{"id":1061985,"date":"2023-09-01T02:35:00","date_gmt":"2023-09-01T09:35:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1061985"},"modified":"2024-09-25T03:44:06","modified_gmt":"2024-09-25T10:44:06","slug":"changho-hwang","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/changho-hwang\/","title":{"rendered":"\u9ec4\u6636\u4e92\uff1a\u575a\u6301\u957f\u671f\u4e3b\u4e49\u7814\u7a76\uff0c\u662f\u4e00\u4e2a\u4e0d\u65ad\u8bf4\u670d\u81ea\u5df1\u7684\u8fc7\u7a0b"},"content":{"rendered":"\n

\u7f16\u8005\u6309\uff1a\u4ed6\uff0c\u521a\u5165\u804c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e00\u5e74\uff0c\u5374\u6709\u7740\u4e30\u5bcc\u7684\u5b66\u672f\u5408\u4f5c\u7ecf\u9a8c\uff1b\u521a\u521a\u535a\u58eb\u6bd5\u4e1a\uff0c\u5c31\u6709\u591a\u7bc7\u8bba\u6587\u88ab\u4e1a\u5185\u9876\u4f1a\u6536\u5f55\u5e76\u83b7\u5956\uff1b\u80fd\u591f\u957f\u671f\u6295\u5165\u5230\u4e00\u9879\u7814\u7a76\u8bfe\u9898\u4e2d\uff0c\u5e76\u6301\u7eed\u8ddf\u8fdb\u4e09\u3001\u56db\u5e74\uff1b\u9009\u5b9a\u4e00\u4e2a\u7814\u7a76\u9886\u57df\uff0c\u5c42\u5c42\u9012\u8fdb\uff0c\u5c55\u5f00\u4e86\u591a\u65b9\u9762\u7684\u7814\u7a76\uff1b\u70ed\u8877\u4e8e\u79d1\u5b66\u521b\u65b0\uff0c\u575a\u6301\u957f\u671f\u4e3b\u4e49\u7684\u7814\u7a76\u7406\u5ff5\u3002\u4ed6\u662f\u6765\u81ea\u97e9\u56fd\u7684\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7814\u7a76\u5458\u9ec4\u6636\u4e92\uff08Changho Hwang\uff09\u3002<\/p>\n\n\n\n

\n\n\n\n

\u201c\u5728\u6210\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u5b9e\u4e60\u751f\u4e4b\u524d\uff0c\u6211\u5bf9\u7814\u7a76\u9662\u7684\u4e86\u89e3\u5c31\u662f ResNet\uff08Residual Network\uff0c\u6b8b\u5dee\u7f51\u7edc\uff09\u7684\u8bba\u6587\u3002\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u9996\u6b21\u5f15\u5165\u4e86\u6b8b\u5dee\u5b66\u4e60\u7684\u601d\u60f3\uff0c\u8ba9 ResNet \u6210\u4e3a\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u6280\u672f\u53d1\u5c55\u7684\u4e00\u4e2a\u91cc\u7a0b\u7891\u3002\u201d\u9ec4\u6636\u4e92\uff08Changho Hwang\uff09\u8bf4\u3002<\/p>\n\n\n\n

\u201c\u524d\u6cbf\u7684\u6280\u672f\u7814\u7a76\uff0c\u9876\u5c16\u7684\u521b\u65b0\u4eba\u624d\u201d\uff0c\u662f\u9ec4\u6636\u4e92\u5bf9\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7b2c\u4e00\u5370\u8c61\u3002<\/p>\n\n\n\n

\u52a0\u5165\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u201c\u4ee5\u6b63\u786e\u7684\u65b9\u5f0f\u505a\u6b63\u786e\u7684\u4e8b\u201d<\/h2>\n\n\n\n

\u535a\u58eb\u5b66\u4e1a\u7684\u7b2c\u4e8c\u5e74\uff0c\u9ec4\u6636\u4e92\u5728\u97e9\u56fd\u79d1\u5b66\u6280\u672f\u9662\uff08KAIST\uff09\u5bfc\u5e08\u7684\u63a8\u8350\u4e0b\uff0c\u6210\u4e3a\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u4e00\u540d\u5b9e\u4e60\u751f\u3002\u57282018\u5e74\u5bd2\u5047\u4e0e2019\u5e74\u6691\u671f\u4e24\u4e2a\u9636\u6bb5\u7684\u5b9e\u4e60\u4e4b\u540e\uff0c\u9ec4\u6636\u4e92\u5bf9\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6709\u4e86\u5168\u65b0\u7684\u8ba4\u8bc6\uff0c\u5e76\u4e14\u786e\u5b9a\u4e86\u81ea\u5df1\u535a\u58eb\u6bd5\u4e1a\u540e\u7684\u804c\u4e1a\u76ee\u6807\u2014\u2014\u52a0\u5165\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u4ece\u4e8b\u66f4\u5177\u524d\u77bb\u6027\u7684\u6280\u672f\u7814\u7a76\u3002\u201c\u5f53\u65f6\u6709\u540c\u5b66\u548c\u540c\u4e8b\u7ed9\u6211\u4ecb\u7ecd\u8fc7\u5176\u4ed6\u5b9e\u9a8c\u5ba4\u548c\u516c\u53f8\uff0c\u4f46\u662f\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u5b9e\u4e60\u7ecf\u5386\u8ba9\u6211\u975e\u5e38\u660e\u786e\uff0c\u6211\u66f4\u559c\u6b22\u8fd9\u91cc\u7684\u5de5\u4f5c\u73af\u5883\u548c\u7814\u7a76\u6c1b\u56f4\uff0c\u5b83\u80fd\u591f\u8ba9\u6211\u4e13\u6ce8\u4e8e\u81ea\u5df1\u611f\u5174\u8da3\u7684\u7814\u7a76\u9886\u57df\u3002\u201d\u9ec4\u6636\u4e92\u8bf4\u3002<\/p>\n\n\n\n

