{"id":1237,"date":"2026-04-20T09:27:33","date_gmt":"2026-04-20T01:27:33","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1237"},"modified":"2026-04-20T09:27:48","modified_gmt":"2026-04-20T01:27:48","slug":"tgraph-a-tensor-centric-graph-processing-framework","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1237","title":{"rendered":"TGraph: A Tensor-centric Graph Processing Framework"},"content":{"rendered":"\n

1. \u6458\u8981\uff08Abstract\uff09<\/strong><\/p>\n\n\n\n

TGraph\u662f\u9996\u4e2a\u57fa\u4e8e\u5f20\u91cf\u7684\u901a\u7528\u56fe\u5904\u7406\u6846\u67b6\uff0c\u65e8\u5728\u89e3\u51b3\u73b0\u6709\u56fe\u7cfb\u7edf\u96be\u4ee5\u8de8\u786c\u4ef6\u540e\u7aef\u8fc1\u79fb\u7684\u95ee\u9898\u3002\u4f20\u7edf\u56fe\u5904\u7406\u7cfb\u7edf\uff08\u5982Ligra\u3001Gunrock\u3001cuGraph\u7b49\uff09\u591a\u9488\u5bf9\u7279\u5b9a\u786c\u4ef6\uff08\u5982NVIDIA GPU\u6216FPGA\uff09\u8fdb\u884c\u6df1\u5ea6\u4f18\u5316\uff0c\u867d\u7136\u5728\u7279\u5b9a\u5e73\u53f0\u4e0a\u6027\u80fd\u7a81\u51fa\uff0c\u4f46\u5e95\u5c42\u5185\u6838\uff08\u5982CUDA\u3001Vitis\uff09\u5bfc\u81f4\u79fb\u690d\u6210\u672c\u6781\u9ad8\uff0c\u65e0\u6cd5\u8f7b\u677e\u9002\u914d\u65b0\u5174\u786c\u4ef6\u52a0\u901f\u5668\uff08\u5982TPU\u3001NPU\u3001AMD GPU\u3001Apple MPS\u7b49\uff09\u3002TGraph\u521b\u65b0\u6027\u5730\u5c06\u56fe\u8ba1\u7b97\u6784\u5efa\u5728\u5f20\u91cf\u8ba1\u7b97\u8fd0\u884c\u65f6\uff08Tensor Computation Runtimes\uff0c\u7b80\u79f0TCRs\uff09\u4e4b\u4e0a\uff0c\u5229\u7528PyTorch\u3001TensorFlow\u7b49\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u53ca\u5176\u7f16\u8bd1\u5668\u548c\u8fd0\u884c\u65f6\u63d0\u4f9b\u7684\u5f20\u91cf\u63a5\u53e3\uff0c\u5b9e\u73b0\u5bf9\u591a\u79cdXPU\uff08\u7edf\u4e00\u79f0\u547c\u786c\u4ef6\u52a0\u901f\u5668\uff09\u7684\u65e0\u7f1d\u652f\u6301\uff0c\u540c\u65f6\u65e0\u9700\u7528\u6237\u6df1\u5165\u5e95\u5c42\u786c\u4ef6\u7279\u6027\u3002<\/p>\n\n\n\n

\u6846\u67b6\u7684\u6838\u5fc3\u8d21\u732e\u5305\u62ec\u4e09\u4e2a\u65b9\u9762\uff1a\u4e00\u662f\u63d0\u51fa\u5f20\u91cf\u4e2d\u5fc3\u7684\u8ba1\u7b97\u6a21\u578b\uff0c\u901a\u8fc7TENSORIZE\u548cCOMPUTE\u4e24\u4e2a\u9ad8\u5c42\u63a5\u53e3\u5c06\u56fe\u7b97\u6cd5\u5206\u89e3\u4e3a\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u652f\u6301BFS\u3001WCC\u3001SSSP\u3001PageRank\u7b49\u591a\u79cd\u7ecf\u5178\u7b97\u6cd5\uff1b\u4e8c\u662f\u62bd\u8c61\u51fa\u4e00\u7ec4\u56fe\u7b97\u5b50\uff08vertexSelect\u3001neighborSelect\u3001reconstruct\u3001aggregate\u3001update\uff09\uff0c\u5c06\u8ba1\u7b97\u6a21\u578b\u4e0e\u5e95\u5c42\u5f20\u91cf\u7b97\u5b50\u89e3\u8026\uff0c\u786e\u4fdd\u6846\u67b6\u5728\u4e0d\u540cTCRs\u95f4\u8f7b\u677e\u8fc1\u79fb\uff1b\u4e09\u662f\u8bbe\u8ba1\u5f20\u91cf\u9a71\u52a8\u7684\u56fe\u538b\u7f29\u7b56\u7565\u548c\u5185\u5b58\u5916\u8ba1\u7b97\u7b56\u7565\uff0c\u89e3\u51b3XPU\u5185\u5b58\u53d7\u9650\u95ee\u9898\uff0c\u5b9e\u73b0\u5bf9\u8d85\u5927\u89c4\u6a21\u56fe\u7684\u9ad8\u6548\u5904\u7406\u3002<\/p>\n\n\n\n

