{"id":481,"date":"2026-07-08T02:59:01","date_gmt":"2026-07-07T18:59:01","guid":{"rendered":"https:\/\/tech-renaissance.cn\/blog\/?p=481"},"modified":"2026-07-08T16:51:30","modified_gmt":"2026-07-08T08:51:30","slug":"%e5%a4%9a%e7%ba%bf%e7%a8%8b%e9%a2%84%e5%a4%84%e7%90%86%e3%80%81numa-%e6%84%9f%e7%9f%a5%e4%b8%8e%e5%bc%82%e6%ad%a5%e5%8f%8c%e7%bc%93%e5%86%b2%ef%bc%9a%e8%ae%a9-gpu-%e4%b8%8d%e5%86%8d%e7%ad%89%e5%be%85","status":"publish","type":"post","link":"https:\/\/tech-renaissance.cn\/blog\/index.php\/2026\/07\/08\/481\/","title":{"rendered":"(10) \u591a\u7ebf\u7a0b\u9884\u5904\u7406\u3001NUMA \u611f\u77e5\u4e0e\u5f02\u6b65\u53cc\u7f13\u51b2\uff1a\u8ba9 GPU \u4e0d\u518d\u7b49\u5f85 CPU"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u5341<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u524d\u9762\u6211\u4eec\u8bb2\u4e86DataLoader\u62bd\u8c61\u3001RAW\/DTS \u53cc\u8def\u5f84\u3001FULLY\/PARTIAL \u52a0\u8f7d\u6a21\u5f0f\u548cDTS\u683c\u5f0f\u7b49\u3002\u4f46\u6570\u636e\u52a0\u8f7d\u7684\u6545\u4e8b\u8fd8\u6ca1\u8bb2\u5b8c\u2014\u2014\u683c\u5f0f\u548c Loader \u53ea\u89e3\u51b3\u4e86&#8221;\u600e\u4e48\u628a\u6570\u636e\u4ece\u78c1\u76d8\u8bfb\u51fa\u6765&#8221;\uff0c\u771f\u6b63\u51b3\u5b9a GPU \u662f\u5426&#8221;\u6328\u997f&#8221;\u7684\uff0c\u662f<strong>CPU \u9884\u5904\u7406\u80fd\u4e0d\u80fd\u8dd1\u5f97\u8db3\u591f\u5feb<\/strong>\uff0c\u4ee5\u53ca<strong>\u5904\u7406\u597d\u7684\u6570\u636e\u80fd\u4e0d\u80fd\u5728\u6b63\u786e\u7684\u65f6\u673a\u9001\u5230 GPU \u4e0a<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u7bc7\u6211\u4eec\u8bb2\u8fd9\u6761\u94fe\u8def\u7684\u6700\u540e\u4e24\u73af\uff1a<strong>\u591a\u7ebf\u7a0b\u9884\u5904\u7406<\/strong>\u4e0e<strong>\u5f02\u6b65\u53cc\u7f13\u51b2\u4f20\u8f93<\/strong>\u3002\u5b83\u4eec\u548c DTS \u683c\u5f0f\u4e00\u8d77\uff0c\u6784\u6210\u4e86 Tech-Renaissance \u6570\u636e\u5f15\u64ce\u7684\u6838\u5fc3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6570\u636e\u9965\u997f\uff1a\u8bad\u7ec3\u74f6\u9888\u5e76\u4e0d\u603b\u5728 GPU \u91cc<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u524d\u9762\u7684\u6587\u7ae0\u91cc\u6211\u591a\u6b21\u63d0\u5230\u4e00\u4e2a\u6982\u5ff5\uff1a<strong>\u6570\u636e\u9965\u997f\uff08data starvation\uff09<\/strong>\u3002\u4e00\u5757 A100 \u5b8c\u6210\u4e00\u6b21 256 \u6837\u672c\u7684 ResNet-50 \u524d\u5411+\u53cd\u5411\u53ef\u80fd\u53ea\u9700\u8981\u51e0\u5341\u6beb\u79d2\uff0c\u4f46 CPU \u4fa7\u8981\u89e3\u7801 256 \u5f20 JPEG\u3001\u505a RandomResizedCrop\u3001ColorJitter\u3001\u5f52\u4e00\u5316\u3001\u7ec4 batch\uff0c\u518d\u53d1\u8d77 H2D \u4f20\u8f93\u3002\u5982\u679c\u8fd9\u4e9b\u6570\u636e\u51c6\u5907\u5de5\u4f5c\u5355\u7ebf\u7a0b\u4e32\u884c\u6267\u884c\uff0c\u8017\u65f6\u5f88\u5bb9\u6613\u8d85\u8fc7 GPU \u8ba1\u7b97\u65f6\u95f4\uff0c\u5bfc\u81f4\u6602\u8d35\u7684\u52a0\u901f\u5668\u7a7a\u95f2\u7b49\u5f85\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u4ee3\u6846\u67b6\u90fd\u610f\u8bc6\u5230\u4e86\u8fd9\u4e00\u70b9\u3002PyTorch \u7684 <code>DataLoader<\/code> \u63d0\u4f9b\u4e86 <code>num_workers<\/code> \u53c2\u6570\uff0c\u7528\u591a\u4e2a\u5b50\u8fdb\u7a0b\u5e76\u884c\u8bfb\u53d6\u548c\u9884\u5904\u7406\uff1b<code>pin_memory=True<\/code> \u4f1a\u628a CPU \u5185\u5b58\u9501\u9875\uff0c\u65b9\u4fbf\u5f02\u6b65 H2D \u4f20\u8f93\uff1b<code>persistent_workers=True<\/code> \u5219\u8ba9\u5b50\u8fdb\u7a0b\u5728\u591a\u4e2a epoch \u4e4b\u95f4\u590d\u7528\uff0c\u907f\u514d\u53cd\u590d\u521b\u5efa\u8fdb\u7a0b\u7684\u5f00\u9500\u3002TensorFlow \u7684 <code>tf.data<\/code> \u901a\u8fc7 <code>map(..., num_parallel_calls)<\/code> \u548c <code>prefetch()<\/code> \u628a\u9884\u5904\u7406\u4e0e GPU \u6d88\u8d39\u6d41\u6c34\u7ebf\u5316\u3002NVIDIA \u7684 DALI \u8d70\u5f97\u66f4\u8fdc\uff0c\u5b83\u628a\u4e00\u90e8\u5206\u9884\u5904\u7406\u76f4\u63a5\u642c\u5230 GPU \u4e0a\u505a\uff0c\u8fdb\u4e00\u6b65\u51cf\u8f7b CPU \u538b\u529b\u3002\u8fd9\u4e9b\u65b9\u6848\u7684\u672c\u8d28\u90fd\u662f\u540c\u4e00\u4e2a\u601d\u60f3\uff1a<strong>\u628a\u6570\u636e\u7ba1\u7ebf\u4e5f\u53d8\u6210\u4e00\u6761\u6d41\u6c34\u7ebf\uff0c\u8ba9 CPU \u9884\u5904\u7406\u3001CPU\u2192GPU \u4f20\u8f93\u3001GPU \u8ba1\u7b97\u4e09\u4e2a\u9636\u6bb5\u5c3d\u53ef\u80fd\u91cd\u53e0\u6267\u884c<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u601d\u60f3\u76f8\u540c\uff0c\u5b9e\u73b0\u7ec6\u8282\u4f1a\u51b3\u5b9a\u5929\u82b1\u677f\u3002PyTorch \u7684 <code>num_workers<\/code> \u662f<strong>\u6bcf\u4e2a GPU \u8fdb\u7a0b<\/strong>\u72ec\u7acb\u914d\u7f6e\u7684\uff0c8 \u5361\u8bad\u7ec3\u65f6\u5982\u679c\u6bcf\u5f20\u5361\u914d 16 \u4e2a worker\uff0c\u5168\u5c40\u5c31\u6709 128 \u4e2a\u8fdb\u7a0b\uff1b\u8fd9\u4e9b\u8fdb\u7a0b\u662f\u5426\u843d\u5728\u5408\u9002\u7684 CPU \u6838\u5fc3\u4e0a\u3001\u662f\u5426\u8bbf\u95ee\u4e86\u672c\u5730 NUMA \u5185\u5b58\u3001worker \u4e4b\u95f4\u662f\u5426\u5b58\u5728\u7ade\u4e89\uff0c\u90fd\u9700\u8981\u7528\u6237\u989d\u5916\u64cd\u5fc3\u3002Tech-Renaissance \u5219\u628a\u8fd9\u4e9b\u95ee\u9898\u6536\u8fdb\u4e86\u6846\u67b6\u5185\u90e8\uff1a\u7528\u4e00\u4e2a\u5168\u5c40\u7edf\u4e00\u7684\u6301\u4e45 worker \u6c60\u3001\u9759\u6001\u6837\u672c\u9886\u53d6\u3001NUMA \u611f\u77e5\u7684\u9501\u9875\u5185\u5b58\u5206\u914d\u3001\u4ee5\u53ca per-GPU \u7684 TransferStation \u53cc\u7f13\u51b2\uff0c\u628a&#8221;\u6570\u636e\u4e0d\u7b49\u5f85&#8221;\u8fd9\u4ef6\u4e8b\u505a\u5230\u7cfb\u7edf\u5316\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u6301\u4e45 worker \u6c60\u4e0e\u9759\u6001\u9886\u53d6\uff1a\u96f6\u7ade\u4e89\u7684\u5e76\u53d1<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>src\/data\/preprocessor.cpp<\/code> \u4e2d\uff0c<code>Preprocessor::start_worker_pool()<\/code> \u542f\u52a8\u4e86\u4e00\u4e2a<strong>\u6301\u4e45\u7ebf\u7a0b\u6c60<\/strong>\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">void Preprocessor::start_worker_pool(DataLoader&amp; loader) {\n    worker_pool_.clear();\n    worker_pool_.reserve(config_.num_workers);\n\n    for (int i = 0; i &lt; config_.num_workers; ++i) {\n        worker_pool_.emplace_back([this, i, &amp;loader]() {\n            worker_func_persistent(i, loader);\n        });\n    }\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e9b worker \u4e00\u65e6\u542f\u52a8\uff0c\u5c31\u4f1a\u4e00\u76f4\u5904\u7406\u591a\u4e2a buffer\uff0c\u76f4\u5230\u6574\u4e2a epoch \u7ed3\u675f\u3002\u65e9\u671f\u5b9e\u73b0\u662f&#8221;\u6bcf\u4e2a buffer \u521b\u5efa\u4e00\u6279\u7ebf\u7a0b\u3001\u5904\u7406\u5b8c\u5c31\u9500\u6bc1&#8221;\uff0c\u5f88\u5feb\u53d1\u73b0\u7ebf\u7a0b\u521b\u5efa\u548c join \u7684\u5f00\u9500\u5728\u9ad8\u5e76\u53d1\u4e0b\u4e0d\u53ef\u5ffd\u7565\u3002\u6539\u6210\u6301\u4e45\u7ebf\u7a0b\u6c60\u540e\uff0c\u7ebf\u7a0b\u53ea\u521b\u5efa\u4e00\u6b21\uff0c\u4e3b\u5faa\u73af\u901a\u8fc7 <code>current_buffer_seq_<\/code> \u539f\u5b50\u53d8\u91cf\u901a\u77e5 worker \u8fdb\u5165\u4e0b\u4e00\u8f6e\uff0c\u5f00\u9500\u88ab\u538b\u5230\u6700\u4f4e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5173\u952e\u7684\u8bbe\u8ba1\u5728\u4e8e<strong>\u9759\u6001\u9886\u53d6<\/strong>\u3002PyTorch DataLoader \u901a\u5e38\u901a\u8fc7\u4e3b\u8fdb\u7a0b sampler \u751f\u6210 index\uff0c\u518d\u7ecf\u7531\u591a\u8fdb\u7a0b\u961f\u5217\u5206\u53d1\u7ed9 worker\uff1b\u5728\u591a\u8fdb\u7a0b\u3001\u961f\u5217\u901a\u4fe1\u548c worker \u8c03\u5ea6\u8fc7\u7a0b\u4e2d\u4ecd\u5b58\u5728\u8c03\u5ea6\u4e0e\u540c\u6b65\u5f00\u9500\u3002Tech-Renaissance \u7684\u505a\u6cd5\u662f\uff1a<strong>\u6bcf\u4e2a worker \u7684\u6837\u672c\u4f4d\u7f6e\u5728\u6570\u5b66\u4e0a\u9884\u5148\u786e\u5b9a<\/strong>\uff0c\u4e0d\u9700\u8981\u62a2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>src\/data\/preprocess_worker.cpp<\/code> \u4e2d\uff0c<code>calculate_write_position()<\/code> \u5b9e\u73b0\u4e86\u8fd9\u4e2a\u903b\u8f91\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">std::pair&lt;int, int> PreprocessWorker::calculate_write_position() const {\n    \/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n    \/\/   M = config_.num_workers_per_engine   \/\/ \u6bcf\u4e2a GPU \u5bf9\u5e94\u7684 PW \u6570\u91cf\n    \/\/   j = config_.pid_in_engine             \/\/ \u8be5 PW \u5728 Engine \u5185\u7684\u7f16\u53f7\n    \/\/   B = config_.local_batch_size          \/\/ \u5355\u5361 batch size\n    \/\/   n = local_sample_id_                  \/\/ \u5f53\u524d PW \u5728\u672c phase \u5df2\u5904\u7406\u6837\u672c\u6570\n\n    int global_seq = n * M + j;\n    int batch_id   = global_seq \/ B;\n    int position   = global_seq % B;\n    return {batch_id, position};\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5047\u8bbe\u6bcf\u4e2a GPU \u5bf9\u5e94 16 \u4e2a worker\uff0c\u90a3\u4e48 worker 0 \u5904\u7406\u7b2c 0\u300116\u300132\u2026\u53f7\u6837\u672c\uff0cworker 1 \u5904\u7406\u7b2c 1\u300117\u300133\u2026\u53f7\u6837\u672c\u3002\u6bcf\u4e2a worker \u53ea\u9700\u8981\u7ef4\u62a4\u81ea\u5df1\u7684 <code>local_sample_id_<\/code>\uff0c\u4e0d\u9700\u8981\u4efb\u4f55\u539f\u5b50\u64cd\u4f5c\u6765\u51b3\u5b9a&#8221;\u4e0b\u4e00\u4e2a\u8be5\u53d6\u8c01&#8221;\u3002\u6e90\u4ee3\u7801\u91cc\u7684\u6ce8\u91ca\u751a\u81f3\u7ed9\u51fa\u4e86\u4e00\u4e2a\u5f62\u5f0f\u5316\u8bc1\u660e\uff1a\u56e0\u4e3a\u4e0d\u540c worker \u7684 <code>j<\/code> \u4e0d\u540c\uff0c\u4efb\u4f55\u4e24\u4e2a worker \u6c38\u8fdc\u4e0d\u4f1a\u843d\u5230\u540c\u4e00\u4e2a <code>(batch_id, position)<\/code> \u69fd\u4f4d\u3002\u8fd9\u5c31\u662f<strong>\u96f6\u7ade\u4e89\u5e76\u53d1<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u96f6\u7ade\u4e89\u5e26\u6765\u7684\u597d\u5904\u662f\u7ebf\u6027\u7684\u6269\u5c55\u6027\u3002\u5728\u5b9e\u6d4b\u4e2d\uff0cTech-Renaissance \u53ef\u4ee5\u628a\u9884\u5904\u7406\u7ebf\u7a0b\u5f00\u5230 128\u3001200 \u751a\u81f3\u66f4\u591a\uff0c\u800c\u4e0d\u4f1a\u56e0\u4e3a\u9501\u7ade\u4e89\u5bfc\u81f4\u6536\u76ca\u9aa4\u964d\u3002\u6bcf\u4e2a worker \u8fd8\u62e5\u6709\u81ea\u5df1\u72ec\u7acb\u7684 Workshop \u5185\u5b58\u3001TurboJPEG 3.x \u53e5\u67c4\u3001PO \u94fe\u514b\u9686\uff0c\u4ee5\u53ca\u72ec\u7acb\u7684\u968f\u673a\u6570\u751f\u6210\u5668\uff0cworker \u4e4b\u95f4\u6ca1\u6709\u5171\u4eab\u72b6\u6001\uff0c\u4e5f\u6ca1\u6709 false sharing \u70ed\u70b9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">worker \u4e0e GPU \u7684\u5bf9\u5e94\u5173\u7cfb\u4e5f\u975e\u5e38\u76f4\u63a5\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   worker_id  = \u5168\u5c40 worker \u7f16\u53f7\n\/\/   world_size = GPU \u5361\u6570\n\nint engine_id    = worker_id % world_size;   \/\/ \u8be5 PW \u670d\u52a1\u54ea\u5f20 GPU\nint pid_in_engine = worker_id \/ world_size;  \/\/ \u5728\u8be5 GPU \u7684 worker \u4e2d\u7684\u7f16\u53f7<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u6709\u670d\u52a1\u4e8e\u540c\u4e00\u5f20 GPU \u7684 worker \u53ea\u5411\u540c\u4e00\u4e2a <code>TransferStation<\/code> \u5199\u5165\uff0c\u4e0d\u540c GPU \u7684\u6570\u636e\u6d41\u5929\u7136\u9694\u79bb\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001NUMA \u611f\u77e5\u4e0e CPU \u7ed1\u6838\uff1a\u8ba9\u6570\u636e\u9760\u8fd1\u8ba1\u7b97<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u591a\u7ebf\u7a0b\u53ea\u662f\u7b2c\u4e00\u6b65\u3002\u5982\u679c 128 \u4e2a worker \u968f\u673a\u6563\u843d\u5728 CPU \u4e0a\uff0c\u800c\u5b83\u4eec\u9891\u7e41\u8bbf\u95ee\u7684\u5185\u5b58\u53c8\u6302\u5728\u53e6\u4e00\u4e2a NUMA \u8282\u70b9\uff0c\u90a3\u4e48\u8de8\u8282\u70b9\u8bbf\u5b58\u5e26\u5bbd\u5f88\u5feb\u5c31\u4f1a\u6210\u4e3a\u65b0\u7684\u74f6\u9888\u3002\u9ad8\u7aef\u8bad\u7ec3\u670d\u52a1\u5668\u901a\u5e38\u662f\u591a\u8def CPU\u3001\u591a\u4e2a NUMA \u8282\u70b9\uff0c\u6bcf\u5f20 GPU \u901a\u8fc7 PCIe \u6302\u5728\u67d0\u4e2a NUMA \u8282\u70b9\u4e0b\u3002\u7406\u60f3\u7684\u72b6\u51b5\u662f\uff1a<strong>\u5904\u7406\u67d0\u5f20 GPU \u6570\u636e\u7684 worker \u7ed1\u5b9a\u5728\u8be5 GPU \u6240\u5728 NUMA \u8282\u70b9\u7684 CPU \u4e0a\uff0c\u5e76\u4e14\u5728\u8be5 NUMA \u8282\u70b9\u672c\u5730\u5206\u914d\u5185\u5b58<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Tech-Renaissance \u5728 <code>TR_SCENE_GPU_CLOUD<\/code> \u573a\u666f\u4e0b\u505a\u4e86\u4e24\u90e8\u5206\u8bbe\u8ba1\uff1a<strong>Staging Buffer \u7684 NUMA \u611f\u77e5\u5206\u914d<\/strong> \u4e0e <strong>worker \u7ebf\u7a0b\u7684\u786e\u5b9a\u6027 CPU \u7ed1\u6838<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 Staging Buffer \u7684 NUMA \u611f\u77e5\u5206\u914d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><code>src\/core\/staging_buffer_pool.cpp<\/code> \u4e3a\u6bcf\u4e2a GPU \u5355\u72ec\u5206\u914d\u4e00\u5757\u9501\u9875\u5185\u5b58\u3002\u5206\u914d\u524d\uff0c\u5b83\u4f1a\u5148\u901a\u8fc7 CUDA API \u67e5\u8be2 GPU \u7684 PCIe DBDF\uff0c\u518d\u8bfb\u53d6 <code>\/sys\/bus\/pci\/devices\/...