{"id":9377,"date":"2024-07-24T14:33:09","date_gmt":"2024-07-24T06:33:09","guid":{"rendered":"\/?p=9377"},"modified":"2024-07-24T14:33:09","modified_gmt":"2024-07-24T06:33:09","slug":"gpu%e5%88%a9%e7%94%a8%e7%8e%87%e9%97%ae%e9%a2%98mps","status":"publish","type":"post","link":"\/?p=9377","title":{"rendered":"GPU\u5229\u7528\u7387\u95ee\u9898MPS"},"content":{"rendered":"

\u53c2\u8003<\/h2>\n

https:\/\/docs.nvidia.com\/deploy\/mps\/#cuda-mps-enable-per-ctx-device-multiprocessor-partitioning<\/a><\/p>\n

https:\/\/zw0610.github.io\/notes-cn\/gpu-sharing-2.html<\/a><\/p>\n

GPU-Util\u548c Compute M.<\/h2>\n

GPU-Util<\/code> \u548c Compute M.<\/code> \u662f\u5728\u4f7f\u7528 nvidia-smi<\/code> \u547d\u4ee4\u67e5\u770b NVIDIA GPU \u72b6\u6001\u65f6\u5e38\u89c1\u7684\u4e24\u9879\u6307\u6807\u3002<\/p>\n

GPU-Util<\/h3>\n

GPU-Util<\/code> \u6307\u7684\u662f GPU \u7684\u5229\u7528\u7387\uff0c\u5b83\u662f\u8861\u91cf GPU \u5f53\u524d\u8d1f\u8f7d\u7684\u4e00\u4e2a\u767e\u5206\u6bd4\u3002\u5f53\u4e00\u4e2a GPU \u6b63\u5728\u5904\u7406\u8ba1\u7b97\u4efb\u52a1\u65f6\uff0c\u5b83\u7684\u5229\u7528\u7387\u4f1a\u589e\u52a0\u3002\u5982\u679c\u4f60\u770b\u5230 GPU-Util<\/code> \u7684\u503c\u63a5\u8fd1 100%\uff0c\u90a3\u4e48\u8fd9\u610f\u5473\u7740 GPU \u51e0\u4e4e\u6ee1\u8d1f\u8377\u5de5\u4f5c\uff0c\u53ef\u80fd\u6b63\u5728\u8fd0\u884c\u56fe\u5f62\u6e32\u67d3\u3001\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u3001\u79d1\u5b66\u8ba1\u7b97\u6216\u5176\u4ed6\u5bc6\u96c6\u578b\u8ba1\u7b97\u4efb\u52a1\u3002<\/p>\n

Compute M.<\/h3>\n

Compute M.<\/code> \u901a\u5e38\u6307\u7684\u662f Compute Mode\uff0c\u5373\u8ba1\u7b97\u6a21\u5f0f\u3002\u8fd9\u662f GPU \u7684\u4e00\u4e2a\u914d\u7f6e\u5c5e\u6027\uff0c\u51b3\u5b9a\u4e86 GPU \u5982\u4f55\u5904\u7406\u5e76\u53d1\u7684\u8ba1\u7b97\u4efb\u52a1\u3002\u4e0d\u540c\u7684\u8ba1\u7b97\u6a21\u5f0f\u5f71\u54cd\u7740 GPU \u4e0a\u53ef\u4ee5\u540c\u65f6\u8fd0\u884c\u7684\u8fdb\u7a0b\u6570\u91cf\u548c\u7c7b\u578b\u3002<\/p>\n

NVIDIA GPU \u7684\u8ba1\u7b97\u6a21\u5f0f\u6709\u4ee5\u4e0b\u51e0\u79cd\uff1a<\/p>\n