\u5927\u91cf\u5b9e\u9a8c\u572813\u4e2a\u771f\u5b9e\u4e16\u754c\u6570\u636e\u96c6\uff08\u5305\u62eccit-Patents\u3001soc-twitter\u7b49\uff09\u4e0a\u8fdb\u884c\uff0c\u4e0e7\u4e2a\u6700\u5148\u8fdb\u56fe\u7cfb\u7edf\uff08Gunrock\u3001cuGraph\u3001GraphBLAST\u3001Subway\u3001Galois\u3001Ligra\u7b49\uff09\u5bf9\u6bd4\uff0c\u7ed3\u679c\u8868\u660eTGraph\u4e0d\u4ec5\u5728\u6027\u80fd\u4e0a\u5168\u9762\u9886\u5148\uff0c\u8fd8\u6210\u529f\u90e8\u7f72\u4e8ePyTorch\u548cTensorFlow\u4e24\u5927\u6846\u67b6\uff0c\u4ee5\u53caNVIDIA GPU\u3001AMD GPU\u3001Apple MPS\u4e09\u79cd\u786c\u4ef6\u540e\u7aef\u3002\u8be5\u5de5\u4f5c\u586b\u8865\u4e86\u5f20\u91cf\u4e2d\u5fc3\u56fe\u5904\u7406\u6846\u67b6\u7684\u7a7a\u767d\uff0c\u4e3a\u5f02\u6784\u786c\u4ef6\u65f6\u4ee3\u56fe\u8ba1\u7b97\u63d0\u4f9b\u4e86\u53ef\u6269\u5c55\u3001\u53ef\u79fb\u690d\u7684\u7edf\u4e00\u89e3\u51b3\u65b9\u6848\uff0c\u5177\u6709\u91cd\u8981\u7684\u7406\u8bba\u521b\u65b0\u4ef7\u503c\u548c\u5de5\u7a0b\u5b9e\u7528\u610f\u4e49\u3002<\/p>\n\n\n\n

\"TGraph:<\/figure>\n\n\n\n
\"TGraph:<\/figure>\n\n\n\n
\"TGraph:<\/figure>\n\n\n\n

2. \u7814\u7a76\u80cc\u666f\u4e0e\u95ee\u9898\u52a8\u673a\uff08Introduction\uff09<\/strong><\/p>\n\n\n\n

\u56fe\u6570\u636e\u5728\u793e\u4ea4\u7f51\u7edc\u3001\u63a8\u8350\u7cfb\u7edf\u3001\u751f\u7269\u4fe1\u606f\u5b66\u7b49\u73b0\u5b9e\u5e94\u7528\u4e2d\u65e0\u5904\u4e0d\u5728\uff0c\u8fc7\u53bb\u5341\u5e74\u6d8c\u73b0\u51fa\u5927\u91cf\u56fe\u5904\u7406\u7cfb\u7edf\uff0c\u5927\u81f4\u53ef\u5206\u4e3a\u5171\u4eab\u5185\u5b58\u7cfb\u7edf\uff08Ligra\u3001Galois\u3001GraphChi\u7b49\uff09\u548c\u5206\u5e03\u5f0f\u7cfb\u7edf\uff08Pregel\u3001PowerGraph\u3001GraphX\u7b49\uff09\u3002\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u786c\u4ef6\u52a0\u901f\u5668\u7684\u53d1\u5c55\uff0cGPU\u3001FPGA\u7b49\u4e13\u7528\u8bbe\u5907\u88ab\u5e7f\u6cdb\u7528\u4e8e\u52a0\u901f\u56fe\u8ba1\u7b97\uff0c\u4ee3\u8868\u6027\u5de5\u4f5c\u5305\u62ecMedusa\u3001CuSha\u3001Gunrock\u3001cuGraph\u3001GraphBLAST\u7b49\u3002\u8fd9\u4e9b\u7cfb\u7edf\u901a\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u8d1f\u8f7d\u5747\u8861\u3001\u5185\u5b58\u8bbf\u95ee\u4f18\u5316\u548c\u5e76\u884c\u539f\u8bed\uff0c\u663e\u8457\u63d0\u5347\u4e86\u56fe\u7b97\u6cd5\u7684\u6267\u884c\u6548\u7387\u3002\u7136\u800c\uff0c\u6240\u6709\u8fd9\u4e9b\u7cfb\u7edf\u90fd\u9762\u4e34\u4e00\u4e2a\u5171\u540c\u7684\u81f4\u547d\u5c40\u9650\uff1a\u9ad8\u5ea6\u786c\u4ef6\u7279\u5b9a\u5316\u3002<\/p>\n\n\n\n

