{"id":405,"date":"2018-08-15T10:57:53","date_gmt":"2018-08-15T02:57:53","guid":{"rendered":"http:\/\/www.luodonggan.com\/wordpress\/?p=405"},"modified":"2018-08-15T10:57:53","modified_gmt":"2018-08-15T02:57:53","slug":"%e5%88%86%e5%b8%83%e5%bc%8f%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0%e5%b9%b3%e5%8f%b0%e5%a4%a7%e6%af%94%e6%8b%bc%ef%bc%9aspark%e3%80%81pmls%e3%80%81tensorflow%e3%80%81mxnet","status":"publish","type":"post","link":"http:\/\/8.152.216.154:8081\/?p=405","title":{"rendered":"\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5927\u6bd4\u62fc\uff1aSpark\u3001PMLS\u3001TensorFlow\u3001MXNet"},"content":{"rendered":"

    \u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u662f\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u4e00\u5927\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u3002\u8fd1\u65e5\u7ebd\u7ea6\u5dde\u7acb\u5927\u5b66\u5e03\u6cd5\u7f57\u5206\u6821\u8ba1\u7b97\u673a\u79d1\u5b66\u4e0e\u5de5\u7a0b\u6559\u6388\u3001Petuum Inc. \u987e\u95ee Murat Demirbas \u548c\u4ed6\u7684\u4e24\u4f4d\u5b66\u751f\u4e00\u8d77\u53d1\u8868\u4e86\u4e00\u7bc7\u5bf9\u6bd4\u73b0\u6709\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u7684\u8bba\u6587\uff0c\u5bf9 Spark\u3001PMLS \u548c TensorFlow \u7b49\u5e73\u53f0\u7684\u67b6\u6784\u548c\u6027\u80fd\u8fdb\u884c\u4e86\u6bd4\u8f83\u548c\u4ecb\u7ecd\u3002<\/span><\/span><\/p>\n

    \u673a\u5668\u5b66\u4e60\uff08\u5c24\u5176\u662f\u6df1\u5ea6\u5b66\u4e60\uff09\u6700\u8fd1\u5df2\u7ecf\u5728\u8bed\u97f3\u8bc6\u522b\u3001\u56fe\u50cf\u8bc6\u522b\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u63a8\u8350\/\u641c\u7d22\u5f15\u64ce\u7b49\u65b9\u9762\u53d6\u5f97\u4e86\u53d8\u9769\u6027\u7684\u6210\u529f\u3002\u8fd9\u4e9b\u6280\u672f\u5728\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u3001\u6570\u5b57\u533b\u7597\u7cfb\u7edf\u3001CRM\u3001\u5e7f\u544a\u3001\u7269\u8054\u7f51\u7b49\u65b9\u9762\u7684\u5e94\u7528\u975e\u5e38\u6709\u524d\u9014\u3002\u5f53\u7136\uff0c\u8d44\u672c\u5e26\u9886\/\u63a8\u52a8\u7740\u673a\u5668\u5b66\u4e60\u52a0\u901f\u53d1\u5c55\uff0c\u6211\u4eec\u770b\u5230\u8fd1\u6bb5\u65f6\u95f4\u4ee5\u6765\u5df2\u7ecf\u8bde\u751f\u4e86\u5f88\u591a\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u3002<\/span><\/p>\n

    \u56e0\u4e3a\u8bad\u7ec3\u8fc7\u7a0b\u6d89\u53ca\u5230\u5de8\u5927\u7684\u6570\u636e\u96c6\u7684\u6a21\u578b\uff0c\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5f80\u5f80\u662f\u5206\u5e03\u5f0f\u7684\uff0c\u5b83\u4eec\u5f80\u5f80\u4f1a\u4f7f\u7528\u5e76\u884c\u7684\u51e0\u5341\u4e2a\u6216\u51e0\u767e\u4e2a\u5de5\u4f5c\u5668\uff08worker\uff09\u6765\u8bad\u7ec3\u6a21\u578b\u3002\u636e\u4f30\u8ba1\uff0c\u5728\u4e0d\u4e45\u7684\u5c06\u6765\uff0c\u6570\u636e\u4e2d\u5fc3\u4e2d\u8fd0\u884c\u7684\u7edd\u5927\u591a\u6570\u4efb\u52a1\u90fd\u5c06\u4f1a\u662f\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u3002<\/span><\/p>\n