\u9ec4\u6636\u4e92\u8868\u793a\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6700\u5438\u5f15\u4ed6\u7684\u4e00\u70b9\u662f\u59cb\u7ec8\u5728\u4ee5\u6b63\u786e\u7684\u65b9\u5f0f\u505a\u6b63\u786e\u7684\u4e8b\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0d\u4f1a\u968f\u6ce2\u9010\u6d41\u5730\u8ffd\u9010\u6280\u672f\u7684\u98ce\u53e3\uff0c\u800c\u662f\u6709\u7740\u72ec\u5230\u7684\u6218\u7565\u548c\u7814\u7a76\u65b9\u5411\uff0c\u5e76\u4e00\u76f4\u5c06\u76ee\u5149\u653e\u5728\u66f4\u5927\u7684\u84dd\u56fe\u4e4b\u4e0a\uff0c\u4e13\u6ce8\u4e8e\u63a2\u7d22\u524d\u6cbf\u7684\u6280\u672f\u7814\u7a76\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0c\u5171\u4e8b\u7684\u540c\u4e8b\u4ee5\u53ca\u7814\u7a76\u9662\u591a\u5143\u7684\u7814\u7a76\u65b9\u5411\uff0c\u4e5f\u662f\u9ec4\u6636\u4e92\u9009\u62e9\u52a0\u5165\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u91cd\u8981\u539f\u56e0\u3002\u7814\u7a76\u9662\u6709\u4e00\u7fa4\u53ef\u7231\u4e14\u6280\u672f\u5b9e\u529b\u6df1\u539a\u7684\u7814\u7a76\u5458\uff0c\u9ec4\u6636\u4e92\u5b9e\u4e60\u671f\u95f4\u7684\u5bfc\u5e08\u5e73\u6613\u8fd1\u4eba\uff0c\u5728\u7814\u7a76\u4e2d\u7ed9\u4e88\u4e86\u4ed6\u81ea\u7531\u7684\u7814\u7a76\u7a7a\u95f4\u548c\u6781\u5927\u7684\u5b66\u672f\u652f\u6301\uff0c\u5728\u5de5\u4f5c\u548c\u751f\u6d3b\u4e2d\uff0c\u540c\u4e8b\u4eec\u4e5f\u90fd\u70ed\u60c5\u76f8\u52a9\uff0c\u8fd9\u8ba9\u8eab\u5904\u5f02\u56fd\u4ed6\u4e61\u7684\u9ec4\u6636\u4e92\u500d\u611f\u6e29\u6696\u3002\u540c\u65f6\uff0c\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6240\u8fdb\u884c\u7684\u524d\u6cbf\u63a2\u7d22\u4e2d\uff0c\u4e0d\u4ec5\u6709\u4e0e\u9ec4\u6636\u4e92\u7684\u7535\u6c14\u5de5\u7a0b\u4e13\u4e1a\u9ad8\u5ea6\u5339\u914d\u7684\u7814\u7a76\u9886\u57df\u548c\u9879\u76ee\uff0c\u8fd8\u6709\u4e0d\u5c11\u6a2a\u8de8\u9886\u57df\/\u884c\u4e1a\u7684\u79d1\u7814\u65b9\u5411\uff0c\u8ba9\u7814\u7a76\u5458\u4eec\u6709\u673a\u4f1a\u62d3\u5c55\u7814\u7a76\u7684\u5e7f\u5ea6\u4e0e\u6df1\u5ea6\u3002\u56e0\u6b64\uff0c2022\u5e74\u535a\u58eb\u6bd5\u4e1a\u540e\uff0c\u9ec4\u6636\u4e92\u6beb\u4e0d\u72b9\u8c6b\u5730\u52a0\u5165\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u6210\u4e3a\u7f51\u7edc\u4e0e\u57fa\u7840\u8bbe\u65bd\u7ec4\uff08Networking Infrastructure Group\uff09\u7684\u4e00\u5458\u3002<\/p>\n\n\n\n

\"a
\u9ec4\u6636\u4e92\uff08\u4e2d\uff09\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7b2c\u4e8c\u6b21\u5b9e\u4e60\u540e\u4e0e\u7ec4\u5185\u5c0f\u4f19\u4f34\u5408\u5f71\u7559\u5ff5<\/figcaption><\/figure>\n\n\n\n

\u805a\u7126\u63d0\u5347AI\u7cfb\u7edf\u6027\u80fd\uff1a\u5c42\u5c42\u9012\u8fdb\u7814\u7a76\uff0c\u6301\u7eed\u6253\u78e8\u6210\u679c<\/h2>\n\n\n\n

\u5b9e\u4e60\u671f\u95f4\uff0c\u9ec4\u6636\u4e92\u6240\u5728\u56e2\u961f\u7684\u91cd\u70b9\u7814\u7a76\u8bfe\u9898\u662f\u4f18\u5316\u652f\u6301\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u8fd0\u884c\u7684 GPU \u7684\u6027\u80fd\uff0c\u5f53\u65f6\u9ec4\u6636\u4e92\u7684\u5de5\u4f5c\u6bd4\u8f83\u660e\u786e\uff0c\u4e3b\u8981\u662f\u63a2\u7d22\u5982\u4f55\u901a\u8fc7\u8f6f\u786c\u4ef6\u534f\u540c\u8bbe\u8ba1\u6765\u63d0\u9ad8\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u541e\u5410\u91cf\u548c\u5229\u7528\u7387\u3002\u7136\u800c\uff0c\u79d1\u5b66\u7814\u7a76\u662f\u4e00\u9879\u957f\u671f\u5de5\u4f5c\uff0c\u6709\u4e9b\u7814\u7a76\u5e76\u4e0d\u4f1a\u5728\u77ed\u671f\u5185\u5c31\u663e\u73b0\u6210\u679c\u3002\u4f5c\u4e3a\u4e00\u540d\u575a\u4fe1\u957f\u671f\u4e3b\u4e49\u7406\u5ff5\u7684\u79d1\u7814\u4eba\u5458\uff0c\u9ec4\u6636\u4e92\u5728\u8fd9\u9879\u7814\u7a76\u4e2d\u5e76\u6ca1\u6709\u628a\u81ea\u5df1\u5f53\u4f5c\u5306\u5306\u8fc7\u5ba2\uff0c\u76f8\u53cd\uff0c\u5728\u5b9e\u4e60\u7ed3\u675f\u56de\u5230\u5b66\u6821\u7ee7\u7eed\u653b\u8bfb\u535a\u58eb\u5b66\u4f4d\u7684\u4e24\u5e74\u91cc\uff0c\u4ed6\u4ecd\u7136\u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u56e2\u961f\u4fdd\u6301\u5408\u4f5c\uff0c\u6301\u7eed\u8ddf\u8fdb\u8fd9\u4e00\u8bfe\u9898\u3002\u6700\u7ec8\uff0c\u4ed6\u548c\u7814\u7a76\u9662\u56e2\u961f\u7684\u7814\u7a76\u6210\u679c\u83b7\u5f97\u4e862022\u5e74 MLArchSys \u5927\u4f1a\u7684\u6700\u4f73\u8bba\u6587\u5956\u3002<\/p>\n\n\n\n