NVIDIA GPU\u7cfb\u7edf\u4f9d\u8d56CUDA\u5185\u6838\uff0cAMD GPU\u6216Apple MPS\u5219\u9700\u5b8c\u5168\u4e0d\u540c\u7684\u7f16\u7a0b\u63a5\u53e3\uff0cFPGA\u7cfb\u7edf\u66f4\u9700Vitis\u7b49\u4e13\u7528\u5de5\u5177\u94fe\u3002\u8fd9\u79cd\u786c\u4ef6\u7ed1\u5b9a\u5bfc\u81f4\u7cfb\u7edf\u79fb\u690d\u6210\u672c\u6781\u9ad8\uff0c\u65e0\u6cd5\u8ddf\u4e0a\u6df1\u5ea6\u5b66\u4e60\u9a71\u52a8\u7684\u65b0\u5174\u52a0\u901f\u5668\uff08\u5982TPU\u3001NPU\uff09\u7684\u5feb\u901f\u53d1\u5c55\u3002\u7528\u6237\u6bcf\u66f4\u6362\u4e00\u6b21\u786c\u4ef6\u540e\u7aef\uff0c\u5c31\u5fc5\u987b\u91cd\u5199\u5927\u91cf\u4f4e\u5c42\u5185\u6838\uff0c\u6781\u5927\u963b\u788d\u4e86\u56fe\u8ba1\u7b97\u5728\u5f02\u6784\u8ba1\u7b97\u73af\u5883\u4e2d\u7684\u666e\u53ca\u3002\u540c\u65f6\uff0c\u73b0\u6709\u7cfb\u7edf\u5728\u8868\u8fbe\u6027\u3001\u53ef\u6269\u5c55\u6027\u548c\u53ef\u79fb\u690d\u6027\u4e0a\u4e5f\u5b58\u5728\u4e0d\u8db3\uff1a\u8bb8\u591a\u7cfb\u7edf\u9488\u5bf9\u7ec6\u7c92\u5ea6\u9876\u70b9\/\u8fb9\u64cd\u4f5c\u8bbe\u8ba1\uff0c\u4e0d\u9002\u5408\u5f20\u91cf\u7ea7\u6279\u91cf\u5e76\u884c\uff1b\u90e8\u5206\u7cfb\u7edf\u867d\u652f\u6301\u5b50\u56fe\u4e2d\u5fc3\u6a21\u578b\uff0c\u4f46\u4ecd\u4f9d\u8d56\u786c\u4ef6\u7279\u5b9a\u6570\u636e\u7ed3\u6784\uff0c\u65e0\u6cd5\u8de8\u5e73\u53f0\u8fc1\u79fb\u3002<\/p>\n\n\n\n

TGraph\u6b63\u662f\u9488\u5bf9\u4e0a\u8ff0\u75db\u70b9\u800c\u8bbe\u8ba1\u3002\u5b83\u9996\u6b21\u5c06\u56fe\u8ba1\u7b97\u5b8c\u5168\u6784\u5efa\u5728\u5f20\u91cf\u4e4b\u4e0a\uff0c\u5229\u7528TCRs\uff08\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u53ca\u5176\u8fd0\u884c\u65f6\uff09\u63d0\u4f9b\u7684\u786c\u4ef6\u65e0\u5173\u5f20\u91cf\u7b97\u5b50\u548c\u81ea\u52a8\u4f18\u5316\u80fd\u529b\uff08\u5982\u7b97\u5b50\u878d\u5408\u3001\u7b97\u5b50\u4e0b\u6c89\u3001\u4ee3\u6570\u7b80\u5316\uff09\uff0c\u5b9e\u73b0\u201c\u4e00\u5904\u7f16\u5199\u3001\u5904\u5904\u8fd0\u884c\u201d\u3002\u901a\u8fc7\u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u578b\u548c\u62bd\u8c61\u56fe\u7b97\u5b50\uff0cTGraph\u65e2\u4fdd\u7559\u4e86\u56fe\u7b97\u6cd5\u7684\u8868\u8fbe\u7075\u6d3b\u6027\uff0c\u53c8\u5b9e\u73b0\u4e86\u5bf9\u591a\u79cdXPU\u7684\u65e0\u7f1d\u9002\u914d\uff0c\u540c\u65f6\u901a\u8fc7\u538b\u7f29\u548c\u5185\u5b58\u5916\u7b56\u7565\u89e3\u51b3\u4e86\u5927\u89c4\u6a21\u56fe\u7684\u5185\u5b58\u74f6\u9888\u3002\u8be5\u5de5\u4f5c\u4e0d\u4ec5\u89e3\u51b3\u4e86\u786c\u4ef6\u5f02\u6784\u6027\u5e26\u6765\u7684\u79fb\u690d\u96be\u9898\uff0c\u4e5f\u4e3a\u672a\u6765\u56fe\u8ba1\u7b97\u5728AI\u9a71\u52a8\u786c\u4ef6\u751f\u6001\u4e2d\u7684\u53d1\u5c55\u63d0\u4f9b\u4e86\u5168\u65b0\u8303\u5f0f\u3002<\/p>\n\n\n\n