    \u6211\u6709\u5206\u5e03\u5f0f\u7cfb\u7edf\u7684\u7814\u7a76\u80cc\u666f\uff0c\u6240\u4ee5\u6211\u4eec\u51b3\u5b9a\u4ece\u5206\u5e03\u5f0f\u7cfb\u7edf\u7684\u89d2\u5ea6\u7814\u7a76\u8fd9\u4e9b\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5e76\u5206\u6790\u5176\u901a\u4fe1\u548c\u63a7\u5236\u5c40\u9650\u3002\u6211\u4eec\u4e5f\u8c03\u67e5\u4e86\u8fd9\u4e9b\u5e73\u53f0\u7684\u5bb9\u9519\u80fd\u529b\u548c\u7f16\u7a0b\u96be\u5ea6\u3002<\/span><\/p>\n

    \u6211\u4eec\u5c06\u8fd9\u4e9b\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5f52\u7c7b\u4e3a\u4e86\u4e09\u5927\u57fa\u672c\u8bbe\u8ba1\u65b9\u6cd5\uff1a<\/span><\/p>\n

1. \u57fa\u672c\u6570\u636e\u6d41\uff08basic dataflow\uff09<\/span><\/p>\n

2. \u53c2\u6570\u670d\u52a1\u5668\u6a21\u578b\uff08parameter-server model\uff09<\/span><\/p>\n

3. \u5148\u8fdb\u6570\u636e\u6d41\uff08advanced dataflow\uff09<\/span><\/p>\n

    \u6211\u4eec\u5bf9\u8fd9\u4e09\u79cd\u65b9\u6cd5\u8fdb\u884c\u4e86\u7b80\u8981\u4ecb\u7ecd\u5e76\u4e3e\u4f8b\u8fdb\u884c\u4e86\u8bf4\u660e\uff0c\u5176\u4e2d\u57fa\u672c\u6570\u636e\u6d41\u65b9\u6cd5\u4f7f\u7528\u4e86 Apache Spark\u3001\u53c2\u6570\u670d\u52a1\u5668\u6a21\u578b\u4f7f\u7528\u4e86 PMLS\uff08Petuum\uff09\u3001\u5148\u8fdb\u6570\u636e\u6d41\u6a21\u578b\u4f7f\u7528\u4e86 TensorFlow \u548c MXNet\u3002\u6211\u4eec\u63d0\u4f9b\u4e86\u51e0\u4e2a\u6bd4\u8f83\u6027\u80fd\u7684\u8bc4\u4f30\u7ed3\u679c\u3002\u8bba\u6587\u91cc\u8fd8\u6709\u66f4\u591a\u8bc4\u4f30\u7ed3\u679c\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u4f5c\u4e3a\u5b66\u672f\u754c\u7684\u4e00\u4e2a\u5c0f\u56e2\u961f\uff0c\u6211\u4eec\u65e0\u6cd5\u8fdb\u884c\u5927\u89c4\u6a21\u7684\u8bc4\u4f30\u3002<\/span><\/p>\n

    \u5728\u672c\u6587\u672b\u5c3e\uff0c\u6211\u7ed9\u51fa\u4e86\u5bf9\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u672a\u6765\u7814\u7a76\u5de5\u4f5c\u7684\u603b\u7ed3\u548c\u5efa\u8bae\u3002\u5982\u679c\u4f60\u5df2\u7ecf\u4e86\u89e3\u8fd9\u4e9b\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\uff0c\u53ef\u4ee5\u76f4\u63a5\u8df3\u81f3\u672b\u5c3e\u67e5\u770b\u7ed3\u8bba\u3002<\/span><\/p>\n

Spark<\/span><\/p>\n

    \u5728 Spark \u4e2d\uff0c\u8ba1\u7b97\u88ab\u5efa\u6a21\u6210\u6709\u5411\u65e0\u73af\u56fe\uff08DAG\uff1adirected acyclic graph\uff09\uff0c\u5176\u4e2d\u6bcf\u4e00\u4e2a\u9876\u70b9\u90fd\u4ee3\u8868\u4e00\u4e2a\u5f39\u6027\u5206\u5e03\u5f0f\u6570\u636e\u96c6\uff08RDD\uff1aResilient Distributed Dataset\uff09\uff0c\u6bcf\u4e00\u6761\u8fb9\u90fd\u4ee3\u8868\u5bf9 RDD \u7684\u4e00\u4e2a\u8fd0\u7b97\u3002RDD \u662f\u88ab\u5206\u5230\u4e86\u4e0d\u540c\u903b\u8f91\u5206\u533a\u7684\u5bf9\u8c61\u7684\u96c6\u5408\uff0c\u8fd9\u4e9b\u903b\u8f91\u5206\u533a\u662f\u4f5c\u4e3a in-memory \u5b58\u50a8\u548c\u5904\u7406\u7684\uff0c\u5e26\u6709\u5230\u78c1\u76d8\u7684 shuffle\/overflow\u3002<\/span><\/p>\n