\u8bba\u6587\u6807\u9898\uff1aTowards GPU-driven Code Execution for Distributed Deep Learning<\/em><\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/chhwang.github.io\/pubs\/mlarchsys22_hwang.pdf<\/em><\/p>\n\n\n\n

\u968f\u7740\u5927\u6a21\u578b\u7684\u53d1\u5c55\uff0cGPU \u6108\u53d1\u6210\u4e3a\u8bad\u7ec3\u548c\u90e8\u7f72\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u5173\u952e\u786c\u4ef6\uff0cGPU \u7684\u6027\u80fd\u548c\u5229\u7528\u6548\u7387\u76f4\u63a5\u5f71\u54cd\u7740\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\u3002\u56e0\u6b64\uff0c\u5728\u6210\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b63\u5f0f\u7684\u7814\u7a76\u5458\u4e4b\u540e\uff0c\u9ec4\u6636\u4e92\u4f9d\u7136\u81f4\u529b\u4e8e\u8fd9\u4e00\u65b9\u5411\u7684\u7814\u7a76\uff0c\u800c\u4ed6\u7684\u89d2\u8272\u4ece\u66fe\u7ecf\u7684\u9879\u76ee\u53c2\u4e0e\u8005\uff0c\u8f6c\u53d8\u4e3a\u4e86\u9879\u76ee\u7684\u4e3b\u5bfc\u8005\u3002<\/p>\n\n\n\n

\u9ec4\u6636\u4e92\u8ba4\u4e3a\uff0c\u5982\u4eca\u6700\u5148\u8fdb\u7684\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u9700\u8981\u5927\u91cf\u5e76\u884c\u7684 GPU \u63d0\u4f9b\u5145\u8db3\u7684\u7b97\u529b\uff0c\u4f46 GPU \u548c CPU \u4e4b\u95f4\u7684\u901a\u4fe1\u6548\u7387\u5374\u5236\u7ea6\u7740\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u6027\u80fd\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u5f53\u524d\u4e3b\u8981\u4f9d\u9760 GPU \u9a71\u52a8\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u901a\u4fe1\u6a21\u5f0f\u4e0b\uff0cCPU \u5374\u626e\u6f14\u7740\u603b\u6307\u6325\u7684\u89d2\u8272\uff0cCPU \u8d1f\u8d23\u7ed9\u591a\u4e2a GPU \u5e03\u7f6e\u4efb\u52a1\uff0c\u800c CPU \u4e0e GPU \u4e4b\u95f4\u7684\u6d88\u606f\u4f20\u9012\u5b58\u5728\u53ef\u89c2\u7684\u5ef6\u8fdf\uff0c\u8fd9\u5c31\u5bfc\u81f4\u4e86\u4efb\u52a1\u6267\u884c\u6548\u7387\u7684\u4f4e\u4e0b\uff0c\u9020\u6210\u4e86 GPU \u8d44\u6e90\u7684\u6d6a\u8d39\u3002<\/p>\n\n\n\n

\u9ec4\u6636\u4e92\u7684\u7814\u7a76\u76ee\u6807\u4e0e\u601d\u8def\u662f\u5e0c\u671b GPU \u53ef\u4ee5\u81ea\u5df1\u6307\u6325\u81ea\u5df1\uff0c\u4ece\u800c\u63d0\u5347\u901a\u4fe1\u6548\u7387\u3002\u4e3a\u6b64\uff0c\u4ed6\u548c\u7ec4\u91cc\u7684\u540c\u4e8b\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u7531 GPU \u9a71\u52a8\u7684\u4ee3\u7801\u6267\u884c\u7cfb\u7edf\uff0c\u5e76\u5f00\u53d1\u4e86\u4e00\u79cd\u80fd\u591f\u88ab GPU \u76f4\u63a5\u9a71\u52a8\u7684 DMA \u5f15\u64ce\uff0c\u8ba9 GPU \u80fd\u591f\u81ea\u5df1\u89e3\u51b3\u539f\u672c\u9700\u8981 CPU \u6307\u6325\u7684\u901a\u4fe1\u95ee\u9898\uff0c\u964d\u4f4e\u4e86\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u901a\u4fe1\u5ef6\u8fdf\uff0c\u63d0\u9ad8\u4e86 GPU \u8ba1\u7b97\u8d44\u6e90\u7684\u5229\u7528\u7387\u3002\u8fd9\u79cd\u65b9\u6cd5\u91ca\u653e\u4e86\u4e4b\u524d\u901a\u4fe1\u6a21\u5f0f\u4e0b\u88ab\u5360\u7528\u7684 CPU \u8d44\u6e90\uff0c\u8ba9 CPU \u4e13\u6ce8\u4e8e\u81ea\u5df1\u7684\u5de5\u4f5c\uff0c\u4e5f\u8ba9 GPU \u5b9e\u73b0\u81ea\u4e3b\u8c03\u5ea6\uff0c\u505a\u5b83\u6700\u64c5\u957f\u7684\u5de5\u4f5c\u2014\u2014\u7ed9\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u63d0\u4f9b\u66f4\u9ad8\u7684\u7b97\u529b\u6027\u80fd\u3002\u8fd9\u9879\u7814\u7a76\u5de5\u4f5c\u9996\u6b21\u8bc1\u660e\u4e86\u57fa\u4e8e\u5206\u5e03\u5f0f GPU \u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u53ef\u4ee5\u7531 GPU \u81ea\u5df1\u5b8c\u6210\u4efb\u52a1\u8c03\u5ea6\uff0c\u76f8\u5173\u8bba\u6587\u5df2\u88ab2023\u5e74 NSDI \u5927\u4f1a\u63a5\u6536\u3002<\/p>\n\n\n\n