3. \u7cfb\u7edf\u67b6\u6784\u4e0e\u9ad8\u5c42\u8bbe\u8ba1\uff08System Overview\uff09<\/strong><\/p>\n\n\n\n

TGraph\u91c7\u7528\u4e09\u5c42\u67b6\u6784\uff0c\u5e95\u5c42\u4e3aTCRs\u548c\u786c\u4ef6\u5c42\uff08\u652f\u6301PyTorch\u3001TensorFlow\u3001TVM\u7b49\u6846\u67b6\u53caNVIDIA GPU\u3001AMD GPU\u3001Apple MPS\u7b49XPU\uff09\uff0c\u4e2d\u95f4\u4e3a\u8ba1\u7b97\u5c42\uff08\u6838\u5fc3\u662f\u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u5757\u548c\u5185\u5b58\u5916\u8ba1\u7b97\u6a21\u5757\uff09\uff0c\u4e0a\u5c42\u4e3a\u5e94\u7528\u5c42\uff08\u652f\u6301BFS\u3001WCC\u3001SSSP\u3001PageRank\u3001HITS\u7b49\u7b97\u6cd5\uff09\u3002\u5b58\u50a8\u5c42\u63d0\u4f9b\u5f20\u91cf\u5316\u7684\u56fe\u8868\u793a\uff08CSR\/COO\uff09\u548c\u538b\u7f29\u6a21\u5757\uff0c\u8fdb\u4e00\u6b65\u8282\u7701\u7a7a\u95f4\u5e76\u52a0\u901f\u8ba1\u7b97\u3002<\/p>\n\n\n\n

\u7cfb\u7edf\u6838\u5fc3\u662f\u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u578b\uff0c\u5c06\u6bcf\u4e2a\u8fed\u4ee3\u5206\u89e3\u4e3aTENSORIZE\uff08\u5c06\u6d3b\u8dc3\u9876\u70b9\u53ca\u5176\u90bb\u5c45\u7ec4\u7ec7\u4e3a\u5927\u5c3a\u5bf8\u4e00\u7ef4\u5f20\u91cf\uff09\u548cCOMPUTE\uff08\u57fa\u4e8e\u5f20\u91cf\u7b97\u5b50\u8fdb\u884c\u805a\u5408\u3001\u66f4\u65b0\u548c\u6d3b\u8dc3\u6027\u5224\u65ad\uff09\u4e24\u4e2a\u6b65\u9aa4\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5145\u5206\u5229\u7528\u5f20\u91cf\u7b97\u5b50\u7684\u6279\u91cf\u5e76\u884c\u80fd\u529b\uff0c\u540c\u65f6\u901a\u8fc7\u62bd\u8c61\u7684\u4e94\u79cd\u56fe\u7b97\u5b50\uff08vertexSelect\u3001neighborSelect\u3001reconstruct\u3001aggregate\u3001update\uff09\u5c4f\u853d\u5e95\u5c42\u5f20\u91cf\u7ec6\u8282\uff0c\u786e\u4fdd\u8ba1\u7b97\u6a21\u578b\u4e0e\u5177\u4f53TCRs\u89e3\u8026\u3002<\/p>\n\n\n\n

\u4e3a\u5904\u7406\u8d85\u5927\u89c4\u6a21\u56fe\uff0cTGraph\u5728\u5b58\u50a8\u5c42\u5f15\u5165\u5f20\u91cf\u9a71\u52a8\u7684\u56fe\u538b\u7f29\u7b56\u7565\uff0c\u901a\u8fc7\u865a\u62df\u9876\u70b9\u9012\u5f52\u66ff\u6362\u91cd\u590d\u90bb\u5c45\u5e8f\u5217\uff0c\u5b9e\u73b0\u591a\u5c42CSR\u8868\u793a\uff1b\u5728\u8ba1\u7b97\u5c42\u8bbe\u8ba1\u5185\u5b58\u5916\u8ba1\u7b97\u7b56\u7565\uff0c\u5305\u62ec\u8fb9\u5747\u8861\u5206\u533a\uff08EBP\uff09\u548c\u826f\u597d\u8fde\u63a5\u5206\u533a\uff08WCP\uff09\uff0c\u7ed3\u5408\u6d41\u6c34\u7ebf\u8c03\u5ea6\u673a\u5236\uff0c\u5b9e\u73b0\u4e3b\u673a\u5185\u5b58\u4e0eXPU\u5185\u5b58\u7684\u9ad8\u6548\u534f\u540c\u3002\u8be5\u67b6\u6784\u65e2\u4fdd\u8bc1\u4e86\u9ad8\u6027\u80fd\uff0c\u53c8\u5b9e\u73b0\u4e86\u8de8\u786c\u4ef6\u3001\u8de8\u6846\u67b6\u7684\u6781\u81f4\u53ef\u6269\u5c55\u6027\uff0c\u662fTGraph\u533a\u522b\u4e8e\u4f20\u7edf\u56fe\u7cfb\u7edf\u7684\u5173\u952e\u521b\u65b0\u70b9\u3002<\/p>\n\n\n\n