\u5728\u4e00\u4e2a DAG \u4e2d\uff0c\u4ece\u9876\u70b9 A \u5230\u9876\u70b9 B \u7684\u8fb9 E \u8868\u793a\uff1aRDD B \u662f\u5728 RDD A \u4e0a\u6267\u884c\u8fd0\u7b97 E \u540e\u5f97\u5230\u7684\u7ed3\u679c\u3002\u8fd0\u7b97\u6709\u4e24\u79cd\uff1a\u53d8\u6362\uff08transformation\uff09\u548c\u52a8\u4f5c\uff08action\uff09\u3002\u53d8\u6362\uff08\u6bd4\u5982\uff1a\u6620\u5c04\u3001\u8fc7\u6ee4\u3001\u8fde\u63a5\uff09\u662f\u6307\u5728\u4e00\u4e2a RDD \u4e0a\u6267\u884c\u4e00\u79cd\u8fd0\u7b97\u751f\u6210\u4e00\u4e2a\u65b0\u7684 RDD\u3002<\/span><\/p>\n

\"image.png\"\/<\/span><\/p>\n

    Spark \u7528\u6237\u9700\u8981\u5c06\u8ba1\u7b97\u5efa\u6a21\u4e3a DAG\uff0c\u4ece\u800c\u5728 RDD \u4e0a\u8fdb\u884c\u53d8\u6362\u6216\u8fd0\u884c\u52a8\u4f5c\u3002DAG \u9700\u8981\u88ab\u7f16\u8bd1\u4e3a stage\u3002\u6bcf\u4e2a stage \u4f5c\u4e3a\u4e00\u7cfb\u5217\u5e76\u884c\u8fd0\u884c\u7684\u4efb\u52a1\u6267\u884c\uff08\u6bcf\u4e2a\u5206\u533a\u6267\u884c\u4e00\u4e2a\u4efb\u52a1\uff09\u3002\u7b80\u5355\u72ed\u7a84\u7684\u4f9d\u8d56\u5173\u7cfb\u6709\u5229\u4e8e\u9ad8\u6548\u6267\u884c\uff0c\u800c\u5bbd\u5e7f\u7684\u4f9d\u8d56\u5173\u7cfb\u4f1a\u5f15\u5165\u74f6\u9888\uff0c\u56e0\u4e3a\u5b83\u4eec\u4f1a\u6270\u4e71\u6d41\u7a0b\uff0c\u800c\u4e14\u9700\u8981\u901a\u4fe1\u5bc6\u96c6\u7684 shuffle \u8fd0\u7b97\u3002<\/span><\/p>\n

\"image.png\"\/<\/span><\/p>\n

<\/p>\n

    Spark \u4e2d\u7684\u5206\u5e03\u5f0f\u6267\u884c\u662f\u901a\u8fc7\u5c06\u8fd9\u79cd DAG stage \u5206\u5272\u5230\u4e0d\u540c\u7684\u673a\u5668\u4e0a\u6267\u884c\u7684\u3002\u8fd9\u5f20\u56fe\u6e05\u6670\u5730\u663e\u793a\u4e86\u8fd9\u79cd master-worker \u67b6\u6784\u3002\u9a71\u52a8\u5668\uff08driver\uff09\u5305\u542b\u4e86\u4efb\u52a1\u548c\u4e24\u4e2a\u8c03\u5ea6\u5668\uff08scheduler\uff09\u7ec4\u4ef6\u2014\u2014DAG \u8c03\u5ea6\u5668\u548c\u4efb\u52a1\u8c03\u5ea6\u5668\uff1b\u5e76\u4e14\u8fd8\u8981\u5c06\u4efb\u52a1\u5bf9\u5e94\u5230\u5de5\u4f5c\u5668\u3002<\/span><\/p>\n