\u8bba\u6587\u6807\u9898\uff1aARK: GPU-driven Code Execution for Distributed Deep Learning<\/em><\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/www.usenix.org\/system\/files\/nsdi23-hwang.pdf<\/em><\/p>\n\n\n\n

\u201c\u7cfb\u7edf\u6027\u80fd\u4f18\u5316\u662f\u4e00\u4e2a\u6c38\u6052\u7684\u8bdd\u9898\u3002\u5728\u8fc7\u53bb\u7684\u5341\u51e0\u5e74\u4e2d\uff0c\u6211\u4eec\u89c1\u8bc1\u4e86\u4eba\u5de5\u667a\u80fd\u7684\u5feb\u901f\u53d1\u5c55\uff0c\u5176\u4e2d\u4e00\u4e2a\u4e3b\u8981\u7684\u9a71\u52a8\u56e0\u7d20\u5c31\u662f\u4e0d\u65ad\u589e\u5f3a\u7684\u7b97\u529b\u652f\u6301\u3002\u5145\u8db3\u7684\u7b97\u529b\u8ba9\u7cfb\u7edf\u6027\u80fd\u6301\u7eed\u63d0\u5347\uff0c\u4e5f\u4f7f\u5f97\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u53d8\u5f97\u8d8a\u6765\u8d8a\u5927\uff0c\u529f\u80fd\u8d8a\u6765\u8d8a\u5f3a\u3002\u5f53\u524d\uff0c\u63d0\u5347\u7cfb\u7edf\u6027\u80fd\u7684\u7814\u7a76\u65b9\u5411\u4e3b\u8981\u6709\u4e24\u4e2a\u5207\u5165\u70b9\uff0c\u4e00\u662f\u63d0\u5347 GPU \u7b49\u786c\u4ef6\u7684\u6027\u80fd\uff0c\u4e8c\u662f\u63d0\u51fa\u65b0\u7684\u4eba\u5de5\u667a\u80fd\u7b97\u6cd5\uff0c\u4f46\u8fd9\u4e24\u79cd\u65b9\u6cd5\u90fd\u76f8\u5f53\u56f0\u96be\uff0c\u5e76\u4e14\u786c\u4ef6\u7684\u8bbe\u8ba1\u548c\u5236\u9020\u6210\u672c\u9ad8\u6602\u3002\u201d\u9ec4\u6636\u4e92\u8bf4\u9053\u3002<\/p>\n\n\n\n

\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\uff0c\u9ec4\u6636\u4e92\u548c\u540c\u4e8b\u4eec\u63d0\u51fa\u4e86\u786c\u4ef6\u4e0e\u7b97\u6cd5\u534f\u540c\u8bbe\u8ba1\u7684\u65b9\u6cd5\uff0c\u8fd9\u6216\u5c06\u662f\u53e6\u4e00\u79cd\u63d0\u5347\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u6027\u80fd\u7684\u6709\u6548\u89e3\u51b3\u65b9\u6848\u3002\u56e0\u6b64\uff0c\u5728\u8bc1\u660e\u4e86 GPU \u53ef\u4ee5\u81ea\u4e3b\u8c03\u5ea6\uff0c\u5b9e\u73b0\u6027\u80fd\u63d0\u5347\u540e\uff0c\u9ec4\u6636\u4e92\u5c06\u7ee7\u7eed\u63a2\u7d22 GPU \u7684\u8c03\u5ea6\u7b97\u6cd5\uff0c\u907f\u514d\u8c03\u5ea6\u51b2\u7a81\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347 GPU \u4e4b\u95f4\u7684\u901a\u4fe1\u6548\u7387\u3002\u9ec4\u6636\u4e92\u8868\u793a\uff0c\u201c\u5e0c\u671b\u672a\u6765 GPU \u4e0d\u518d\u9700\u8981\u989d\u5916\u7684 DMA \u5f15\u64ce\u5c31\u80fd\u5b9e\u73b0\u81ea\u4e3b\u8c03\u5ea6\uff0c\u4ece\u800c\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u6027\u80fd\u518d\u4e0a\u4e00\u4e2a\u53f0\u9636\u3002\u201d<\/p>\n\n\n\n

\u201c\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u6211\u53ef\u4ee5\u81ea\u7531\u9009\u62e9\u7814\u7a76\u65b9\u5411\u201d<\/h2>\n\n\n\n

\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e00\u76f4\u4ee5\u6765\u6240\u62e5\u6709\u7684\u5f00\u653e\u3001\u5305\u5bb9\u3001\u591a\u5143\u7684\u7814\u7a76\u6587\u5316\uff0c\u4e5f\u5bf9\u9ec4\u6636\u4e92\u6709\u7740\u5de8\u5927\u7684\u5438\u5f15\u529b\u3002\u5728\u7814\u7a76\u9662\u5de5\u4f5c\u4e00\u5e74\u6709\u4f59\u7684\u9ec4\u6636\u4e92\u5bf9\u8fd9\u91cc\u6709\u4e86\u66f4\u6df1\u523b\u7684\u8ba4\u8bc6\uff0c\u201c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u66f4\u50cf\u662f\u4e00\u4e2a\u5b9e\u9a8c\u5ba4\uff0c\u4e00\u4e2a\u771f\u6b63\u7684\u7814\u7a76\u673a\u6784\uff0c\u5728\u8fd9\u91cc\uff0c\u6240\u6709\u4eba\u90fd\u662f\u5e73\u7b49\u7684\uff0c\u6240\u505a\u7684\u5de5\u4f5c\u90fd\u662f\u900f\u660e\u7684\uff0c\u5927\u5bb6\u4e86\u89e3\u5f7c\u6b64\u7684\u60f3\u6cd5\uff0c\u601d\u60f3\u4e0a\u4e5f\u80fd\u591f\u4fdd\u6301\u540c\u6b65\u3002\u5728\u7814\u7a76\u9662\uff0c\u6211\u4eec\u6709\u66f4\u5927\u7684\u81ea\u7531\u5ea6\u6765\u9009\u62e9\u81ea\u5df1\u7684\u7814\u7a76\u65b9\u5411\u3002\u201d<\/p>\n\n\n\n