\"TGraph:<\/figure>\n\n\n\n
\"TGraph:<\/figure>\n\n\n\n

4. \u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u578b\u4e0e\u56fe\u7b97\u5b50\uff08Tensor-centric Computation Model\uff09<\/strong><\/p>\n\n\n\n

TGraph\u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u63d0\u51fa\u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u578b\uff0c\u5c06\u56fe\u7b97\u6cd5\u62bd\u8c61\u4e3a\u8fed\u4ee3\u8fc7\u7a0b\uff1aTENSORIZE\u8d1f\u8d23\u4ece\u5b8c\u6574\u56fe\u4e2d\u63d0\u53d6\u6d3b\u8dc3\u5b50\u56fe\u5e76\u7ec4\u7ec7\u4e3a\u5f20\u91cf\uff0cCOMPUTE\u5219\u5728\u5b50\u56fe\u4e0a\u6267\u884c\u805a\u5408\u3001\u66f4\u65b0\u548c\u6d3b\u8dc3\u6027\u6807\u8bb0\u3002\u7b97\u6cd51\u7ed9\u51fa\u4e86\u6982\u5ff5\u6d41\u7a0b\uff1a\u9996\u5148\u521d\u59cb\u5316\u9876\u70b9\u6570\u636evData\u548c\u6d3b\u8dc3\u63a9\u7801actMask\uff0c\u968f\u540e\u5faa\u73af\u8c03\u7528TENSORIZE\u548cCOMPUTE\uff0c\u76f4\u81f3\u65e0\u6d3b\u8dc3\u9876\u70b9\u3002<\/p>\n\n\n\n

\u4e3a\u65b9\u4fbf\u7528\u6237\u5b9e\u73b0\uff0cTGraph\u8fdb\u4e00\u6b65\u62bd\u8c61\u4e94\u79cd\u56fe\u7b97\u5b50\u3002vertexSelect\u7528\u4e8e\u4ece\u9876\u70b9\u96c6\u4e2d\u9009\u53d6\u5b50\u96c6\uff1bneighborSelect\u9ad8\u6548\u63d0\u53d6\u6307\u5b9a\u9876\u70b9\u7684\u90bb\u5c45\uff1breconstruct\u5c06\u9876\u70b9\u5b50\u96c6\u53ca\u5176\u90bb\u5c45\u91cd\u6784\u4e3a\u5f20\u91cf\u5b50\u56fe\uff1baggregate\u652f\u6301push\/pull\u4e24\u79cd\u6a21\u5f0f\u5bf9\u90bb\u5c45\u6570\u636e\u8fdb\u884cmin\/sum\u7b49\u805a\u5408\uff1bupdate\u5219\u6839\u636e\u805a\u5408\u7ed3\u679c\u66f4\u65b0\u9876\u70b9\u503c\u548c\u6d3b\u8dc3\u63a9\u7801\u3002\u8fd9\u4e9b\u7b97\u5b50\u5168\u90e8\u57fa\u4e8e\u6807\u51c6\u5f20\u91cf\u7b97\u5b50\uff08\u5982index_select\u3001repeat_interleave\u3001scatter_reduce\u3001segment_csr\u7b49\uff09\u5b9e\u73b0\uff0c\u65e2\u4fdd\u8bc1\u4e86\u8868\u8fbe\u6027\uff0c\u53c8\u5b9e\u73b0\u4e86\u4e0e\u5e95\u5c42TCRs\u7684\u5b8c\u5168\u89e3\u8026\u3002<\/p>\n\n\n\n

\u7b97\u6cd5\u8868\u8fbe\u6027\u5206\u6790\u8868\u660e\uff0c\u8fd9\u4e9b\u7b97\u5b50\u53ef\u8986\u76d6\u7ebf\u6027\u4ee3\u6570\u56fe\u7cfb\u7edf\uff08\u5982GraphBLAS\u3001GraphBLAST\uff09\u7684\u5927\u591a\u6570\u539f\u8bed\uff0c\u652f\u6301\u5e7f\u6cdb\u7684\u8fed\u4ee3\u7c7b\u56fe\u7b97\u6cd5\u3002\u5b9e\u9645\u5b9e\u73b0\u4e2d\uff0cTGraph\u901a\u8fc7\u52a8\u6001\u5207\u6362push\/pull\u6a21\u5f0f\u548c\u4f18\u5316\u5173\u952e\u5f20\u91cf\u7b97\u5b50\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u3002\u4f8b\u5982\uff0caggregate\u5728\u6d3b\u8dc3\u9876\u70b9\u8f83\u591a\u65f6\u81ea\u52a8\u5207\u6362\u5230pull\u6a21\u5f0f\u907f\u514d\u539f\u5b50\u64cd\u4f5c\uff0c\u5728\u6d3b\u8dc3\u9876\u70b9\u8f83\u5c11\u65f6\u4f7f\u7528push\u6a21\u5f0f\u51cf\u5c11\u65e0\u6548\u8ba1\u7b97\u3002\u8fd9\u79cd\u8bbe\u8ba1\u65e2\u7b80\u5316\u4e86\u7528\u6237\u7f16\u7a0b\uff0c\u53c8\u6700\u5927\u5316\u4e86\u5f20\u91cf\u5e76\u884c\u6f5c\u529b\u3002<\/p>\n\n\n\n