    \n
  1. Default<\/strong>\uff1a\u9ed8\u8ba4\u6a21\u5f0f\uff0c\u5141\u8bb8\u540c\u65f6\u8fd0\u884c\u56fe\u5f62\u548c\u8ba1\u7b97\u4efb\u52a1\uff0c\u4f46\u4e0d\u5141\u8bb8\u4e24\u4e2a\u8ba1\u7b97\u4efb\u52a1\u540c\u65f6\u4f7f\u7528\u540c\u4e00\u4e2a GPU\u3002<\/li>\n
  2. Exclusive Process<\/strong>\uff1a\u72ec\u5360\u6a21\u5f0f\uff0c\u53ea\u5141\u8bb8\u4e00\u4e2a\u8ba1\u7b97\u8fdb\u7a0b\u4f7f\u7528 GPU\uff0c\u76f4\u5230\u8be5\u8fdb\u7a0b\u7ed3\u675f\u6216\u653e\u5f03 GPU\u3002<\/li>\n
  3. Prohibited<\/strong>\uff1a\u7981\u6b62\u6a21\u5f0f\uff0c\u963b\u6b62\u4efb\u4f55\u8ba1\u7b97\u4efb\u52a1\u4f7f\u7528 GPU\uff0c\u4ec5\u5141\u8bb8\u56fe\u5f62\u4efb\u52a1\u3002<\/li>\n
  4. Exclusive Thread<\/strong>\uff1a\u7ebf\u7a0b\u72ec\u5360\u6a21\u5f0f\uff0c\u5141\u8bb8\u4e00\u4e2a\u8ba1\u7b97\u4efb\u52a1\u72ec\u5360 GPU\uff0c\u76f4\u5230\u5b83\u91ca\u653e GPU \u6216\u88ab\u5176\u4ed6\u7ebf\u7a0b\u62a2\u5360\u3002<\/li>\n<\/ol>\n

    \u901a\u8fc7 nvidia-smi<\/code>\uff0c\u4f60\u53ef\u4ee5\u67e5\u770b\u5f53\u524d GPU \u7684\u8ba1\u7b97\u6a21\u5f0f\u3002\u5982\u679c\u9700\u8981\u6539\u53d8\u8ba1\u7b97\u6a21\u5f0f\uff0c\u901a\u5e38\u9700\u8981\u4f7f\u7528 NVIDIA \u7684 nvidia-smi<\/code> \u547d\u4ee4\u884c\u5de5\u5177\uff0c\u5177\u4f53\u547d\u4ee4\u683c\u5f0f\u5982\u4e0b\uff1a<\/p>\n