\u9664\u4e86\u5728\u5185\u90e8\u8425\u9020\u81ea\u7531\u7684\u5b66\u672f\u6c1b\u56f4\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u8fd8\u5728\u5b66\u672f\u4ea4\u6d41\u548c\u4eba\u624d\u57f9\u517b\u65b9\u9762\uff0c\u4e0e\u5305\u62ec\u97e9\u56fd\u5728\u5185\u7684\u5168\u7403\u5b66\u672f\u754c\u6301\u7eed\u4fdd\u6301\u7740\u7d27\u5bc6\u7684\u5408\u4f5c\u3002\u4f8b\u5982\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u8054\u5408\u6e05\u534e\u5927\u5b66\u3001\u5317\u4eac\u5927\u5b66\u3001\u65b0\u52a0\u5761\u56fd\u7acb\u5927\u5b66\u3001\u9996\u5c14\u56fd\u7acb\u5927\u5b66\u7b49\u591a\u6240\u4e9a\u6d32\u5730\u533a\u9ad8\u6821\u6210\u7acb\u4e86 OpenNetLab \u5f00\u653e\u7f51\u7edc\u5e73\u53f0\u8054\u76df\uff0c\u4ee5\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u5728\u7f51\u7edc\u7814\u7a76\u4e2d\u7684\u5e94\u7528\u4e0e\u53d1\u5c55\uff0c\u9ec4\u6636\u4e92\u5728 KAIST \u6c42\u5b66\u65f6\u7684\u5bfc\u5e08\u5c31\u53c2\u4e0e\u5176\u4e2d\u3002\u518d\u5982\uff0c\u6301\u7eed\u4e86\u5341\u591a\u5e74\u7684\u9762\u5411\u97e9\u56fd\u9ad8\u6821\u4eba\u624d\u57f9\u517b\u548c\u5b66\u672f\u7814\u7a76\u7684 MSIT \u9879\u76ee\uff0c\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0e\u97e9\u56fd\u5b66\u672f\u754c\u642d\u5efa\u4e86\u5b66\u672f\u4ea4\u6d41\u7684\u6865\u6881\uff0c\u901a\u8fc7\u5408\u4f5c\u9879\u76ee\uff0c\u5b66\u8005\u4eec\u5f00\u5c55\u4e86\u6df1\u5165\u7684\u79d1\u7814\u5408\u4f5c\uff0c\u5e76\u4e30\u5bcc\u4e86\u5168\u7403\u8ba1\u7b97\u673a\u9886\u57df\u7684\u4eba\u624d\u50a8\u5907\u3002\u9ec4\u6636\u4e92\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5b9e\u4e60\u540e\u4e5f\u53c2\u4e0e\u4e86\u4e00\u4e2a\u5b66\u672f\u5408\u4f5c\u9879\u76ee\uff0c\u76f8\u5173\u8bba\u6587\u8fd8\u83b7\u5f97\u4e862021\u5e74 APSys \u5927\u4f1a\u7684\u6700\u4f73\u8bba\u6587\u5956\u3002<\/p>\n\n\n\n

\u8bba\u6587\u6807\u9898\uff1aAccelerating GNN Training with Locality-Aware Partial Execution<\/em><\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/dl.acm.org\/doi\/10.1145\/3476886.3477515<\/em><\/p>\n\n\n\n

\"graphical<\/figure>\n\n\n\n

\u4f5c\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e43\u81f3\u6574\u4e2a\u8ba1\u7b97\u673a\u5b66\u672f\u751f\u6001\u4f53\u7cfb\u7684\u4e00\u90e8\u5206\uff0c\u8fd9\u4e9b\u591a\u6837\u7684\u4ea4\u6d41\u4e0e\u5408\u4f5c\u9879\u76ee\u4e0d\u4ec5\u4ea7\u51fa\u4e86\u4f17\u591a\u524d\u6cbf\u7684\u79d1\u7814\u6210\u679c\uff0c\u4e5f\u6210\u4e3a\u4e86\u4f17\u591a\u5b66\u8005\u548c\u5b66\u751f\u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7ed3\u7f18\u7684\u8d77\u70b9\u3002\u4ee5\u97e9\u56fd\u4e3a\u4f8b\uff0c\u622a\u81f3\u76ee\u524d\u5df2\u6709\u8d85\u8fc7150\u591a\u540d\u6765\u81ea\u97e9\u56fd\u7684\u8de8\u5b66\u79d1\u4eba\u624d\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u8fdb\u884c\u8fc7\u5b9e\u4e60\uff0c\u4e5f\u5438\u5f15\u4e86\u50cf\u9ec4\u6636\u4e92\u8fd9\u6837\u7684\u4f18\u79c0\u4eba\u624d\uff0c\u6210\u4e3a\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u6b63\u5f0f\u5458\u5de5\u3002<\/p>\n\n\n\n

\u575a\u6301\u957f\u671f\u4e3b\u4e49\u7814\u7a76\u7684\u5fc3\u5f97<\/h2>\n\n\n\n

\u79d1\u5b66\u7814\u7a76\u4e4b\u8def\u9053\u963b\u4e14\u957f\uff0c\u575a\u6301\u957f\u671f\u4e3b\u4e49\u7814\u7a76\u3001\u884c\u800c\u4e0d\u8f8d\u5e76\u975e\u4e00\u4ef6\u6613\u4e8b\u3002\u9664\u4e86\u672c\u8eab\u6267\u8457\u7684\u6027\u683c\u4e4b\u5916\uff0c\u9ec4\u6636\u4e92\u4e5f\u6709\u81ea\u5df1\u7684\u65b9\u6cd5\u548c\u5fc3\u5f97\u3002<\/p>\n\n\n\n

\u9ec4\u6636\u4e92\u8ba4\u4e3a\u4ece\u4e8b\u79d1\u7814\u5de5\u4f5c\uff0c\u9996\u5148\u8981\u5bf9\u7814\u7a76\u4e8b\u4e1a\u4fdd\u6301\u9ad8\u5ea6\u7684\u70ed\u60c5\uff0c\u6bd4\u5982\u4ed6\u81ea\u8eab\u5c31\u5341\u5206\u4eab\u53d7\u79d1\u5b66\u7814\u7a76\u4e2d\u53d1\u73b0\u95ee\u9898\u3001\u89e3\u51b3\u95ee\u9898\u7684\u6574\u4e2a\u8fc7\u7a0b\u3002\u201c\u6709\u4e9b\u5de5\u4f5c\u7684\u76ee\u6807\u662f\u627e\u5230\u907f\u5f00\u95ee\u9898\u7684\u6700\u4f73\u65b9\u5f0f\uff1b\u800c\u79d1\u5b66\u7814\u7a76\u7684\u76ee\u6807\u662f\u627e\u5230\u95ee\u9898\u3001\u76f4\u9762\u95ee\u9898\u3001\u89e3\u51b3\u95ee\u9898\u3002\u6211\u66f4\u4eab\u53d7\u4ece\u53d1\u73b0\u95ee\u9898\u5230\u89e3\u51b3\u95ee\u9898\u7684\u79d1\u7814\u8fc7\u7a0b\u3002\u201d\u9ec4\u6636\u4e92\u8bf4\u3002<\/p>\n\n\n\n