\"TGraph:<\/figure>\n\n\n\n

5. \u6269\u5c55\u7b56\u7565\uff1a\u56fe\u538b\u7f29\u4e0e\u5185\u5b58\u5916\u8ba1\u7b97\uff08Scaling Strategies\uff09<\/strong><\/p>\n\n\n\n

\u9488\u5bf9XPU\u5185\u5b58\u53d7\u9650\u95ee\u9898\uff0cTGraph\u63d0\u51fa\u4e24\u5927\u6269\u5c55\u7b56\u7565\u3002\u9996\u5148\u662f\u5f20\u91cf\u9a71\u52a8\u7684\u56fe\u538b\u7f29\uff1a\u8bc6\u522b\u9ad8\u9891\u91cd\u590d\u90bb\u5c45\u5e8f\u5217\u5e76\u7528\u865a\u62df\u9876\u70b9\u66ff\u6362\uff0c\u5f62\u6210\u591a\u5c42\u5d4c\u5957CSR\u8868\u793a\u3002\u538b\u7f29\u540e\u56fe\u4ee5\u5206\u5c42\u5f20\u91cfCSR\u5b58\u50a8\uff0c\u5e76\u5f15\u5165virtualMask\u6807\u8bb0\u865a\u62df\u9876\u70b9\u3002\u8ba1\u7b97\u65f6\u9700\u6309\u5c42\u7ea7\u4ece\u4f4e\u5230\u9ad8\uff08pull\u6a21\u5f0f\uff09\u6216\u9ad8\u5230\u4f4e\uff08push\u6a21\u5f0f\uff09\u904d\u5386\uff0c\u786e\u4fdd\u6570\u636e\u4f9d\u8d56\u6b63\u786e\u3002\u8be5\u7b56\u7565\u5728AR\u548cIT\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u4e864.35~4.65\u500d\u538b\u7f29\u6bd4\uff0c\u663e\u8457\u964d\u4f4e\u5b58\u50a8\u5f00\u9500\u5e76\u52a0\u901f\u8ba1\u7b97\u3002<\/p>\n\n\n\n

\u5176\u6b21\u662f\u5185\u5b58\u5916\u8ba1\u7b97\u7b56\u7565\uff1a\u5148\u901a\u8fc7\u8fb9\u5747\u8861\u5206\u533a\uff08EBP\uff0c\u4f7f\u7528searchsorted\u5b9e\u73b0\u5feb\u901f\u5e73\u8861\u8fb9\u6570\uff09\u6216\u826f\u597d\u8fde\u63a5\u5206\u533a\uff08WCP\uff0c\u901a\u8fc7\u591a\u8f6e\u591a\u6e90BFS+\u5408\u5e76\u751f\u6210\u9ad8\u5185\u805a\u5b50\u56fe\uff09\u5c06\u56fe\u5212\u5206\u4e3a\u591a\u4e2a\u5b50\u56fe\uff0c\u968f\u540e\u91c7\u7528\u6d41\u6c34\u7ebf\u8c03\u5ea6\u673a\u5236\u3002LoadQueue\u5b58\u50a8\u5f85\u8c03\u5ea6\u5b50\u56feID\uff0c\u52a0\u8f7d\u7ebf\u7a0b\u8d1f\u8d23\u4ece\u4e3b\u673a\u5185\u5b58\u52a0\u8f7d\u5b50\u56fe\u5230ComputeQueue\uff0c\u8ba1\u7b97\u7ebf\u7a0b\u5219\u5728XPU\u4e0a\u6267\u884c\u5b50\u56fe\u8ba1\u7b97\uff0c\u540c\u65f6\u5229\u7528\u5171\u4eab\u6570\u636e\u533a\u5b9e\u73b0\u5b50\u56fe\u95f4\u6d88\u606f\u4f20\u9012\u3002\u8fd9\u79cd\u91cd\u53e0\u52a0\u8f7d\u4e0e\u8ba1\u7b97\u7684\u8bbe\u8ba1\uff0c\u6781\u5927\u51cf\u5c11\u4e86\u6570\u636e\u4f20\u8f93\u5f00\u9500\uff0c\u5b9e\u73b0\u4e86\u5bf9TB\u7ea7\u56fe\u7684\u9ad8\u6548\u5904\u7406\u3002<\/p>\n\n\n\n