    nvidia-smi -i <GPU-ID> -cm <mode><\/code><\/pre>\n
    \u5176\u4e2d <GPU-ID><\/code> \u662f\u4f60\u60f3\u8981\u914d\u7f6e\u7684 GPU \u7684 ID\uff0c\u800c <mode><\/code> \u662f\u4f60\u60f3\u8981\u8bbe\u7f6e\u7684\u8ba1\u7b97\u6a21\u5f0f\u3002\u4f8b\u5982\uff0c\u8981\u5c06 GPU 0 \u8bbe\u7f6e\u4e3a Exclusive Process \u6a21\u5f0f\uff0c\u53ef\u4ee5\u4f7f\u7528\uff1a<\/p>\n
    nvidia-smi -i 0 -cm 1<\/code><\/pre>\n
    \u4f46\u662f\uff0c\u8bf7\u6ce8\u610f\uff0c\u6539\u53d8\u8ba1\u7b97\u6a21\u5f0f\u53ef\u80fd\u9700\u8981\u7ba1\u7406\u5458\u6743\u9650\uff0c\u5e76\u4e14\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u53ef\u80fd\u9700\u8981\u91cd\u542f\u7cfb\u7edf\u624d\u80fd\u751f\u6548\u3002<\/p>\n
    MPS<\/h2>\n
    \u4e0e\u5355\u6838CPU\u7684\u8c03\u5ea6\u65b9\u5f0f\u7c7b\u4f3c\uff0c\u5728\u5355\u4e00\u65f6\u95f4\u7247\u5185\uff0cGPU\u4e2d\u53ea\u4f1a\u6709\u4e00\u4e2aGPU\u8fdb\u7a0b\u5728\u8fd0\u884c\uff0c\u5f53\u591a\u4e2a\u8fdb\u7a0b\u540c\u65f6\u628aCUDA\u4efb\u52a1\u53d1\u5c04\u5230GPU\u65f6\uff0cGPU\u4f7f\u7528\u65f6\u95f4\u7247\u8f6e\u8f6c\u8c03\u5ea6\u7684\u65b9\u5f0f\uff0c\u591a\u4e2aGPU\u8fdb\u7a0b\u4e4b\u95f4\u5728\u5fae\u89c2\u5c42\u9762\u4e0a\u662f\u4ea4\u66ff\u8fd0\u884c\u7684\u3002\u8fd9\u4e5f\u5bfc\u81f4\uff0c\u5728\u67d0\u4e00\u4e2a\u65f6\u95f4\u7247\u5185\uff0c\u5982\u679c\u6b63\u5728\u8fd0\u884c\u7684GPU\u8fdb\u7a0b\u6ca1\u6709\u5f88\u597d\u5730\u5229\u7528\u8ba1\u7b97\u8d44\u6e90\uff0c\u90a3\u4e48\u7a7a\u95f2\u7684\u8ba1\u7b97\u8d44\u6e90\u5c31\u662f\u6d6a\u8d39\u6389\u7684\u3002\u4e5f\u5c31\u662f\u8bf4\uff0cGPU\u5e76\u6ca1\u6709\u771f\u6b63\u5730\u8fdb\u884c\u5e76\u53d1\u8ba1\u7b97\u3002\u518d\u52a0\u4e0a\u4e0d\u540c\u8fdb\u7a0b\u7684\u4e0a\u4e0b\u6587\u5207\u6362\uff0c\u4e5f\u5e26\u6765\u4e86\u66f4\u591a\u7684\u65f6\u95f4\u5f00\u9500\u3002<\/p>\n
    Nvidia\u9488\u5bf9\u591a\u8fdb\u7a0b\u5e76\u53d1\u6267\u884c\u7684\u573a\u666f\u63a8\u51fa\u4e86\u591a\u8fdb\u7a0b\u670d\u52a1\u89e3\u51b3\u65b9\u6848-MPS\uff0c\u8be5\u65b9\u6848\u53ef\u4ee5\u505a\u5230\u7a7a\u5206\u590d\u7528\u3002<\/p>\n
    MPS\u7684\u8fd0\u884c\u6a21\u5f0f\u4e3a\u4e00\u4e2aMPS Server\u548c\u591a\u4e2aMPS Client\u3002MPS Server\u901a\u8fc7\u4e00\u4e2aCUDA Context\u7ba1\u7406GPU\u786c\u4ef6\u8d44\u6e90\uff0c\u6bcf\u4e2aMPS Client\u5bf9\u5e94\u4e00\u4e2aGPU\u8fdb\u7a0b\uff0c\u591a\u4e2aMPS Client\u4f1a\u5c06\u5b83\u4eec\u7684\u4efb\u52a1\u901a\u8fc7MPS Server\u4f20\u5165GPU\uff0cMPS Server\u53ef\u4ee5\u628a\u591a\u4e2a\u8fdb\u7a0b\u7684\u4e0a\u4e0b\u6587\u8fdb\u884c\u878d\u5408\uff0c\u5408\u5e76\u540e\u7684\u8fdb\u7a0b\u5c06\u591a\u4e2a\u8fdb\u7a0b\u7684Kernel\u4ea4\u7ec7\u5230\u4e00\u8d77\u8fdb\u884c\u53d1\u5c04\uff0c\u4ece\u800c\u8d8a\u8fc7\u4e86\u786c\u4ef6\u65f6\u95f4\u5206\u7247\u8c03\u5ea6\u7684\u9650\u5236\uff0c\u4f7f\u5f97\u5b83\u4eec\u7684CUDAkernels\u5b9e\u73b0\u771f\u6b63\u610f\u4e49\u4e0a\u7684\u5e76\u884c\uff0c\u8fd9\u53ef\u4ee5\u5e26\u6765\u4ee5\u4e0b\u597d\u5904\uff1a<\/p>\n
    > \u8fdb\u7a0b\u4e4b\u95f4\u65e0\u9700\u4e0a\u4e0b\u6587\u5207\u6362\uff0c\u51cf\u5c11\u4e86\u4e0a\u4e0b\u6587\u5207\u6362\u7684\u5f00\u9500\u3002<\/p>\n