\"\u9ec4\u6636\u4e92\"<\/figure>\n\n\n\n

\u800c\u5728\u957f\u671f\u7814\u7a76\u4e2d\u96be\u514d\u4f1a\u9047\u5230\u963b\u788d\uff0c\u6216\u8005\u7ed3\u679c\u8fbe\u4e0d\u5230\u9884\u671f\uff0c\u4f8b\u5982\u9ec4\u6636\u4e92\u7684\u7814\u7a76\u8bba\u6587\u4e5f\u66fe\u88ab\u6240\u6295\u5927\u4f1a\u4e00\u6b21\u6b21\u62d2\u4e4b\u95e8\u5916\u3002\u9762\u5bf9\u8fd9\u79cd\u60c5\u51b5\uff0c\u201c\u4e0d\u8981\u6c14\u9981\u6216\u662f\u6028\u5929\u5c24\u4eba\uff0c\u800c\u662f\u8981\u53cd\u601d\u81ea\u5df1\uff0c\u590d\u76d8\u5df2\u6709\u5de5\u4f5c\uff0c\u627e\u51fa\u5176\u4e2d\u7684\u95ee\u9898\uff0c\u518d\u6295\u5165\u65b0\u7684\u7814\u7a76\u3002\u201d\u9ec4\u6636\u4e92\u8ba4\u4e3a\uff0c\u201c\u8fd9\u662f\u4e00\u4e2a\u8bf4\u670d\u81ea\u5df1\u7684\u8fc7\u7a0b\uff0c\u8981\u8ba9\u81ea\u5df1\u770b\u5230\u7814\u7a76\u7684\u4ef7\u503c\u3002\u201d<\/p>\n\n\n\n