    Spark \u662f\u4e3a\u4e00\u822c\u7684\u6570\u636e\u5904\u7406\u8bbe\u8ba1\u7684\uff0c\u5e76\u4e0d\u7279\u5b9a\u4e8e\u673a\u5668\u5b66\u4e60\u3002\u4f46\u662f\u4f7f\u7528 MLlib for Spark\uff0c\u4e5f\u53ef\u4ee5\u5728 Spark \u4e0a\u8fdb\u884c\u673a\u5668\u5b66\u4e60\u3002\u5728\u57fa\u672c\u7684\u8bbe\u7f6e\u4e2d\uff0cSpark \u5c06\u6a21\u578b\u53c2\u6570\u5b58\u50a8\u5728\u9a71\u52a8\u5668\u8282\u70b9\uff0c\u5de5\u4f5c\u5668\u4e0e\u9a71\u52a8\u5668\u901a\u4fe1\u4ece\u800c\u5728\u6bcf\u6b21\u8fed\u4ee3\u540e\u66f4\u65b0\u8fd9\u4e9b\u53c2\u6570\u3002\u5bf9\u4e8e\u5927\u89c4\u6a21\u90e8\u7f72\u800c\u8a00\uff0c\u8fd9\u4e9b\u6a21\u578b\u53c2\u6570\u53ef\u80fd\u5e76\u4e0d\u9002\u5408\u9a71\u52a8\u5668\uff0c\u5e76\u4e14\u4f1a\u4f5c\u4e3a\u4e00\u4e2a RDD \u800c\u8fdb\u884c\u7ef4\u62a4\u66f4\u65b0\u3002\u8fd9\u4f1a\u5e26\u6765\u5927\u91cf\u989d\u5916\u5f00\u9500\uff0c\u56e0\u4e3a\u6bcf\u6b21\u8fed\u4ee3\u90fd\u9700\u8981\u521b\u9020\u4e00\u4e2a\u65b0\u7684 RDD \u6765\u4fdd\u5b58\u66f4\u65b0\u540e\u7684\u6a21\u578b\u53c2\u6570\u3002\u66f4\u65b0\u6a21\u578b\u6d89\u53ca\u5230\u5728\u6574\u4e2a\u673a\u5668\/\u78c1\u76d8\u4e0a\u91cd\u6392\u6570\u636e\uff0c\u8fd9\u5c31\u9650\u5236\u4e86 Spark \u7684\u6269\u5c55\u6027\u3002\u8fd9\u662f Spark \u7684\u57fa\u672c\u6570\u636e\u6d41\u6a21\u578b\uff08DAG\uff09\u7684\u4e0d\u8db3\u4e4b\u5904\u3002Spark \u5e76\u4e0d\u80fd\u5f88\u597d\u5730\u652f\u6301\u673a\u5668\u5b66\u4e60\u6240\u9700\u7684\u8fed\u4ee3\u3002<\/span><\/p>\n

PMLS<\/span><\/p>\n

PMLS \u662f\u4e13\u4e3a\u673a\u5668\u5b66\u4e60\u8bbe\u8ba1\u7684\uff0c\u6ca1\u6709\u5176\u5b83\u6742\u4e71\u7684\u5386\u53f2\u3002\u5b83\u5f15\u5165\u4e86\u53c2\u6570\u670d\u52a1\u5668\uff08PS\uff1a parameter-server\uff09\u7684\u62bd\u8c61\u6982\u5ff5\uff0c\u652f\u6301\u5bc6\u96c6\u8fed\u4ee3\u7684\u673a\u5668\u5b66\u4e60\u8bad\u7ec3\u8fc7\u7a0b\u3002<\/span><\/p>\n

\"image.png\"\/<\/span><\/p>\n

<\/p>\n

    \u5176\u4e2d PS\uff08\u56fe\u4e2d\u7eff\u8272\u65b9\u6846\uff09\u88ab\u7528\u4f5c\u5206\u5e03\u5f0f\u7684\u5185\u5b58\u952e\u503c\u5b58\u50a8\uff08distributed in-memory key-value store\uff09\u3002\u5b83\u4f1a\u88ab\u590d\u5236\u548c\u5171\u4eab\uff1a\u6bcf\u4e2a\u8282\u70b9\u90fd\u88ab\u7528\u4f5c\u8fd9\u4e2a\u6a21\u578b\uff08\u53c2\u6570\u7a7a\u95f4\uff09\u4e00\u4e2a\u5206\u7247\u7684\u4e3b\u8282\u70b9\u4ee5\u53ca\u5176\u5b83\u5206\u7247\u7684\u6b21\u8981\u8282\u70b9\/\u526f\u672c\u3002\u56e0\u6b64\u5728\u8282\u70b9\u6570\u91cf\u65b9\u9762\uff0cPS \u53ef\u4ee5\u5f88\u597d\u5730\u6269\u5c55\u3002<\/span><\/p>\n