\"TGraph:<\/figure>\n\n\n\n
\"TGraph:<\/figure>\n\n\n\n

6. \u5b9e\u9a8c\u8bc4\u4f30\u4e0e\u6027\u80fd\u5206\u6790\uff08Experiments\uff09<\/strong><\/p>\n\n\n\n

\u5b9e\u9a8c\u572813\u4e2a\u771f\u5b9e\u4e16\u754c\u6570\u636e\u96c6\uff08\u4ececit-Patents\u5230sk-2005\uff0c\u8fb9\u6570\u4ece\u5343\u4e07\u5230\u6570\u5341\u4ebf\uff09\u4e0a\u8fdb\u884c\uff0c\u4e0e7\u4e2a\u6700\u5148\u8fdb\u7cfb\u7edf\uff08Gunrock\u3001cuGraph\u3001GraphBLAST\u3001Subway\u3001Galois\u3001Ligra\u7b49\uff09\u5168\u9762\u5bf9\u6bd4\u3002\u786c\u4ef6\u73af\u5883\u4e3aIntel Xeon Gold 6330 CPU + NVIDIA RTX 3090 GPU\uff0824GB\u663e\u5b58\uff09\uff0c\u8f6f\u4ef6\u57fa\u4e8ePyTorch 1.11\u3002<\/p>\n\n\n\n

\u6574\u4f53\u6027\u80fd\u7ed3\u679c\u663e\u793a\uff0cTGraphG\u5728PR\u3001HITS\u7b49\u8ba1\u7b97\u5bc6\u96c6\u578b\u7b97\u6cd5\u4e0a\u9886\u5148\u6700\u591a\uff0c\u5728BFS\u3001WCC\u7b49\u5185\u5b58\u5bc6\u96c6\u578b\u7b97\u6cd5\u4e0a\u4e5f\u4fdd\u6301\u7ade\u4e89\u529b\u3002Nsight Compute\u6027\u80fd\u5256\u6790\u8bc1\u5b9eTGraphG\u901a\u8fc7\u52a8\u6001push\/pull\u5207\u6362\u907f\u514d\u539f\u5b50\u64cd\u4f5c\uff0c\u663e\u8457\u63d0\u5347L1\/L2\/DRAM\u7f13\u5b58\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n

\u53ef\u6269\u5c55\u6027\u8bc4\u4f30\u8868\u660e\uff0c\u56fe\u538b\u7f29\u7b56\u7565\u5728AR\/IT\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0\u9ad8\u6548\u5904\u7406\uff1b\u5185\u5b58\u5916\u8ba1\u7b97\u5728TW\/GS\/SK\u7b49\u8d85\u5927\u56fe\u4e0a\u6027\u80fd\u4f18\u4e8eLigra\u3001Galois\u548cSubway\u3002\u8de8\u786c\u4ef6\/\u6846\u67b6\u6d4b\u8bd5\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86TGraph\u5728AMD GPU\u3001Apple MPS\u3001V100\u4ee5\u53caTensorFlow\u4e0a\u7684\u65e0\u7f1d\u90e8\u7f72\uff0c\u5145\u5206\u8bc1\u660e\u4e86\u6846\u67b6\u7684\u6781\u81f4\u53ef\u79fb\u690d\u6027\u3002\u6d88\u878d\u5b9e\u9a8c\u548c\u6210\u672c\u6548\u76ca\u5206\u6790\u4e5f\u663e\u793a\uff0cTGraph\u5728\u6027\u80fd\u3001\u5185\u5b58\u548c\u5f00\u53d1\u6548\u7387\u4e0a\u5747\u53d6\u5f97\u826f\u597d\u5e73\u8861\u3002<\/p>\n\n\n\n

\"TGraph:<\/figure>\n\n\n\n
\"TGraph:<\/figure>\n\n\n\n

7. \u8d21\u732e\u4e0e\u7ed3\u8bba\uff08Contributions and Conclusion\uff09<\/strong><\/p>\n\n\n\n

TGraph\u7684\u4e3b\u8981\u8d21\u732e\u5305\u62ec\uff1a\uff081\uff09\u63d0\u51fa\u5f20\u91cf\u4e2d\u5fc3\u8ba1\u7b97\u6a21\u578b\u53caTENSORIZE\/COMPUTE\u63a5\u53e3\uff0c\u652f\u6301\u9ad8\u6548\u56fe\u7b97\u6cd5\u5b9e\u73b0\uff1b\uff082\uff09\u62bd\u8c61\u4e94\u79cd\u56fe\u7b97\u5b50\uff0c\u5b9e\u73b0\u8ba1\u7b97\u6a21\u578b\u4e0e\u5e95\u5c42\u5f20\u91cf\u7b97\u5b50\u7684\u5b8c\u5168\u89e3\u8026\uff0c\u786e\u4fdd\u8de8TCRs\u8fc1\u79fb\uff1b\uff083\uff09\u8bbe\u8ba1\u5f20\u91cf\u9a71\u52a8\u7684\u56fe\u538b\u7f29\u548c\u5185\u5b58\u5916\u8ba1\u7b97\u7b56\u7565\uff0c\u89e3\u51b3\u5927\u89c4\u6a21\u56fe\u5904\u7406\u96be\u9898\uff1b\uff084\uff09\u5b9e\u73b0\u9996\u4e2a\u53ef\u8de8DL\u6846\u67b6\u548cXPU\u90e8\u7f72\u7684\u56fe\u5904\u7406\u6846\u67b6\uff1b\uff085\uff09\u901a\u8fc7\u5168\u9762\u5b9e\u9a8c\u9a8c\u8bc1TGraph\u5728\u6027\u80fd\u3001\u53ef\u6269\u5c55\u6027\u548c\u53ef\u79fb\u690d\u6027\u4e0a\u7684\u5168\u9762\u4f18\u52bf\u3002<\/p>\n\n\n\n