\u5f53\u9762\u5bf9\u7814\u7a76\u56f0\u5883\u65f6\uff0c\u9ec4\u6636\u4e92\u8868\u793a\uff0c\u4e0d\u80fd\u753b\u5730\u4e3a\u7262\uff0c\u5c06\u81ea\u5df1\u56f0\u4e8e\u5f53\u524d\u7684\u95ee\u9898\u4e2d\uff0c\u800c\u662f\u8981\u5b66\u4f1a\u653e\u677e\u81ea\u5df1\uff0c\u4f8b\u5982\u4ed6\u4f1a\u5f39\u5f39\u94a2\u7434\uff0c\u6216\u8005\u4e0e\u4ed6\u4eba\u8c08\u5fc3\u4ea4\u6d41\uff0c\u4ee5\u6b64\u6765\u6446\u8131\u684e\u688f\uff0c\u8f6c\u6362\u601d\u8def\u4e5f\u8bb8\u95ee\u9898\u5c31\u80fd\u8fce\u5203\u800c\u89e3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u7f16\u8005\u6309\uff1a\u4ed6\uff0c\u521a\u5165\u804c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e00\u5e74\uff0c\u5374\u6709\u7740\u4e30\u5bcc\u7684\u5b66\u672f\u5408\u4f5c\u7ecf\u9a8c\uff1b\u521a\u521a\u535a\u58eb\u6bd5\u4e1a\uff0c\u5c31\u6709\u591a\u7bc7\u8bba\u6587\u88ab\u4e1a\u5185\u9876\u4f1a\u6536\u5f55\u5e76\u83b7\u5956\uff1b\u80fd\u591f\u957f\u671f\u6295\u5165\u5230\u4e00\u9879\u7814\u7a76\u8bfe\u9898\u4e2d\uff0c\u5e76\u6301\u7eed\u8ddf\u8fdb\u4e09\u3001\u56db\u5e74\uff1b\u9009\u5b9a\u4e00\u4e2a\u7814\u7a76\u9886\u57df\uff0c\u5c42\u5c42\u9012\u8fdb\uff0c\u5c55\u5f00\u4e86\u591a\u65b9\u9762\u7684\u7814\u7a76\uff1b\u70ed\u8877\u4e8e\u79d1\u5b66\u521b\u65b0\uff0c\u575a\u6301\u957f\u671f\u4e3b\u4e49\u7684\u7814\u7a76\u7406\u5ff5\u3002\u4ed6\u662f\u6765\u81ea\u97e9\u56fd\u7684\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7814\u7a76\u5458\u9ec4\u6636\u4e92\uff08Changho Hwang\uff09\u3002 \u201c\u5728\u6210\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u5b9e\u4e60\u751f\u4e4b\u524d\uff0c\u6211\u5bf9\u7814\u7a76\u9662\u7684\u4e86\u89e3\u5c31\u662f ResNet\uff08Residual Network\uff0c\u6b8b\u5dee\u7f51\u7edc\uff09\u7684\u8bba\u6587\u3002\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u9996\u6b21\u5f15\u5165\u4e86\u6b8b\u5dee\u5b66\u4e60\u7684\u601d\u60f3\uff0c\u8ba9 ResNet \u6210\u4e3a\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u6280\u672f\u53d1\u5c55\u7684\u4e00\u4e2a\u91cc\u7a0b\u7891\u3002\u201d\u9ec4\u6636\u4e92\uff08Changho Hwang\uff09\u8bf4\u3002 \u201c\u524d\u6cbf\u7684\u6280\u672f\u7814\u7a76\uff0c\u9876\u5c16\u7684\u521b\u65b0\u4eba\u624d\u201d\uff0c\u662f\u9ec4\u6636\u4e92\u5bf9\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7b2c\u4e00\u5370\u8c61\u3002 \u535a\u58eb\u5b66\u4e1a\u7684\u7b2c\u4e8c\u5e74\uff0c\u9ec4\u6636\u4e92\u5728\u97e9\u56fd\u79d1\u5b66\u6280\u672f\u9662\uff08KAIST\uff09\u5bfc\u5e08\u7684\u63a8\u8350\u4e0b\uff0c\u6210\u4e3a\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u4e00\u540d\u5b9e\u4e60\u751f\u3002\u57282018\u5e74\u5bd2\u5047\u4e0e2019\u5e74\u6691\u671f\u4e24\u4e2a\u9636\u6bb5\u7684\u5b9e\u4e60\u4e4b\u540e\uff0c\u9ec4\u6636\u4e92\u5bf9\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6709\u4e86\u5168\u65b0\u7684\u8ba4\u8bc6\uff0c\u5e76\u4e14\u786e\u5b9a\u4e86\u81ea\u5df1\u535a\u58eb\u6bd5\u4e1a\u540e\u7684\u804c\u4e1a\u76ee\u6807\u2014\u2014\u52a0\u5165\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u4ece\u4e8b\u66f4\u5177\u524d\u77bb\u6027\u7684\u6280\u672f\u7814\u7a76\u3002\u201c\u5f53\u65f6\u6709\u540c\u5b66\u548c\u540c\u4e8b\u7ed9\u6211\u4ecb\u7ecd\u8fc7\u5176\u4ed6\u5b9e\u9a8c\u5ba4\u548c\u516c\u53f8\uff0c\u4f46\u662f\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u5b9e\u4e60\u7ecf\u5386\u8ba9\u6211\u975e\u5e38\u660e\u786e\uff0c\u6211\u66f4\u559c\u6b22\u8fd9\u91cc\u7684\u5de5\u4f5c\u73af\u5883\u548c\u7814\u7a76\u6c1b\u56f4\uff0c\u5b83\u80fd\u591f\u8ba9\u6211\u4e13\u6ce8\u4e8e\u81ea\u5df1\u611f\u5174\u8da3\u7684\u7814\u7a76\u9886\u57df\u3002\u201d\u9ec4\u6636\u4e92\u8bf4\u3002 \u9ec4\u6636\u4e92\u8868\u793a\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6700\u5438\u5f15\u4ed6\u7684\u4e00\u70b9\u662f\u59cb\u7ec8\u5728\u4ee5\u6b63\u786e\u7684\u65b9\u5f0f\u505a\u6b63\u786e\u7684\u4e8b\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0d\u4f1a\u968f\u6ce2\u9010\u6d41\u5730\u8ffd\u9010\u6280\u672f\u7684\u98ce\u53e3\uff0c\u800c\u662f\u6709\u7740\u72ec\u5230\u7684\u6218\u7565\u548c\u7814\u7a76\u65b9\u5411\uff0c\u5e76\u4e00\u76f4\u5c06\u76ee\u5149\u653e\u5728\u66f4\u5927\u7684\u84dd\u56fe\u4e4b\u4e0a\uff0c\u4e13\u6ce8\u4e8e\u63a2\u7d22\u524d\u6cbf\u7684\u6280\u672f\u7814\u7a76\u3002 \u6b64\u5916\uff0c\u5171\u4e8b\u7684\u540c\u4e8b\u4ee5\u53ca\u7814\u7a76\u9662\u591a\u5143\u7684\u7814\u7a76\u65b9\u5411\uff0c\u4e5f\u662f\u9ec4\u6636\u4e92\u9009\u62e9\u52a0\u5165\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u91cd\u8981\u539f\u56e0\u3002\u7814\u7a76\u9662\u6709\u4e00\u7fa4\u53ef\u7231\u4e14\u6280\u672f\u5b9e\u529b\u6df1\u539a\u7684\u7814\u7a76\u5458\uff0c\u9ec4\u6636\u4e92\u5b9e\u4e60\u671f\u95f4\u7684\u5bfc\u5e08\u5e73\u6613\u8fd1\u4eba\uff0c\u5728\u7814\u7a76\u4e2d\u7ed9\u4e88\u4e86\u4ed6\u81ea\u7531\u7684\u7814\u7a76\u7a7a\u95f4\u548c\u6781\u5927\u7684\u5b66\u672f\u652f\u6301\uff0c\u5728\u5de5\u4f5c\u548c\u751f\u6d3b\u4e2d\uff0c\u540c\u4e8b\u4eec\u4e5f\u90fd\u70ed\u60c5\u76f8\u52a9\uff0c\u8fd9\u8ba9\u8eab\u5904\u5f02\u56fd\u4ed6\u4e61\u7684\u9ec4\u6636\u4e92\u500d\u611f\u6e29\u6696\u3002\u540c\u65f6\uff0c\u5728\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6240\u8fdb\u884c\u7684\u524d\u6cbf\u63a2\u7d22\u4e2d\uff0c\u4e0d\u4ec5\u6709\u4e0e\u9ec4\u6636\u4e92\u7684\u7535\u6c14\u5de5\u7a0b\u4e13\u4e1a\u9ad8\u5ea6\u5339\u914d\u7684\u7814\u7a76\u9886\u57df\u548c\u9879\u76ee\uff0c\u8fd8\u6709\u4e0d\u5c11\u6a2a\u8de8\u9886\u57df\/\u884c\u4e1a\u7684\u79d1\u7814\u65b9\u5411\uff0c\u8ba9\u7814\u7a76\u5458\u4eec\u6709\u673a\u4f1a\u62d3\u5c55\u7814\u7a76\u7684\u5e7f\u5ea6\u4e0e\u6df1\u5ea6\u3002\u56e0\u6b64\uff0c2022\u5e74\u535a\u58eb\u6bd5\u4e1a\u540e\uff0c\u9ec4\u6636\u4e92\u6beb\u4e0d\u72b9\u8c6b\u5730\u52a0\u5165\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\uff0c\u6210\u4e3a\u7f51\u7edc\u4e0e\u57fa\u7840\u8bbe\u65bd\u7ec4\uff08Networking