    PS \u8282\u70b9\u4f1a\u5b58\u50a8\u548c\u66f4\u65b0\u6a21\u578b\u53c2\u6570\u4ee5\u53ca\u54cd\u5e94\u6765\u81ea\u5de5\u4f5c\u5668\u7684\u8bf7\u6c42\u3002\u5de5\u4f5c\u5668\u4f1a\u8bf7\u6c42\u6765\u81ea\u5b83\u4eec\u7684\u5c40\u90e8 PS \u526f\u672c\u7684\u6700\u65b0\u6a21\u578b\u53c2\u6570\uff0c\u5e76\u5728\u5206\u914d\u7ed9\u5b83\u4eec\u7684\u6570\u636e\u96c6\u90e8\u5206\u4e0a\u6267\u884c\u8ba1\u7b97\u3002<\/span><\/p>\n

    PMLS \u8fd8\u91c7\u7528\u4e86 SSP\uff08Stale Synchronous Parallelism\uff09\u6a21\u578b\uff0c\u8fd9\u6bd4 BSP\uff08Bulk Synchronous Parellelism\uff09\u6a21\u578b\u66f4\u5bbd\u677e\u2014\u2014\u5176\u4e2d\u5de5\u4f5c\u5668\u5728\u6bcf\u6b21\u8fed\u4ee3\u7ed3\u675f\u65f6\u540c\u6b65\u3002SSP \u4e3a\u5de5\u4f5c\u5668\u7684\u540c\u6b65\u51cf\u5c11\u4e86\u9ebb\u70e6\uff0c\u786e\u4fdd\u6700\u5feb\u7684\u5de5\u4f5c\u5668\u4e0d\u80fd\u8d85\u8fc7\u6700\u6162\u7684\u5de5\u4f5c\u5668 s \u6b21\u8fed\u4ee3\u3002\u5bbd\u677e\u7684\u4e00\u81f4\u6027\u6a21\u578b\u4ecd\u7136\u53ef\u4ee5\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u8bad\u7ec3\uff0c\u56e0\u4e3a\u8fd9\u4e2a\u8fc7\u7a0b\u6709\u4e00\u5b9a\u7684\u566a\u58f0\u5bb9\u9519\u80fd\u529b\uff0c\u6211\u5728 2016 \u5e74 4 \u6708\u7684\u8fd9\u7bc7\u6587\u7ae0\u4e2d\u8c08\u8fc7\u8fd9\u4e2a\u95ee\u9898\uff1ahttps:\/\/muratbuffalo.blogspot.com\/2016\/04\/petuum-new-platform-for-distributed.html<\/span><\/p>\n

TensorFlow<\/span><\/p>\n

    \u8c37\u6b4c\u6709\u4e00\u4e2a\u57fa\u4e8e\u53c2\u6570\u670d\u52a1\u5668\u6a21\u578b\u7684\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0 DistBelief\u3002\u53c2\u9605\u6211\u5bf9 DistBelief \u8bba\u6587\u7684\u8bc4\u8bba\uff1ahttps:\/\/muratbuffalo.blogspot.com\/2017\/01\/google-distbelief-paper-large-scale.html\u3002\u5728\u6211\u770b\u6765\uff0cDistBelief \u7684\u4e3b\u8981\u7f3a\u9677\u662f\uff1a\u4e3a\u4e86\u7f16\u5199\u673a\u5668\u5b66\u4e60\u5e94\u7528\uff0c\u9700\u8981\u64cd\u4f5c\u4f4e\u7ea7\u4ee3\u7801\u3002\u8c37\u6b4c\u60f3\u8981\u81ea\u5df1\u7684\u6240\u6709\u5458\u5de5\u65e0\u9700\u7cbe\u901a\u5206\u5e03\u5f0f\u6267\u884c\u5c31\u80fd\u7f16\u5199\u673a\u5668\u5b66\u4e60\u4ee3\u7801\u2014\u2014\u57fa\u4e8e\u540c\u6837\u7684\u7406\u7531\uff0c\u8c37\u6b4c\u4e3a\u5927\u6570\u636e\u5904\u7406\u7f16\u5199\u4e86 MapReduce \u6846\u67b6\u3002<\/span><\/p>\n

    \u6240\u4ee5\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u8c37\u6b4c\u8bbe\u8ba1\u4e86 TensorFlow\u3002TensorFlow \u91c7\u7528\u4e86\u6570\u636e\u6d41\u8303\u5f0f\uff0c\u4f46\u662f\u662f\u4e00\u79cd\u66f4\u9ad8\u7ea7\u7684\u7248\u672c\u2014\u2014\u5176\u4e2d\u8ba1\u7b97\u56fe\u65e0\u9700\u662f DAG\uff0c\u800c\u4e14\u5305\u542b\u5faa\u73af\u4e14\u652f\u6301\u53ef\u53d8\u72b6\u6001\u3002\u6211\u8ba4\u4e3a Naiad \u8bbe\u8ba1\u53ef\u80fd\u5bf9 TensorFlow \u8bbe\u8ba1\u6709\u6240\u5f71\u54cd\u3002<\/span><\/p>\n

