/174\/revisions"}],"predecessor-version":[{"id":182,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/174\/revisions\/182"}],"wp:attachment":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=174"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=174"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=174"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}