Infrastructure Group\uff09\u7684\u4e00\u5458\u3002 \u5b9e\u4e60\u671f\u95f4\uff0c\u9ec4\u6636\u4e92\u6240\u5728\u56e2\u961f\u7684\u91cd\u70b9\u7814\u7a76\u8bfe\u9898\u662f\u4f18\u5316\u652f\u6301\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u8fd0\u884c\u7684 GPU \u7684\u6027\u80fd\uff0c\u5f53\u65f6\u9ec4\u6636\u4e92\u7684\u5de5\u4f5c\u6bd4\u8f83\u660e\u786e\uff0c\u4e3b\u8981\u662f\u63a2\u7d22\u5982\u4f55\u901a\u8fc7\u8f6f\u786c\u4ef6\u534f\u540c\u8bbe\u8ba1\u6765\u63d0\u9ad8\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u541e\u5410\u91cf\u548c\u5229\u7528\u7387\u3002\u7136\u800c\uff0c\u79d1\u5b66\u7814\u7a76\u662f\u4e00\u9879\u957f\u671f\u5de5\u4f5c\uff0c\u6709\u4e9b\u7814\u7a76\u5e76\u4e0d\u4f1a\u5728\u77ed\u671f\u5185\u5c31\u663e\u73b0\u6210\u679c\u3002\u4f5c\u4e3a\u4e00\u540d\u575a\u4fe1\u957f\u671f\u4e3b\u4e49\u7406\u5ff5\u7684\u79d1\u7814\u4eba\u5458\uff0c\u9ec4\u6636\u4e92\u5728\u8fd9\u9879\u7814\u7a76\u4e2d\u5e76\u6ca1\u6709\u628a\u81ea\u5df1\u5f53\u4f5c\u5306\u5306\u8fc7\u5ba2\uff0c\u76f8\u53cd\uff0c\u5728\u5b9e\u4e60\u7ed3\u675f\u56de\u5230\u5b66\u6821\u7ee7\u7eed\u653b\u8bfb\u535a\u58eb\u5b66\u4f4d\u7684\u4e24\u5e74\u91cc\uff0c\u4ed6\u4ecd\u7136\u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u56e2\u961f\u4fdd\u6301\u5408\u4f5c\uff0c\u6301\u7eed\u8ddf\u8fdb\u8fd9\u4e00\u8bfe\u9898\u3002\u6700\u7ec8\uff0c\u4ed6\u548c\u7814\u7a76\u9662\u56e2\u961f\u7684\u7814\u7a76\u6210\u679c\u83b7\u5f97\u4e862022\u5e74 MLArchSys \u5927\u4f1a\u7684\u6700\u4f73\u8bba\u6587\u5956\u3002 \u8bba\u6587\u6807\u9898\uff1aTowards GPU-driven Code Execution for Distributed Deep Learning \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/chhwang.github.io\/pubs\/mlarchsys22_hwang.pdf \u968f\u7740\u5927\u6a21\u578b\u7684\u53d1\u5c55\uff0cGPU \u6108\u53d1\u6210\u4e3a\u8bad\u7ec3\u548c\u90e8\u7f72\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u5173\u952e\u786c\u4ef6\uff0cGPU \u7684\u6027\u80fd\u548c\u5229\u7528\u6548\u7387\u76f4\u63a5\u5f71\u54cd\u7740\u4eba\u5de5\u667a\u80fd\u7684\u53d1\u5c55\u3002\u56e0\u6b64\uff0c\u5728\u6210\u4e3a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b63\u5f0f\u7684\u7814\u7a76\u5458\u4e4b\u540e\uff0c\u9ec4\u6636\u4e92\u4f9d\u7136\u81f4\u529b\u4e8e\u8fd9\u4e00\u65b9\u5411\u7684\u7814\u7a76\uff0c\u800c\u4ed6\u7684\u89d2\u8272\u4ece\u66fe\u7ecf\u7684\u9879\u76ee\u53c2\u4e0e\u8005\uff0c\u8f6c\u53d8\u4e3a\u4e86\u9879\u76ee\u7684\u4e3b\u5bfc\u8005\u3002 \u9ec4\u6636\u4e92\u8ba4\u4e3a\uff0c\u5982\u4eca\u6700\u5148\u8fdb\u7684\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u9700\u8981\u5927\u91cf\u5e76\u884c\u7684 GPU \u63d0\u4f9b\u5145\u8db3\u7684\u7b97\u529b\uff0c\u4f46 GPU \u548c CPU \u4e4b\u95f4\u7684\u901a\u4fe1\u6548\u7387\u5374\u5236\u7ea6\u7740\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u6027\u80fd\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u5f53\u524d\u4e3b\u8981\u4f9d\u9760 GPU \u9a71\u52a8\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u901a\u4fe1\u6a21\u5f0f\u4e0b\uff0cCPU \u5374\u626e\u6f14\u7740\u603b\u6307\u6325\u7684\u89d2\u8272\uff0cCPU \u8d1f\u8d23\u7ed9\u591a\u4e2a GPU \u5e03\u7f6e\u4efb\u52a1\uff0c\u800c CPU \u4e0e GPU \u4e4b\u95f4\u7684\u6d88\u606f\u4f20\u9012\u5b58\u5728\u53ef\u89c2\u7684\u5ef6\u8fdf\uff0c\u8fd9\u5c31\u5bfc\u81f4\u4e86\u4efb\u52a1\u6267\u884c\u6548\u7387\u7684\u4f4e\u4e0b\uff0c\u9020\u6210\u4e86 GPU \u8d44\u6e90\u7684\u6d6a\u8d39\u3002 \u9ec4\u6636\u4e92\u7684\u7814\u7a76\u76ee\u6807\u4e0e\u601d\u8def\u662f\u5e0c\u671b GPU \u53ef\u4ee5\u81ea\u5df1\u6307\u6325\u81ea\u5df1\uff0c\u4ece\u800c\u63d0\u5347\u901a\u4fe1\u6548\u7387\u3002\u4e3a\u6b64\uff0c\u4ed6\u548c\u7ec4\u91cc\u7684\u540c\u4e8b\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u7531 GPU \u9a71\u52a8\u7684\u4ee3\u7801\u6267\u884c\u7cfb\u7edf\uff0c\u5e76\u5f00\u53d1\u4e86\u4e00\u79cd\u80fd\u591f\u88ab […]<\/p>\n","protected":false},"author":34512,"featured_media":1061988,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1061985","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1061985","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":1,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1061985\/revisions"}],"predecessor-version":[{"id":1062000,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1061985\/revisions\/1062000"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1061988"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1061985"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1061985"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1061985"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1061985"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}