    TensorFlow \u4f7f\u7528\u8282\u70b9\u548c\u8fb9\u7684\u6709\u5411\u56fe\u6765\u8868\u793a\u8ba1\u7b97\u3002\u8282\u70b9\u8868\u793a\u8ba1\u7b97\uff0c\u72b6\u6001\u53ef\u53d8\u3002\u800c\u8fb9\u5219\u8868\u793a\u591a\u7ef4\u6570\u636e\u6570\u7ec4\uff08\u5f20\u91cf\uff09\uff0c\u5728\u8282\u70b9\u4e4b\u95f4\u4f20\u8f93\u3002TensorFlow \u9700\u8981\u7528\u6237\u9759\u6001\u58f0\u660e\u8fd9\u79cd\u7b26\u53f7\u8ba1\u7b97\u56fe\uff0c\u5e76\u5bf9\u8be5\u56fe\u4f7f\u7528\u590d\u5199\u548c\u5206\u533a\uff08rewrite & partitioning\uff09\u5c06\u5176\u5206\u914d\u5230\u673a\u5668\u4e0a\u8fdb\u884c\u5206\u5e03\u5f0f\u6267\u884c\u3002\uff08MXNet\uff0c\u5c24\u5176\u662f DyNet \u4f7f\u7528\u4e86\u56fe\u7684\u52a8\u6001\u58f0\u660e\uff0c\u8fd9\u6539\u5584\u4e86\u7f16\u7a0b\u7684\u96be\u5ea6\u548c\u7075\u6d3b\u6027\u3002\uff09<\/span><\/p>\n

\"image.png\"\/<\/p>\n

    TensorFlow \u4e2d\u7684\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u8bad\u7ec3\u4f7f\u7528\u4e86\u5982\u56fe\u6240\u793a\u7684\u53c2\u6570\u670d\u52a1\u5668\u65b9\u6cd5\u3002\u5f53\u4f60\u5728 TensorFlow \u4e2d\u4f7f\u7528 PS \u62bd\u8c61\u65f6\uff0c\u4f60\u5c31\u7528\u5230\u4e86\u53c2\u6570\u670d\u52a1\u5668\u548c\u6570\u636e\u5e76\u884c\u3002TensorFlow \u8ba9\u4f60\u8fd8\u80fd\u505a\u66f4\u590d\u6742\u7684\u4e8b\u60c5\uff0c\u4f46\u90a3\u9700\u8981\u7f16\u5199\u81ea\u5b9a\u4e49\u4ee3\u7801\u5e76\u8fdb\u5165\u5168\u65b0\u7684\u7586\u57df\u3002<\/span><\/p>\n

\u4e00\u4e9b\u8bc4\u4f30\u7ed3\u679c<\/span><\/p>\n

    \u6211\u4eec\u7684\u8bc4\u4f30\u4f7f\u7528\u4e86 Amazon EC2 m4.xlarge \u5b9e\u4f8b\u3002\u6bcf\u4e2a\u5b9e\u4f8b\u5305\u542b 4 \u4e2a\u7531 Intel Xeon E5-2676 v3 \u9a71\u52a8\u7684 vCPU \u548c 16 GiB RAM\u3002EBS \u5e26\u5bbd\u4e3a 750Mbps\u3002\u6211\u4eec\u4f7f\u7528\u4e86\u4e24\u4e2a\u5e38\u89c1\u7684\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u8fdb\u884c\u8bc4\u4f30\uff1a\u4e8c\u5206\u7c7b logistic \u56de\u5f52\u548c\u4f7f\u7528\u591a\u5c42\u795e\u7ecf\u7f51\u7edc\u7684\u56fe\u50cf\u5206\u7c7b\u3002\u6211\u5728\u8fd9\u91cc\u4ec5\u7ed9\u51fa\u4e86\u51e0\u5f20\u56fe\uff0c\u67e5\u770b\u6211\u4eec\u7684\u8bba\u6587\u53ef\u4ee5\u4e86\u89e3\u66f4\u591a\u5b9e\u9a8c\u3002\u4f46\u6211\u4eec\u7684\u5b9e\u9a8c\u8fd8\u6709\u4e00\u4e9b\u5c40\u9650\u6027\uff1a\u6211\u4eec\u4f7f\u7528\u4e86\u5c11\u91cf\u673a\u5668\uff0c\u4e0d\u80fd\u5927\u89c4\u6a21\u6d4b\u8bd5\u3002\u6211\u4eec\u4e5f\u9650\u5236\u4e86 CPU \u8ba1\u7b97\uff0c\u6ca1\u6709\u6d4b\u8bd5 GPU\u3002<\/span><\/p>\n