\u8bba\u6587\u7ed3\u8bba\u6307\u51fa\uff0cTGraph\u6210\u529f\u5c06\u6df1\u5ea6\u5b66\u4e60\u751f\u6001\u4e2d\u7684\u5f20\u91cf\u8ba1\u7b97\u80fd\u529b\u5f15\u5165\u56fe\u5904\u7406\u9886\u57df\uff0c\u4e3a\u5f02\u6784\u786c\u4ef6\u65f6\u4ee3\u63d0\u4f9b\u4e86\u7edf\u4e00\u3001\u9ad8\u6548\u3001\u53ef\u79fb\u690d\u7684\u56fe\u8ba1\u7b97\u89e3\u51b3\u65b9\u6848\u3002\u5c3d\u7ba1\u5f53\u524d\u4e3b\u8981\u805a\u7126\u5355\u673a\u56fe\u5206\u6790\uff0c\u4f46\u6846\u67b6\u7684\u5f00\u653e\u8bbe\u8ba1\u4e3a\u672a\u6765\u5206\u5e03\u5f0f\u3001\u591a\u6a21\u6001\u56fe\u8ba1\u7b97\u7b49\u6269\u5c55\u5960\u5b9a\u4e86\u575a\u5b9e\u57fa\u7840\u3002\u8be5\u5de5\u4f5c\u4e0d\u4ec5\u5177\u6709\u91cd\u8981\u7684\u5b66\u672f\u521b\u65b0\u4ef7\u503c\uff0c\u4e5f\u4e3a\u5de5\u4e1a\u754c\u5728AI\u786c\u4ef6\u751f\u6001\u4e2d\u90e8\u7f72\u56fe\u8ba1\u7b97\u63d0\u4f9b\u4e86\u53ef\u843d\u5730\u7684\u6280\u672f\u8def\u5f84\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

1. \u6458\u8981\uff08Abstract\uff09 TGraph\u662f\u9996\u4e2a\u57fa\u4e8e\u5f20\u91cf\u7684\u901a\u7528\u56fe\u5904\u7406\u6846\u67b6\uff0c\u65e8\u5728\u89e3\u51b3\u73b0\u6709\u56fe\u7cfb\u7edf\u96be\u4ee5\u8de8\u786c\u4ef6\u540e\u7aef\u8fc1\u79fb\u7684\u95ee\u9898\u3002\u4f20\u7edf\u56fe\u5904\u7406\u7cfb\u7edf\uff08\u5982Ligra\u3001Gunrock\u3001cuGraph\u7b49\uff09\u591a\u9488\u5bf9\u7279\u5b9a\u786c\u4ef6\uff08\u5982NVIDIA GPU\u6216FPGA\uff09\u8fdb\u884c\u6df1\u5ea6\u4f18\u5316\uff0c\u867d\u7136\u5728\u7279\u5b9a\u5e73\u53f0\u4e0a\u6027\u80fd\u7a81\u51fa\uff0c\u4f46\u5e95\u5c42\u5185\u6838\uff08\u5982CUDA\u3001Vitis\uff09\u5bfc\u81f4\u79fb\u690d\u6210\u672c\u6781\u9ad8\uff0c\u65e0\u6cd5\u8f7b\u677e\u9002\u914d\u65b0\u5174\u786c\u4ef6\u52a0\u901f\u5668\uff08\u5982TPU\u3001NPU\u3001AMD GPU\u3001Apple MPS\u7b49\uff09\u3002TGraph\u521b\u65b0\u6027\u5730\u5c06\u56fe\u8ba1\u7b97\u6784\u5efa\u5728\u5f20\u91cf\u8ba1\u7b97\u8fd0\u884c\u65f6\uff08Tensor C … \u7ee7\u7eed\u9605\u8bfb →<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1248,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,5],"tags":[14],"class_list":["post-1237","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-icn","category-rengongzhineng","tag-ien"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1237","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1237"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1237\/revisions"}],"predecessor-version":[{"id":1249,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1237\/revisions\/1249"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1248"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1237"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1237"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1237"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}