\/numa_node<\/code> \u786e\u5b9a GPU \u6240\u5c5e\u7684 NUMA \u8282\u70b9\uff1b\u7136\u540e\u4e3a\u6bcf\u4e2a GPU \u542f\u52a8\u4e00\u4e2a\u72ec\u7acb\u5206\u914d\u7ebf\u7a0b\uff0c\u5728\u7ebf\u7a0b\u5185\u901a\u8fc7 <code>numa_run_on_node<\/code> \u548c <code>numa_set_preferred<\/code> \u628a\u7ebf\u7a0b\u8fc1\u79fb\u5230\u8be5 GPU \u5bf9\u5e94\u7684 NUMA \u8282\u70b9\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   gpu_id   = \u76ee\u6807 GPU \u7f16\u53f7\n\/\/   numa_node = GPU \u6240\u5c5e NUMA \u8282\u70b9\uff08\u901a\u8fc7 \/sys\/bus\/pci\/devices\/&lt;DBDF>\/numa_node \u8bfb\u53d6\uff09\n\/\/   bytes    = \u8be5 GPU \u9700\u8981\u7684 Staging Buffer \u5b57\u8282\u6570\n\nvoid StagingBufferPool::allocate_worker(int gpu_id, int numa_node,\n                                        size_t bytes, void** out_ptr) {\n    if (numa_node >= 0 &amp;&amp; numa_available() >= 0) {\n        numa_run_on_node(numa_node);    \/\/ \u7ebf\u7a0b\u7ed1\u5b9a\u5230\u76ee\u6807 NUMA \u8282\u70b9\n        numa_set_preferred(numa_node);  \/\/ \u4f18\u5148\u4ece\u8be5\u8282\u70b9\u5206\u914d\u5185\u5b58\n    }\n    cudaSetDevice(gpu_id);\n    cudaMallocHost(out_ptr, bytes);     \/\/ \u5206\u914d\u9501\u9875\u5185\u5b58\n    std::memset(*out_ptr, 0, bytes);    \/\/ \u9632\u5fa1\u6027\u6e05\u96f6\u4e0e\u9875\u8868\u9884\u70ed\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u91cc\u7684\u5173\u952e\u662f\uff1a<strong>\u9501\u9875\u5185\u5b58\u5728 <code>cudaMallocHost<\/code> \u5206\u914d\uff08\u9501\u9875\uff09\u65f6\u5373\u88ab\u63d0\u4ea4\u5e76\u9501\u5b9a\uff0c\u5176 NUMA \u5f52\u5c5e\u7531\u5206\u914d\u7ebf\u7a0b\u5f53\u65f6\u7684 NUMA \u7b56\u7565\u51b3\u5b9a<\/strong>\u3002\u56e0\u6b64\u5fc5\u987b\u5728\u7ed1\u5b9a\u5230\u76ee\u6807 NUMA \u8282\u70b9\u540e\u518d\u8c03\u7528 <code>cudaMallocHost<\/code>\uff1b\u968f\u540e\u7684 <code>memset<\/code> \u4e3b\u8981\u8d77\u6e05\u96f6\u4e0e\u9875\u8868\u9884\u70ed\u4f5c\u7528\uff0c\u8ba9 worker \u540e\u7eed\u8bbf\u95ee\u65f6\u4e0d\u4f1a\u89e6\u53d1\u989d\u5916\u7684\u7f3a\u9875\u4e2d\u65ad\u3002\u5982\u679c NUMA \u7ed1\u5b9a\u5728\u5206\u914d\u4e4b\u540e\u624d\u505a\uff0c\u5df2\u7ecf\u65e0\u6cd5\u6539\u53d8\u5df2\u9501\u5b9a\u7684\u7269\u7406\u9875\u4f4d\u7f6e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\uff0c\u4ece CPU \u5230 GPU \u7684 H2D \u4f20\u8f93\u8def\u5f84\u6700\u77ed\uff1aworker \u5728\u672c\u5730 NUMA \u6838\u5fc3\u4e0a\u9884\u5904\u7406\uff0c\u628a\u7ed3\u679c\u5199\u5165\u672c\u5730 NUMA \u7684\u9501\u9875\u5185\u5b58\uff0c\u518d\u901a\u8fc7\u540c NUMA \u8282\u70b9\u4e0b\u7684 PCIe Root Complex \u5f02\u6b65\u62f7\u8d1d\u5230 GPU\uff0c\u5c3d\u91cf\u51cf\u5c11\u8de8\u8282\u70b9\u8bbf\u95ee\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 Worker \u7ebf\u7a0b\u7684\u786e\u5b9a\u6027 CPU \u7ed1\u6838<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 <code>TR_SCENE_GPU_CLOUD<\/code> \u573a\u666f\u4e0b\uff0c\u6bcf\u4e2a worker \u7ebf\u7a0b\u542f\u52a8\u65f6\u4f1a\u8c03\u7528 <code>Preprocessor::bind_worker_to_cpu()<\/code>\uff0c\u901a\u8fc7 <code>pthread_setaffinity_np<\/code> \u7ed1\u5b9a\u5230\u56fa\u5b9a\u903b\u8f91 CPU\u3002\u5f53\u524d\u5b9e\u73b0\u91c7\u7528<strong>\u786e\u5b9a\u6027 Round-Robin<\/strong>\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   total_workers = \u9884\u5904\u7406 worker \u603b\u6570\n\/\/   ncpus         = \u7cfb\u7edf\u5728\u7ebf CPU \u6838\u5fc3\u6570\uff08sysconf(_SC_NPROCESSORS_ONLN)\uff09\n\nfor (int w = 0; w &lt; total_workers; ++w) {\n    int target_cpu = w % ncpus;\n    binding_map[w] = target_cpu;\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u79cd\u7ed1\u6cd5\u7684\u4f18\u70b9\u662f\u6bcf\u4e2a worker \u90fd\u6709\u56fa\u5b9a\u7684&#8221;\u5ea7\u4f4d&#8221;\uff0c\u4e0d\u4f1a\u51fa\u73b0\u591a\u4e2a worker \u540c\u65f6\u6324\u5728\u540c\u4e00\u4e2a\u6838\u5fc3\u4e0a\u4e89\u593a\u6267\u884c\u5355\u5143\u7684\u60c5\u51b5\uff1b\u540c\u65f6\uff0c\u7531\u4e8e Workshop \u5185\u5b58\u3001TurboJPEG \u53e5\u67c4\u7b49\u5c40\u90e8\u72b6\u6001\u90fd\u5728\u5df2\u7ecf\u7ed1\u5b9a\u7684\u7ebf\u7a0b\u4e2d\u521b\u5efa\uff0c\u64cd\u4f5c\u7cfb\u7edf\u9ed8\u8ba4\u7684 first-touch \u7b56\u7565\u4f1a\u8ba9\u8fd9\u4e9b\u70ed\u6570\u636e\u4f18\u5148\u843d\u5728\u8be5 CPU \u6240\u5728\u7684 NUMA \u8282\u70b9\u4e0a\uff0c\u63d0\u5347\u7f13\u5b58\u4e0e\u5185\u5b58\u5c40\u90e8\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u8bf4\u660e\u7684\u662f\uff0c\u8fd9\u4e00\u5c42\u7ed1\u6838\u76ee\u524d\u91c7\u7528\u7684\u662f\u5168\u5c40 Round-Robin\uff0c\u800c\u4e0d\u662f\u6309 GPU-NUMA \u8282\u70b9\u505a\u4e25\u683c\u6620\u5c04\u3002\u5b83\u5df2\u7ecf\u53ef\u4ee5\u6d88\u9664\u540c\u6838\u7ade\u4e89\uff0c\u5e76\u5728\u591a\u6570\u60c5\u51b5\u4e0b\u83b7\u5f97\u4e0d\u9519\u7684\u5c40\u90e8\u6027\uff1b\u5982\u679c\u672a\u6765\u9700\u8981\u66f4\u6781\u81f4\u7684 NUMA \u4eb2\u548c\u6027\uff0c\u6846\u67b6\u4e2d\u5df2\u7ecf\u4fdd\u7559\u4e86 <code>HardwareTopology<\/code> \/ <code>CPUAdvisor<\/code> \u7b49\u57fa\u7840\u8bbe\u65bd\uff0c\u53ef\u4ee5\u5728\u6b64\u57fa\u7840\u4e0a\u8fdb\u4e00\u6b65\u7ec6\u5316\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001TransferStation\uff1aCPU \u4e0e GPU \u4e4b\u95f4\u7684\u53cc\u7f13\u51b2\u9a7f\u7ad9<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u9884\u5904\u7406\u5b8c\u6210\u4e4b\u540e\uff0c\u6570\u636e\u8fd8\u8981\u5230\u8fbe GPU\u3002Tech-Renaissance \u6ca1\u6709\u8ba9\u7528\u6237\u624b\u52a8 <code>tensor.to(device)<\/code>\uff0c\u800c\u662f\u628a H2D \u4f20\u8f93\u5c01\u88c5\u8fdb\u4e86\u6846\u67b6\u5185\u90e8\uff0c\u6838\u5fc3\u662f <code>TransferStation<\/code>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a GPU \u5bf9\u5e94\u4e00\u4e2a <code>TransferStation<\/code>\uff0c\u5b83\u4ece <code>StagingBufferPool<\/code> \u62ff\u5230\u4e24\u5757\u5df2\u7ecf\u5206\u914d\u597d\u7684\u9501\u9875\u5185\u5b58\u3002\u6bcf\u5757\u5185\u5b58\u6309 <code>[labels][padding][image_data]<\/code> \u5e03\u5c40\uff0c\u6807\u7b7e\u533a\u4e0e\u56fe\u50cf\u6570\u636e\u533a\u90fd\u6309 256 \u5b57\u8282\u5bf9\u9f50\uff0c\u5e76\u989d\u5916\u9884\u7559 16 \u5b57\u8282\u4f5c\u4e3a SIMD \u8d8a\u754c\u4fdd\u62a4\u3002\u8fd9\u4e2a\u5e03\u5c40\u4e0e GPU \u7aef <code>DTensor<\/code> \u7684\u5185\u5b58\u5e03\u5c40\u5b8c\u5168\u517c\u5bb9\uff0cH2D \u5f02\u6b65\u4f20\u8f93\u540e\u4e0d\u9700\u8981\u518d\u505a\u6570\u636e\u91cd\u6392\u3002\u4ee5 AMP\uff08FP16\uff09\u8def\u5f84\u4e3a\u4f8b\uff0c\u5355\u533a\u5927\u5c0f\u8ba1\u7b97\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   n            = local_batch_size\n\/\/   max_sample_bytes = \u5355\u4e2a\u6837\u672c\u6700\u5927\u5b57\u8282\u6570\n\/\/   label_raw    = n * sizeof(int32_t)\n\/\/   data_raw     = n * max_sample_bytes\n\n\/\/ labels \u533a\u5bf9\u9f50\u540e\u5927\u5c0f\uff08FP32\/INT32 \u901a\u7528\u516c\u5f0f\uff09\nsize_t label_aligned = 2 * align_up_256(label_raw \/ 2 + 16);\n\n\/\/ image \u6570\u636e\u533a\u5bf9\u9f50\u540e\u5927\u5c0f\uff08AMP\/FP16 \u8def\u5f84\uff09\nsize_t data_aligned  = align_up_256(data_raw + 16);\n\n\/\/ \u5355\u533a\u603b\u5927\u5c0f = label_aligned + data_aligned\n\/\/ A \u533a\u8d77\u59cb\uff1abase\n\/\/ B \u533a\u8d77\u59cb\uff1abase + per_zone\uff0c\u5176\u4e2d per_zone = label_aligned + data_aligned<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\"><code>TransferStation<\/code> \u7ef4\u62a4\u4e24\u4e2a buffer\uff1aCPU \u7684 PreprocessWorker \u5411\u5f53\u524d buffer \u5199\u5165\u6837\u672c\uff0c\u5f53\u5f53\u524d buffer \u88ab\u4e00\u4e2a batch \u586b\u6ee1\u6216\u6240\u6709 worker \u62a5\u544a\u6837\u672c\u8017\u5c3d\u65f6\uff0c\u8c03\u7528 <code>execute_transfer_locked()<\/code> \u628a\u5b83\u6807\u8bb0\u4e3a readable\uff0c\u5e76\u5207\u6362\u5230\u53e6\u4e00\u5757 buffer \u7ee7\u7eed\u5199\u3002GPU \u4fa7\u7684\u6df1\u5ea6\u5b66\u4e60\u5f15\u64ce\u5219\u5728 CUDA Graph \u4e2d\u8c03\u7528 <code>wait_buffer_readable()<\/code> \u7b49\u5f85\u53ef\u8bfb\uff0c\u8bfb\u53d6\u5b8c\u6bd5\u540e\u8c03\u7528 <code>set_buffer_writeable()<\/code> \u628a buffer \u4ea4\u8fd8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5173\u952e\u4ee3\u7801\u5728 <code>src\/data\/transfer_station.cpp<\/code> \u4e2d\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">void TransferStation::execute_transfer_locked(int samples_count,\n                                              bool fill_before_transfer) {\n    int buf_id = current_buffer_.load(std::memory_order_acquire);\n\n    \/\/ \u82e5\u9700\u8981 filling\uff0c\u590d\u5236\u9996\u6837\u672c\u586b\u6ee1\u6700\u540e\u4e00\u4e2a\u4e0d\u5b8c\u6574 batch\n    int samples_count_after_filling = samples_count;\n    if (need_filling_ &amp;&amp; fill_before_transfer) {\n        samples_count_after_filling++;\n        std::memcpy(buffer_data_[buf_id] + snapshot_sample_bytes * samples_count,\n                    filling_sample_data_, snapshot_sample_bytes);\n        buffer_labels_[buf_id][samples_count] = filling_label_;\n    }\n\n    set_buffer_writeable(buf_id, false);\n    buffer_actual_transfer_samples_[buf_id] = samples_count_after_filling;\n    set_buffer_readable(buf_id, true);\n\n    \/\/ \u7b49\u5f85\u53e6\u4e00\u5757 buffer \u53ef\u5199\uff0c\u7136\u540e\u5207\u6362\n    int next_buf = 1 - buf_id;\n    while (!buffer_is_writeable(next_buf)) {\n        wait_buffer_writeable(next_buf);\n    }\n    set_buffer_readable(next_buf, false);\n    current_buffer_.store(next_buf, std::memory_order_release);\n    