\"image.png\"\/<\/p>\n

    \u8fd9\u5e45\u56fe\u5c55\u793a\u4e86\u5404\u5e73\u53f0\u7684 logistic \u56de\u5f52\u6267\u884c\u901f\u5ea6\u3002Spark \u8868\u73b0\u4e0d\u9519\uff0c\u4f46\u843d\u540e\u4e8e PMLS \u548c MXNet\u3002<\/span><\/p>\n

\"image.png\"\/<\/p>\n

    \u8fd9\u5e45\u56fe\u5c55\u793a\u4e86\u5404\u5e73\u53f0\u7684\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff08DNN\uff09\u6267\u884c\u901f\u5ea6\u3002\u76f8\u6bd4\u4e8e\u5355\u5c42\u7684 logistic \u56de\u5f52\uff0cSpark \u5728\u4e24\u5c42\u795e\u7ecf\u7f51\u7edc\u4e0a\u6709\u66f4\u5927\u7684\u6027\u80fd\u635f\u5931\u3002\u8fd9\u662f\u56e0\u4e3a\u4e24\u5c42\u7f51\u7edc\u9700\u8981\u66f4\u591a\u8fed\u4ee3\u8ba1\u7b97\u3002\u5728 Spark \u4e2d\u6211\u4eec\u5c06\u53c2\u6570\u4fdd\u5b58\u5728\u9a71\u52a8\u5668\u4e2d\uff0c\u8fd9\u6837\u5b83\u4eec\u53ef\u4ee5\u62df\u5408\uff1b\u5982\u679c\u6211\u4eec\u5c06\u53c2\u6570\u4fdd\u5b58\u5728\u4e00\u4e2a RDD \u4e2d\u5e76\u4e14\u5728\u6bcf\u6b21\u8fed\u4ee3\u540e\u66f4\u65b0\uff0c\u60c5\u51b5\u8fd8\u4f1a\u53d8\u5f97\u66f4\u52a0\u7cdf\u7cd5\u3002<\/span><\/p>\n

\"image.png\"\/<\/p>\n

    \u8fd9\u5e45\u56fe\u7ed9\u51fa\u4e86\u5404\u5e73\u53f0\u7684 CPU \u5229\u7528\u7387\u3002Spark \u5e94\u7528\u4f3c\u4e4e\u6709\u660e\u663e\u5f88\u9ad8\u7684 CPU \u5229\u7528\u7387\uff0c\u8fd9\u4e3b\u8981\u662f\u56e0\u4e3a\u5e8f\u5217\u5316\uff08serialization\uff09\u7684\u989d\u5916\u5f00\u9500\u3002\u6211\u4eec\u66f4\u65e9\u671f\u7684\u5de5\u4f5c\u5df2\u7ecf\u6307\u51fa\u4e86\u8fd9\u4e00\u95ee\u9898\uff1ahttps:\/\/muratbuffalo.blogspot.com\/2017\/05\/paper-summary-making-sense-of.html<\/span><\/p>\n

\u603b\u7ed3\u4e0e\u672a\u6765\u65b9\u5411<\/span><\/p>\n

    \u673a\u5668\u5b66\u4e60\/\u6df1\u5ea6\u5b66\u4e60\u5e94\u7528\u7684\u5e76\u884c\u5904\u7406\u8ba9\u4eba\u4e3a\u96be\uff0c\u800c\u4e14\u4ece\u5e76\u53d1\u7b97\u6cd5\uff08concurrent algorithms\uff09\u7684\u89d2\u5ea6\u770b\u5e76\u4e0d\u975e\u5e38\u6709\u8da3\u3002\u53ef\u4ee5\u76f8\u5f53\u80af\u5b9a\u5730\u8bf4\u53c2\u6570\u670d\u52a1\u5668\u65b9\u6cd5\u5728\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u7684\u8bad\u7ec3\u4e0a\u66f4\u597d\u3002<\/span><\/p>\n

    \u81f3\u4e8e\u5c40\u9650\u6027\u65b9\u9762\uff0c\u7f51\u7edc\u4ecd\u7136\u662f\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e94\u7528\u7684\u4e00\u4e2a\u74f6\u9888\u3002\u63d0\u4f9b\u66f4\u597d\u7684\u6570\u636e\/\u6a21\u578b\u5206\u7ea7\u6bd4\u66f4\u5148\u8fdb\u7684\u901a\u7528\u6570\u636e\u6570\u636e\u6d41\u5e73\u53f0\u66f4\u6709\u7528\uff1b\u5e94\u8be5\u5c06\u6570\u636e\/\u6a21\u578b\u770b\u4f5c\u5934\u7b49\u516c\u6c11\u3002<\/span><\/p>\n