    > \u540c\u4e00\u4e2a\u65f6\u95f4\u7247\u91cc\uff0c\u591a\u4e2a\u8fdb\u7a0b\u7684kernel\u4e00\u8d77\u6267\u884c\uff0c\u63d0\u5347\u4e86GPU\u8ba1\u7b97\u8d44\u6e90\u7684\u5229\u7528\u7387\u3002<\/p>\n
    MPS\u5728\u5355\u8fdb\u7a0b\u5bf9GPU\u5229\u7528\u7387\u4e0d\u9ad8\u7684\u60c5\u51b5\u4e0b\u662f\u975e\u5e38\u6709\u7528\u7684\uff0cMPS\u7684\u7f3a\u70b9\u5219\u5728\u4e8e\u6545\u969c\u9694\u79bb\u95ee\u9898\uff0c\u672c\u6587\u5ffd\u7565\u3002<\/p>\n
    \u7ec8\u6b62\u5ba2\u6237\u7aef\u5e94\u7528\u7a0b\u5e8f\u7684\u901a\u7528\u5de5\u4f5c\u6d41\uff1a<\/p>\n
    \u4f7f\u7528 control \u547d\u4ee4\u83b7\u53d6\u5f53\u524d\u6d3b\u52a8 MPS \u5ba2\u6237\u7aef\u7684\u72b6\u6001<\/p>\n
    $ echo "ps" | nvidia-cuda-mps-control\n\nPID ID SERVER DEVICE NAMESPACE COMMAND\n\n9741 0 6472 GPU-cb1213a3-d6a4-be7f 4026531836 .\/nbody\n\n9743 0 6472 GPU-cb1213a3-d6a4-be7f 4026531836 .\/matrixMul<\/code><\/pre>\n
    \u4f7f\u7528\u4e3b\u673a PID \u547d\u540d\u7a7a\u95f4\u4e2d\u7684 PID \u7ec8\u6b62\uff0c\u5982\u4e0b\u6240\u8ff0\uff1a<\/p>\n
    $ echo "terminate_client 6472 9741" | nvidia-cuda-mps-control\n\n#wait until terminate_client to return\n\n#upon successful termination 0 is returned\n\n0<\/code><\/pre>\n
    \u73b0\u5728kill\u662f\u5b89\u5168\u7684\uff1a<\/p>\n
    $ kill -9 9741<\/code><\/pre>\n
    Tegra \u5e73\u53f0\u4e0d\u652f\u6301 MPS \u5ba2\u6237\u7aef\u7ec8\u6b62\u3002<\/p>\n
    export CUDA_MPS_ENABLE_PER_CTX_DEVICE_MULTIPROCESSOR_PARTITIONING=1\nexport CUDA_MPS_ACTIVE_THREAD_PERCENTAGE=70<\/code><\/pre>\n
    nvidia-cuda-mps-\u63a7\u5236\uf0c1<\/a><\/h3>\n
    \u6b64\u63a7\u4ef6\u5b88\u62a4\u7a0b\u5e8f\u901a\u5e38\u5b58\u50a8\u5728 Linux \u7cfb\u7edf\u4e0b\uff0c\u901a\u5e38\u4ee5\u8d85\u7ea7\u7528\u6237\u6743\u9650\u8fd0\u884c\uff0c\u7528\u4e8e\u7ba1\u7406\u4e0b\u4e00\u8282\u4e2d\u63cf\u8ff0\u7684\u5185\u5bb9\u3002\u4ee5\u4e0b\u662f\u76f8\u5173\u7528\u4f8b\uff1a\/usr\/bin<\/code>nvidia-cuda-mps-server<\/code><\/p>\n
    man nvidia-cuda-mps-control          # Describes usage of this utility.\n\nnvidia-cuda-mps-control -d           # Start daemon in background process.\n\nps -ef | grep mps                    # Check if the MPS daemon is running.\n\necho quit | nvidia-cuda-mps-control  # Shut the daemon down.\n\nnvidia-cuda-mps-control -f           # Start daemon in foreground.\n\nnvidia-cuda-mps-control -v           # Print version of control daemon executable (applicable on Tegra platforms only).<\/code><\/pre>\n
    nvidia-smi<\/h3>\n
    man nvidia-smi                        # Describes usage of this utility.\nnvidia-smi -L                         # List the GPU's on node.\nnvidia-smi -q                         # List GPU state and configuration information.\nnvidia-smi -q -d compute              # Show the compute mode of each GPU.\nnvidia-smi -i 0 -c EXCLUSIVE_PROCESS  # Set GPU 0 to exclusive mode, run as root.\nnvidia-smi -i 0 -c DEFAULT            # Set GPU 0 to default mode, run as root. (SHARED_PROCESS)\nnvidia-smi -i 0 -r                    # Reboot GPU 0 with the new setting.<\/code><\/pre>\n