current_batch_id_.fetch_add(1, std::memory_order_release);\n    cv_batch_ready_.notify_all();\n}<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6bb5\u4ee3\u7801\u5b8c\u6210\u4e86\u4e09\u4ef6\u4e8b\uff1a\u628a\u521a\u5199\u6ee1\u7684 buffer \u4ea4\u7ed9 GPU\uff1b\u7b49\u5f85\u53e6\u4e00\u5757 buffer \u88ab GPU \u6d88\u8d39\u5b8c\uff1b\u5207\u6362\u5f53\u524d buffer \u5e76\u5524\u9192\u6240\u6709\u7b49\u5f85\u7684 worker\u3002\u7531\u4e8e\u4e24\u5757 buffer \u4ea4\u66ff\u5de5\u4f5c\uff0cCPU \u9884\u5904\u7406\u548c GPU \u8ba1\u7b97\u53ef\u4ee5\u6d41\u6c34\u7ebf\u91cd\u53e0\uff1aGPU \u5728\u8ba1\u7b97 buffer A \u65f6\uff0cCPU \u6b63\u5728\u5f80 buffer B \u91cc\u586b\u4e0b\u4e00\u4e2a batch\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u6b65\u72b6\u6001\u7531\u56db\u4e2a\u539f\u5b50\u5e03\u5c14\u6807\u5fd7\u7ba1\u7406\uff1a<code>buffer_0_is_readable_<\/code>\u3001<code>buffer_1_is_readable_<\/code>\u3001<code>buffer_0_is_writeable_<\/code>\u3001<code>buffer_1_is_writeable_<\/code>\uff0c\u914d\u5408 <code>std::condition_variable<\/code> \u5b9e\u73b0\u7b49\u5f85\uff0c\u907f\u514d\u5fd9\u8f6e\u8be2\u6d6a\u8d39 CPU\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u4e0d\u5b8c\u6574 batch \u7684 filling \u673a\u5236<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5206\u5e03\u5f0f\u8bad\u7ec3\u4e2d\uff0c\u8bad\u7ec3\u96c6\u6837\u672c\u6570\u4e0d\u4e00\u5b9a\u80fd\u88ab <code>local_batch_size \u00d7 world_size<\/code> \u6574\u9664\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c<code>TransferStation<\/code> \u5b9e\u73b0\u4e86 <strong>filling \u673a\u5236<\/strong>\uff1a\u5728\u7b2c\u4e00\u4e2a batch \u4f20\u8f93\u65f6\uff0c\u4fdd\u5b58\u7b2c\u4e00\u4e2a\u6837\u672c\u7684\u6570\u636e\u548c\u6807\u7b7e\uff1b\u5f53\u6700\u540e\u4e00\u4e2a batch \u4e0d\u6ee1\u65f6\uff0c\u7528\u8fd9\u4e2a\u4fdd\u5b58\u7684\u6837\u672c\u586b\u5145\u5230\u6ee1 batch\u3002\u5224\u65ad\u903b\u8f91\u5728 <code>configure()<\/code> \u4e2d\u9884\u5148\u8ba1\u7b97\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   num_train_samples = \u8bad\u7ec3\u96c6\u603b\u6837\u672c\u6570\n\/\/   world_size        = GPU \u5361\u6570\n\/\/   engine_id         = \u5f53\u524d TransferStation \u5bf9\u5e94\u7684 GPU \u7f16\u53f7\n\nint remainder = num_train_samples % world_size;\nbool need_filling = (remainder != 0) &amp;&amp; (engine_id >= remainder);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u6837\uff0c\u6bcf\u4e2a epoch \u7684\u6240\u6709 batch \u5927\u5c0f\u4e00\u81f4\uff0c\u6df1\u5ea6\u5b66\u4e60\u5f15\u64ce\u4e0d\u9700\u8981\u5904\u7406\u53d8\u957f batch\uff0cCUDA Graph \u4e5f\u4e0d\u9700\u8981\u4e3a\u4e0d\u5b8c\u6574 batch \u5355\u72ec\u6355\u83b7\u56fe\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u53ef\u590d\u73b0\u6027\u5f00\u5173\uff1a\u6027\u80fd\u4e0e\u786e\u5b9a\u6027\u7684\u6743\u8861<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u517c\u987e\u6027\u80fd\u4e0e\u53ef\u590d\u73b0\u6027\uff0c<code>TransferStation<\/code> \u8fd8\u5b9e\u73b0\u4e86\u4e24\u79cd\u6a21\u5f0f\uff0c\u901a\u8fc7 <code>GlobalRegistry::instance().ensure_reproducibility(true\/false)<\/code> \u5207\u6362\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728<strong>\u53ef\u590d\u73b0\u6a21\u5f0f<\/strong>\u4e0b\uff0cworker \u5fc5\u987b\u6309\u7167 <code>calculate_write_position()<\/code> \u7b97\u51fa\u7684\u56fa\u5b9a\u69fd\u4f4d\u5199\u5165\uff0c\u4f20\u8f93\u89e6\u53d1\u65f6\u673a\u4e25\u683c\u7531 batch \u6ee1\u51b3\u5b9a\uff1b\u5feb worker \u5fc5\u987b\u7b49\u5f85\u6162 worker\uff0c\u6240\u6709 worker \u5c31\u4f4d\u540e\u624d\u89e6\u53d1\u4f20\u8f93\u3002\u8fd9\u4fdd\u8bc1\u4e86\u6bcf\u6b21\u8fd0\u884c\u5199\u5165\u987a\u5e8f\u5b8c\u5168\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728<strong>\u975e\u53ef\u590d\u73b0\u6a21\u5f0f<\/strong>\u4e0b\uff0c\u4f7f\u7528 <code>request_count_<\/code> \/ <code>written_count_<\/code> \u539f\u5b50\u8ba1\u6570\u5668\u5feb\u901f\u5206\u914d slot\uff0c\u53ea\u6709\u5728\u771f\u6b63\u89e6\u53d1\u4f20\u8f93\u65f6\u624d\u52a0\u9501\uff1a<\/p>\n\n\n\n<pre class=\"EnlighterJSRAW\" data-enlighter-language=\"cpp\" data-enlighter-theme=\"\" data-enlighter-highlight=\"\" data-enlighter-linenumbers=\"\" data-enlighter-lineoffset=\"\" data-enlighter-title=\"\" data-enlighter-group=\"\">\/\/ \u53d8\u91cf\u8bf4\u660e\uff1a\n\/\/   local_batch_size_ = \u5355\u5361 batch size\n\nsize_t my_request = request_count_.fetch_add(1, std::memory_order_relaxed);\nsize_t slot     = my_request % local_batch_size_;\nint    my_batch = static_cast&lt;int>(my_request \/ local_batch_size_);<\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u751f\u4ea7\u8bad\u7ec3\u901a\u5e38\u7528\u975e\u53ef\u590d\u73b0\u6a21\u5f0f\u6362\u53d6\u6781\u9650\u541e\u5410\uff0c\u8c03\u8bd5\u548c\u6d88\u878d\u5b9e\u9a8c\u5219\u7528\u53ef\u590d\u73b0\u6a21\u5f0f\u4fdd\u8bc1\u7ed3\u679c\u7a33\u5b9a\u3002\u8fd9\u4e0e\u7b2c 22 \u7bc7\u5c06\u8981\u4ecb\u7ecd\u7684 Philox \u786e\u5b9a\u6027\u968f\u673a\u6570\u751f\u6210\u5668\u4e00\u8d77\uff0c\u6784\u6210\u4e86 Tech-Renaissance \u7aef\u5230\u7aef\u786e\u5b9a\u6027\u8bad\u7ec3\u7684\u5b8c\u6574\u62fc\u56fe\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u4e3a\u4ec0\u4e48\u6bd4 PyTorch DataLoader \u66f4\u7a33\u3001\u66f4\u53ef\u9884\u6d4b<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5199\u5230\u8fd9\u91cc\uff0c\u53ef\u80fd\u53c8\u4f1a\u6709\u4eba\u95ee\uff1a\u8fd9\u4e0d\u5c31\u662f C++ \u7248\u7684 PyTorch DataLoader \u5417\uff1f\u529f\u80fd\u4e0a\u786e\u5b9e\u6709\u76f8\u4f3c\u4e4b\u5904\uff0c\u4f46\u5b9e\u73b0\u5b9a\u4f4d\u4e0d\u540c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e00\uff0cworker \u603b\u6570\u7684\u8bed\u4e49\u4e0d\u540c<\/strong>\u3002PyTorch \u7684 <code>num_workers<\/code> \u662f<strong>\u6bcf\u4e2a GPU \u8fdb\u7a0b<\/strong>\u7684 worker \u6570\uff1bTech-Renaissance \u7684 <code>preprocess_workers<\/code> \u662f<strong>\u8de8\u6240\u6709 GPU \u7684\u603b\u7ebf\u7a0b\u6570<\/strong>\u3002\u4f8b\u5982 8 \u5361\u670d\u52a1\u5668\u3001PyTorch \u6bcf\u5f20\u5361 16 worker\uff0c\u603b\u5171 128 \u4e2a\u8fdb\u7a0b\uff1b\u5728 Tech-Renaissance \u91cc\u76f4\u63a5\u5199 <code>.preprocess_workers(128)<\/code>\uff0c\u6846\u67b6\u5185\u90e8\u901a\u8fc7 <code>engine_id = worker_id % world_size<\/code> \u628a\u6837\u672c\u5212\u5206\u5230\u5404\u4e2a GPU\u3002\u8fd9\u4e2a\u8bbe\u8ba1\u8ba9\u591a\u5361\u573a\u666f\u4e0b\u7684 CPU \u8d44\u6e90\u5206\u914d\u66f4\u7edf\u4e00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e8c\uff0c\u6837\u672c\u5206\u914d\u65b9\u5f0f\u4e0d\u540c<\/strong>\u3002PyTorch \u7684 worker \u901a\u5e38\u7531\u4e3b\u8fdb\u7a0b sampler \u751f\u6210 index\uff0c\u518d\u7ecf\u7531\u591a\u8fdb\u7a0b\u961f\u5217\u5206\u53d1\u7ed9 worker\uff0c\u8fd9\u4e2a\u5206\u53d1\u3001\u8c03\u5ea6\u4e0e\u540c\u6b65\u8fc7\u7a0b\u4ecd\u5b58\u5728\u8fd0\u884c\u65f6\u5f00\u9500\uff1bTech-Renaissance \u901a\u8fc7\u9759\u6001\u9886\u53d6\u628a\u6bcf\u4e2a worker \u7684\u6837\u672c\u4f4d\u7f6e\u9884\u5148\u786e\u5b9a\uff0c\u4e0d\u4f9d\u8d56\u8fd0\u884c\u65f6\u961f\u5217\u5206\u53d1\uff0c\u6d88\u9664\u4e86\u8fd9\u4e2a\u7ade\u4e89\u70ed\u70b9\uff0c\u4e5f\u8ba9 200+ worker \u7684\u6269\u5c55\u6027\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e09\uff0cStaging Buffer \u7684 NUMA \u611f\u77e5\u662f\u6846\u67b6\u5185\u7f6e\u7684<\/strong>\u3002PyTorch \u4e0d\u4f1a\u81ea\u52a8\u6839\u636e GPU \u62d3\u6251\u628a\u9501\u9875\u5185\u5b58\u5206\u914d\u5728\u672c\u5730 NUMA \u8282\u70b9\uff1bTech-Renaissance \u5728 <code>TR_SCENE_GPU_CLOUD<\/code> \u4e0b\u628a\u62d3\u6251\u53d1\u73b0\u3001\u672c\u5730\u8282\u70b9\u7ed1\u6838\u5206\u914d\u3001\u9875\u8868\u9884\u70ed\u4e32\u6210\u4e86\u4e00\u6761\u94fe\u8def\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u56db\uff0cH2D \u4f20\u8f93\u5185\u805a<\/strong>\u3002PyTorch DataLoader \u628a\u9884\u5904\u7406\u540e\u7684 tensor \u4ea4\u7ed9 Python \u4e3b\u8fdb\u7a0b\uff0c\u7528\u6237\u518d\u81ea\u5df1 <code>to(device)<\/code> \u53d1\u8d77 H2D\uff1bTech-Renaissance \u7684 <code>TransferStation<\/code> \u76f4\u63a5\u548c\u6df1\u5ea6\u5b66\u4e60\u5f15\u64ce\u5171\u4eab staging memory\uff0cH2D \u7531 CUDA Graph \u5b50\u56fe\u7edf\u4e00\u8c03\u5ea6\uff0c\u4e0d\u9700\u8981 Python \u4fa7\u53c2\u4e0e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7b2c\u4e94\uff0c\u53ef\u590d\u73b0\u6027\u5f00\u5173<\/strong>\u3002PyTorch \u8981\u8fbe\u5230\u7aef\u5230\u7aef\u53ef\u590d\u73b0\uff0c\u9700\u8981\u4ed4\u7ec6\u8bbe\u7f6e worker seed\u3001<code>persistent_workers<\/code>\u3001<code>generator<\/code> \u7b49\u591a\u4e2a\u7ec6\u8282\uff1bTech-Renaissance \u628a reproducibility \u505a\u6210\u4e86\u9876\u5c42\u5f00\u5173\uff0c\u9759\u6001\u9886\u53d6\u548c Philox \u8ba1\u6570\u5668\u968f\u673a\u6570\u5171\u540c\u4fdd\u8bc1\u9ad8\u5e76\u53d1\u4e0b\u4ecd\u53ef\u590d\u73b0\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u5c0f\u7ed3\uff1a\u628a\u6570\u636e\u7ba1\u7ebf\u505a\u5230\u6781\u81f4\u662f\u4e00\u95e8\u7cfb\u7edf\u5de5\u7a0b<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u52a0\u8f7d\u7ba1\u7ebf\u770b\u8d77\u6765\u4e0d\u5982\u7b97\u5b50\u878d\u5408\u3001CUDA Graph \u90a3\u4e48&#8221;\u786c\u6838&#8221;\uff0c\u4f46\u5b83\u51b3\u5b9a\u4e86 GPU \u80fd\u4e0d\u80fd\u6ee1\u8d1f\u8377\u8fd0\u8f6c\u3002Tech-Renaissance \u5728\u8fd9\u4e00\u5c42\u7684\u8bbe\u8ba1\u53ef\u4ee5\u6982\u62ec\u4e3a\u56db\u53e5\u8bdd\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6301\u4e45\u7ebf\u7a0b\u6c60<\/strong>\uff1aworker \u53ea\u521b\u5efa\u4e00\u6b21\uff0c\u8de8 buffer \u590d\u7528\uff0c\u907f\u514d\u7ebf\u7a0b\u521b\u5efa\u5f00\u9500\uff1b<\/li>\n\n\n\n<li><strong>\u9759\u6001\u9886\u53d6<\/strong>\uff1a\u6bcf\u4e2a worker \u7684\u6837\u672c\u4f4d\u7f6e\u9884\u5148\u786e\u5b9a\uff0c\u5b9e\u73b0\u96f6\u7ade\u4e89\u5e76\u53d1\uff1b<\/li>\n\n\n\n<li><strong>NUMA \u611f\u77e5\u7684 Staging Buffer<\/strong>\uff1a\u901a\u8fc7\u62d3\u6251\u53d1\u73b0\u3001\u672c\u5730\u8282\u70b9\u7ed1\u6838\u5206\u914d\u3001\u9875\u8868\u9884\u70ed\uff0c\u8ba9 H2D \u4f20\u8f93\u8def\u5f84\u6700\u77ed\uff1bworker \u7ebf\u7a0b\u505a\u786e\u5b9a\u6027 CPU \u7ed1\u6838\uff0c\u907f\u514d\u540c\u6838\u7ade\u4e89\uff1b<\/li>\n\n\n\n<li><strong>\u5f02\u6b65\u53cc\u7f13\u51b2<\/strong>\uff1a<code>TransferStation<\/code> \u8ba9 CPU \u9884\u5904\u7406\u3001H2D \u4f20\u8f93\u3001GPU \u8ba1\u7b97\u4e09\u6bb5\u6d41\u6c34\u7ebf\u91cd\u53e0\uff0c\u540c\u65f6\u7528\u53ef\u590d\u73b0\u6027\u5f00\u5173\u5728\u6027\u80fd\u4e0e\u786e\u5b9a\u6027\u4e4b\u95f4\u7075\u6d3b\u5207\u6362\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u5957\u673a\u5236\u5728\u5b9e\u6d4b\u4e2d\u53ef\u4ee5\u628a ImageNet \u9884\u5904\u7406\u7ebf\u7a0b\u63a8\u5230\u5f88\u9ad8\uff0c\u540c\u65f6\u4fdd\u6301\u7ed3\u679c\u53ef\u590d\u73b0\u3001\u4e0d\u51fa\u73b0\u6b7b\u9501\u3001\u4e0d\u4f1a\u56e0\u4e3a NUMA \u8fdc\u7aef\u8bbf\u5b58\u800c\u6389\u901f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53\u7136\uff0c\u9884\u5904\u7406\u518d\u5feb\uff0c\u5982\u679c\u6700\u540e\u4e00\u6b65\u5f52\u4e00\u5316\u3001\u7c7b\u578b\u8f6c\u6362\u3001\u968f\u673a\u7ffb\u8f6c\u8fd8\u8981\u53cd\u590d\u8bfb\u5199\u5185\u5b58\uff0c\u4e5f\u8fd8\u662f\u4f1a\u6d6a\u8d39\u5e26\u5bbd\u3002\u4e0b\u4e00\u7bc7\u300aFusedNormalization\uff1a\u628a\u6570\u636e\u589e\u5f3a\u7684\u6700\u540e\u4e00\u6b65\u878d\u8fdb\u4e00\u6b21\u5185\u5b58\u904d\u5386\u300b\uff0c\u6211\u4eec\u6765\u8bb2 Tech-Renaissance \u5982\u4f55\u628a ToTensor\u3001Normalize\u3001RandomHorizontalFlip\u3001RandomErasing \u8fd9\u4e9b\u539f\u672c\u72ec\u7acb\u7684\u6536\u5c3e\u64cd\u4f5c\uff0c\u878d\u5408\u6210\u4e00\u6b21\u5185\u5b58\u904d\u5386\uff0c\u8ba9\u6570\u636e\u4fa7\u7684\u6027\u80fd\u518d\u4e0a\u4e00\u4e2a\u53f0\u9636\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u2014\u2014\u201c\u4e00\u4e2a\u4eba\u7528AI\u5982\u4f55\u5199\u51fa\u6bd4PyTorch\u66f4\u5feb\u7684\u81ea\u7814\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u201d\u7cfb\u5217\u6587\u7ae0\u4e4b\u5341 \u524d\u9762\u6211\u4eec\u8bb2\u4e86DataLoade [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":683,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-container-style":"default","site-container-layout":"default","site-sidebar-layout":"default","disable-article-header":"default","disable-site-header":"default","disable-site-footer":"default","disable-content-area-spacing":"default","footnotes":""},"categories":[15],"tags":[],"class_list":["post-481","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-15"],"_links":{"self":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/481","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/comments?post=481"}],"version-history":[{"count":3,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/481\/revisions"}],"predecessor-version":[{"id":682,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/posts\/481\/revisions\/682"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media\/683"}],"wp:attachment":[{"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/media?parent=481"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/categories?post=481"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tech-renaissance.cn\/blog\/index.php\/wp-json\/wp\/v2\/tags?post=481"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}