    \u4f46\u662f\uff0c\u53ef\u80fd\u4f1a\u6709\u4e00\u4e9b\u8ba9\u4eba\u60ca\u5947\u548c\u5fae\u5999\u7684\u5730\u65b9\u3002\u5728 Spark \u4e2d\uff0cCPU \u5f00\u9500\u4f1a\u5148\u4e8e\u7f51\u7edc\u9650\u5236\u53d8\u6210\u74f6\u9888\u3002Spark \u4f7f\u7528\u7684\u7f16\u7a0b\u8bed\u8a00 Scala\/JVM \u663e\u8457\u5f71\u54cd\u4e86\u5176\u6027\u80fd\u8868\u73b0\u3002\u56e0\u6b64\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5c24\u5176\u9700\u8981\u66f4\u597d\u7684\u76d1\u63a7\u548c\/\u6216\u6027\u80fd\u9884\u6d4b\u5de5\u5177\u3002\u6700\u8fd1\u5df2\u7ecf\u6709\u4eba\u63d0\u51fa\u4e86\u4e00\u4e9b\u89e3\u51b3 Spark \u6570\u636e\u5904\u7406\u5e94\u7528\u7684\u95ee\u9898\u7684\u5de5\u5177\uff0c\u6bd4\u5982 Ernest \u548c CherryPick\u3002<\/span><\/p>\n

    \u5728\u673a\u5668\u5b66\u4e60\u8fd0\u884c\u65f6\u7684\u5206\u5e03\u5f0f\u7cfb\u7edf\u652f\u6301\u4e0a\u8fd8\u6709\u5f88\u591a\u60ac\u800c\u672a\u51b3\u7684\u95ee\u9898\uff0c\u6bd4\u5982\u8d44\u6e90\u8c03\u5ea6\u548c\u8fd0\u884c\u65f6\u7684\u6027\u80fd\u63d0\u5347\u3002\u5bf9\u5e94\u7528\u4f7f\u7528\u8fd0\u884c\u65f6\u76d1\u63a7\/\u6027\u80fd\u5206\u6790\uff0c\u4e0b\u4e00\u4ee3\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e73\u53f0\u5e94\u8be5\u4f1a\u63d0\u4f9b\u4efb\u52a1\u8fd0\u884c\u7684\u8ba1\u7b97\u3001\u5185\u5b58\u3001\u7f51\u7edc\u8d44\u6e90\u7684\u8be6\u7ec6\u7684\u8fd0\u884c\u65f6\u5f39\u6027\u914d\u7f6e\/\u8c03\u5ea6\u3002<\/span><\/p>\n

    \u6700\u540e\uff0c\u5728\u7f16\u7a0b\u548c\u8f6f\u4ef6\u5de5\u7a0b\u652f\u6301\u65b9\u9762\u4e5f\u6709\u4e00\u4e9b\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002\u4ec0\u4e48\u6837\u7684\uff08\u5206\u5e03\u5f0f\uff09\u7f16\u7a0b\u62bd\u8c61\u601d\u60f3\u9002\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u5e94\u7528\uff1f\u53e6\u5916\u5728\u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u5e94\u7528\u7684\u68c0\u9a8c\u548c\u9a8c\u8bc1\uff08\u5c24\u5176\u662f\u4f7f\u7528\u6709\u95ee\u9898\u7684\u8f93\u5165\u6765\u6d4b\u8bd5 DNN\uff09\u4e0a\u4e5f\u8fd8\u9700\u8981\u66f4\u591a\u7814\u7a76\u3002<\/span><\/p>\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u5206\u5e03\u5f0f\u673a\u5668\u5b66\u4e60\u662f\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u4e00\u5927\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u3002\u8fd1\u65e5\u7ebd\u7ea6\u5dde\u7acb\u5927 […]<\/p>\n","protected":false},"author":1,"featured_media":406,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[5],"tags":[29],"_links":{"self":[{"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/posts\/405"}],"collection":[{"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=405"}],"version-history":[{"count":1,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/posts\/405\/revisions"}],"predecessor-version":[{"id":407,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/posts\/405\/revisions\/407"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=\/wp\/v2\/media\/406"}],"wp:attachment":[{"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=405"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=405"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/8.152.216.154:8081\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=405"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}