    \u63a7\u5236\u5b88\u62a4\u7a0b\u5e8f\u521b\u5efa\u4e00\u4e2a\u6587\u4ef6\uff0c\u5176\u4e2d\u5305\u542b\u63a7\u5236\u5b88\u62a4\u7a0b\u5e8f\u8fdb\u7a0b\u7684 PID \u8fd9\u3002\u5f53\u6709\u591a\u4e2a\u5e76\u884c\u8fd0\u884c\u7684\u63a7\u5236\u5b88\u62a4\u7a0b\u5e8f\u5b9e\u4f8b\u65f6\uff0c\u4e00<\/p>\n
    demo<\/h2>\n
    \u4f7f\u7528pytorch\u7f16\u5199\u4e00\u4e2a\u7b80\u5355\u7684\u8bad\u7ec3\u4efb\u52a1\uff0c\u4e14\u80fd\u8f83\u957f\u65f6\u95f4\u3001\u8f83\u5927\u5360\u7528\u5730\u5229\u7528GPU\uff1a<\/p>\n
    #!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\n\n# \u8bbe\u7f6e\u968f\u673a\u79cd\u5b50\u4ee5\u83b7\u5f97\u53ef\u91cd\u590d\u7684\u7ed3\u679c\ntorch.manual_seed(0)\n\n# \u521b\u5efa\u4e00\u4e2a\u7b80\u5355\u7684\u5168\u8fde\u63a5\u795e\u7ecf\u7f51\u7edc\nclass SimpleNet(nn.Module):\n    def __init__(self):\n        super(SimpleNet, self).__init__()\n        self.fc1 = nn.Linear(1000, 512)\n        self.fc2 = nn.Linear(512, 256)\n        self.fc3 = nn.Linear(256, 10)\n\n    def forward(self, x):\n        x = torch.relu(self.fc1(x))\n        x = torch.relu(self.fc2(x))\n        x = self.fc3(x)\n        return x\n\n# \u5b9e\u4f8b\u5316\u6a21\u578b\nmodel = SimpleNet().cuda()   # \u5c06\u6a21\u578b\u79fb\u52a8\u5230 GPU \u4e0a\n\n# \u8bbe\u7f6e\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)\n\n# \u8bbe\u7f6e\u8bad\u7ec3\u53c2\u6570\nbatch_size = 512*5\nnum_epochs = 1000*1000*1000\ndata_dim = 1000\nnum_classes = 10\n\n# \u521b\u5efa\u968f\u673a\u6570\u636e\ninputs = torch.randn(batch_size, data_dim).cuda()\nlabels = torch.randint(0, num_classes, (batch_size,)).cuda()\n\n# \u8bad\u7ec3\u5faa\u73af\nfor epoch in range(num_epochs):\n    optimizer.zero_grad()\n    outputs = model(inputs)\n    loss = criterion(outputs, labels)\n    loss.backward()\n    optimizer.step()\n\n    if (epoch + 1) % 100 == 0:\n        print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n\nprint('Finished Training')\n<\/code><\/pre>\n
    CUDA_VISIBLE_DEVICES=2 python app.py --avatar_id avator_1 --model musetalk --batch_size 8 --listenport 8056\nCUDA_VISIBLE_DEVICES=2 python app2.py --avatar_id avator_1 --model musetalk --batch_size 8 --listenport 8057<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
    \u53c2\u8003 https:\/\/docs.nvidia.com\/deploy\/mps\/#cuda-mps-enable-per-ctx-devic   \u2026 …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0},"categories":[224],"tags":[],"_links":{"self":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9377"}],"collection":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9377"}],"version-history":[{"count":1,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9377\/revisions"}],"predecessor-version":[{"id":9378,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9377\/revisions\/9378"}],"wp:attachment":[{"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9377"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9377"},{"taxonomy":"post_tag","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9377"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}