diff --git a/demo-serving/BCLOUD b/demo-serving/BCLOUD index e95c7e3d455e5b34f3af404199ccadf15cc70321..b4d6ea752b20f285d30198c309c69d8fa4fc9784 100644 --- a/demo-serving/BCLOUD +++ b/demo-serving/BCLOUD @@ -41,7 +41,7 @@ cpp_source_dirs.append('proto/*.proto') # DELETE_AUTO_LIBS('$OUT_ROOT/third-64/glog/output/lib/libglog.a') # DELETE_AUTO_LIBS('$OUT_ROOT/third-64/gflags/output/lib/libgflags.a') # DELETE_AUTO_LIBS('$OUT_ROOT/third-64/gflags/output/lib/libgflags_nothreads.a') -# DELETE_AUTO_LIBS('$OUT_ROOT/baidu/paddlepaddle/paddle/output/lib/libpaddle_fluid_noavx_openblas.a') +DELETE_AUTO_LIBS('$OUT_ROOT/baidu/paddlepaddle/paddle/output/lib/libpaddle_fluid_noavx_openblas.a') PROTOFLAGS( '--plugin=protoc-gen-pdcodegen=../pdcodegen/plugin/pdcodegen', diff --git a/doc/BENCHMARKING.md b/doc/BENCHMARKING.md index 4218f0c28e745aed4d251116531e23ccb3535811..3544498e2ee7f8bdf6c32d7e16d226f7c4d81b84 100644 --- a/doc/BENCHMARKING.md +++ b/doc/BENCHMARKING.md @@ -10,7 +10,7 @@ 下图是一个对serving请求的耗时阶段的不完整分析。图中对brpc的开销，只列出了bthread创建和启动开销。 -![](http://paddle-serving.bj.bcebos.com/doc/serving-timings.png) +![](serving-timings.png) (右键在新窗口中浏览大图) @@ -280,7 +280,7 @@ Serving扩展能力的测试是指，在不同模型上： 下图是Paddle Serving在BOW模型上QPS随serving端线程数增加而变化的图表。可以看出当线程数较少时(4线程/8线程/12线程)，QPS的变化规律非常杂乱；当线程数较多时，QPS曲线又基本趋于一致，基本无线性增长关系。 -![](https://paddle-serving.bj.bcebos.com/doc/qps-threads-bow.png) +![](qps-threads-bow.png) (右键在新窗口中浏览大图) @@ -438,7 +438,7 @@ Serving扩展能力的测试是指，在不同模型上： 下图是Paddle Serving在CNN模型上QPS随serving端线程数增加而变化的图表。可以看出，随着线程数变大，Serving QPS有较为明显的线性增长关系。可以这样解释此图表：例如，线程数为16时，基本在20个并发时达到最大QPS，此后再增加并发压力QPS基本保持稳定；当线程能够数为24线程时，基本在28并发时达到最大QPS，此后再增大并发压力qps基本保持稳定。 -![](https://paddle-serving.bj.bcebos.com/doc/qps-threads-cnn.png) +![](qps-threads-cnn.png) (右键在新窗口中浏览大图) @@ -596,7 +596,7 @@ Serving扩展能力的测试是指，在不同模型上： 下图是Paddle Serving在LSTM模型上QPS随serving端线程数增加而变化的图表。可以看出，随着线程数变大，Serving QPS有较为明显的线性增长关系。可以这样解释此图表：例如，线程数为16时，基本在20个并发时达到最大QPS，此后再增加并发压力QPS基本保持稳定；当线程能够数为24线程时，基本在28并发时达到最大QPS，此后再增大并发压力qps基本保持稳定。 -![](https://paddle-serving.bj.bcebos.com/doc/qps-threads-lstm.png) +![](qps-threads-lstm.png) (右键在新窗口中浏览大图) diff --git a/doc/GPU_BENCHMARKING.md b/doc/GPU_BENCHMARKING.md new file mode 100755 index 0000000000000000000000000000000000000000..d2978ec8458a32149668f3c3bf5e88cb3919dd92 --- /dev/null +++ b/doc/GPU_BENCHMARKING.md @@ -0,0 +1,1086 @@ +Paddle 第一版发布的demo中已经包含图像分类任务，不过当时Serving还只支持CPU预测，而图像分类的计算量太大，因此只提供了demo，没有特别强调Serving响应时间。 + +本次提测的Serving版本，支持GPU预测，希望以此任务为例，对Paddle Serving支持GPU预测的性能给出测试数据。 + + + +Paddle Serving依赖PaddlePaddle预测库进行实际的计算，因此Serving的性能很大程度上依赖预测库。Paddle预测库在GPU上有如下限制： + +1) PaddlePaddle预测库只能单卡预测 + +2) 单个Serving进程内，所有预测任务共用同一个GPU stream，多个并发请求的GPU操作会被串行到stream内，因此不是完全并行；最坏情况下，2个并发请求会是串行计算。 + +这样当多个请求到达时，预期会出现较为严重的排队情况。 + + + +基于以上已知的问题，设计以下几种测试场景： + +**1) 单卡单并发，随batch size变大的响应时间变化情况** + +**2) 单卡多并发，固定batch size，观察响应时间随并发数变大的变化情况** + +这是为了观察和验证Serving进程内请求排队的情况。 + +**3) 单卡多进程，随batch size变大的响应时间变化情况** + +单卡多进程：启动N个Serving实例，他们绑定到同一张GPU卡。 + +如果单个Serving进程不能将GPU卡的算力打满（例如，只有不到50%），那理论上多个Serving进程使用同1张卡，可以提高的GPU的使用率。本测试项就是为了观察和验证这种场景。 + + (1) 每个Seving实例的请求并发数为1 + + (2) 每个Serving实例的请求并发数为N + +**4) 多卡多并发，随batch size变大的响应时间变化情况** + +多卡多并发：启动N个Serving实例，每个实例绑定一张GPU卡，观察每个实例上响应时间和qps等数据随请求并发数变大的变化情况 + +理论上，系统上多张卡之间应该互不干扰。本测试项是为了观察和验证多张卡同时跑任务时的性能。 + + + +# 1. 测试环境说明 + +| | GPU | 显存 | CPU | 内存 | +|----------|---------|----------|----------------------------------------------|------| +| Serving端 | P40 x 4 | 22912MiB | Intel(R) Xeon(R) Gold 5117 CPU @ 2.00GHz 26核 | 216G | +| Client端 | P40 x 4 | 22912MiB | Intel(R) Xeon(R) Gold 5117 CPU @ 2.00GHz 26核 | 216G | + +网络时延：2.63毫秒 + +# 2. 单机性能测试 +在进行Serving的性能测试前，先要摸清图像分类GPU单机预测的性能 + +## 2.1 单卡单并发 +确保没有其他进程占用GPU。 + +启动单个测试进程，绑定GPU0，concurrency = 1, 持续增加batch size（1, 4, 8, 12…）记录平均响应时间；直到平响突然急剧增大。这时可以认为GPU kernel已经全部用满。 + + +| 并发数 | batch size | 总时间 | qps | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|--------|----------|----------|----------|----------|----------|------------| +| 1 | 2 | 2224 | 43.61511 | 22.92784 | 23 | 23 | 25 | 25 | +| 1 | 4 | 3006 | 32.2688 | 30.98969 | 31 | 31 | 33 | 33 | +| 1 | 6 | 3751 | 25.85977 | 38.67011 | 39 | 39 | 50 | 50 | +| 1 | 8 | 4831 | 20.07866 | 49.80412 | 50 | 50 | 50 | 50 | +| 1 | 10 | 5492 | 17.66205 | 56.61856 | 57 | 57 | 58 | 58 | +| 1 | 12 | 6283 | 15.43849 | 64.77319 | 65 | 65 | 66 | 66 | +| 1 | 14 | 6815 | 14.23331 | 70.25773 | 71 | 71 | 71 | 71 | +| 1 | 16 | 7729 | 12.55014 | 79.68041 | 80 | 80 | 91 | 91 | +| 1 | 18 | 8545 | 11.35167 | 88.09278 | 88 | 89 | 90 | 90 | +| 1 | 20 | 9292 | 10.43909 | 95.79382 | 96 | 96 | 97 | 97 | +| 1 | 22 | 10049 | 9.652702 | 103.5979 | 104 | 104 | 107 | 107 | +| 1 | 24 | 10927 | 8.877093 | 112.6495 | 113 | 113 | 115 | 115 | +| 1 | 26 | 13555 | 7.156031 | 139.7423 | 140 | 141 | 144 | 144 | +| 1 | 28 | 14255 | 6.80463 | 146.9588 | 148 | 148 | 149 | 149 | +| 1 | 30 | 15095 | 6.425969 | 155.6186 | 156 | 157 | 176 | 176 | +| 1 | 32 | 16076 | 6.033839 | 165.732 | 166 | 167 | 170 | 170 | +| 1 | 34 | 16646 | 5.827226 | 171.6082 | 172 | 173 | 178 | 178 | +| 1 | 36 | 17345 | 5.59239 | 178.8144 | 179 | 182 | 187 | 187 | +| 1 | 38 | 18366 | 5.281498 | 189.3402 | 191 | 193 | 216 | 216 | +| 1 | 40 | 18870 | 5.140435 | 194.5361 | 195 | 195 | 198 | 198 | +| 1 | 42 | 19677 | 4.929613 | 202.8557 | 203 | 204 | 214 | 214 | +| 1 | 44 | 20327 | 4.771978 | 209.5567 | 210 | 210 | 211 | 211 | +| 1 | 46 | 21281 | 4.558056 | 219.3918 | 220 | 220 | 221 | 221 | +| 1 | 48 | 22140 | 4.38121 | 228.2474 | 229 | 231 | 241 | 241 | +| 1 | 50 | 26627 | 3.642919 | 274.5052 | 275 | 276 | 279 | 279 | +| 1 | 52 | 27507 | 3.526375 | 283.5773 | 284 | 286 | 292 | 292 | +| 1 | 54 | 28311 | 3.42623 | 291.866 | 292 | 293 | 296 | 296 | +| 1 | 56 | 27004 | 3.59206 | 278.3918 | 279 | 281 | 296 | 296 | +| 1 | 58 | 30151 | 3.21714 | 310.8351 | 312 | 315 | 336 | 336 | +| 1 | 60 | 31092 | 3.119774 | 320.5361 | 324 | 329 | 359 | 359 | +| 1 | 62 | 31783 | 3.051946 | 327.6598 | 328 | 329 | 333 | 333 | +| 1 | 64 | 30644 | 3.165383 | 315.9175 | 316 | 320 | 342 | 342 | +| 1 | 66 | 34820 | 2.785755 | 358.9691 | 362 | 362 | 364 | 364 | +| 1 | 68 | 35824 | 2.707682 | 369.3196 | 371 | 373 | 381 | 381 | +| 1 | 70 | 36945 | 2.625524 | 380.8763 | 383 | 387 | 436 | 436 | +| 1 | 72 | 34425 | 2.81772 | 354.8969 | 357 | 361 | 369 | 369 | +| 1 | 74 | 38227 | 2.537474 | 394.0928 | 396 | 396 | 437 | 437 | +| 1 | 76 | 39079 | 2.482152 | 402.8763 | 405 | 405 | 415 | 415 | +| 1 | 78 | 40135 | 2.416843 | 413.7629 | 416 | 421 | 441 | 441 | +| 1 | 80 | 37924 | 2.557747 | 390.9691 | 393 | 400 | 417 | 417 | +| 1 | 82 | 41943 | 2.312662 | 432.4021 | 437 | 442 | 462 | 462 | +| 1 | 84 | 43172 | 2.246827 | 445.0722 | 452 | 458 | 473 | 473 | +| 1 | 86 | 43468 | 2.231527 | 448.1237 | 449 | 451 | 471 | 471 | +| 1 | 88 | 41309 | 2.348157 | 425.866 | 427 | 434 | 484 | 484 | +| 1 | 90 | 45175 | 2.147205 | 465.7216 | 467 | 469 | 491 | 491 | +| 1 | 92 | 46655 | 2.079091 | 480.9794 | 485 | 490 | 519 | 519 | +| 1 | 94 | 47312 | 2.05022 | 487.7526 | 490 | 491 | 493 | 493 | +| 1 | 96 | 44671 | 2.171431 | 460.5258 | 461 | 465 | 487 | 487 | +| 1 | 98 | 50541 | 1.919234 | 521.0413 | 526 | 535 | 575 | 575 | +| 1 | 100 | 51237 | 1.893163 | 528.2165 | 531 | 542 | 554 | 554 | +| 1 | 102 | 51835 | 1.871322 | 534.3815 | 535 | 536 | 537 | 537 | +| 1 | 104 | 48575 | 1.996912 | 500.7732 | 501 | 502 | 508 | 508 | +| 1 | 106 | 53755 | 1.804483 | 554.1752 | 555 | 556 | 567 | 567 | +| 1 | 108 | 54626 | 1.775711 | 563.1547 | 564 | 568 | 585 | 585 | +| 1 | 110 | 55572 | 1.745483 | 572.9072 | 573 | 576 | 595 | 595 | +| 1 | 112 | 52238 | 1.856886 | 538.5361 | 539 | 540 | 555 | 555 | +| 1 | 114 | 57482 | 1.687485 | 592.598 | 594 | 603 | 609 | 609 | +| 1 | 116 | 58634 | 1.65433 | 604.4742 | 612 | 617 | 654 | 654 | +| 1 | 118 | 59014 | 1.643678 | 608.3917 | 609 | 610 | 627 | 627 | +| 1 | 120 | 55744 | 1.740098 | 574.6804 | 576 | 576 | 582 | 582 | +| 1 | 122 | 61507 | 1.577056 | 634.0928 | 633 | 636 | 704 | 704 | +| 1 | 124 | 62304 | 1.556882 | 642.3093 | 643 | 646 | 664 | 664 | +| 1 | 126 | 63553 | 1.526285 | 655.1855 | 663 | 676 | 689 | 689 | +| 1 | 128 | 59575 | 1.6282 | 614.1752 | 615 | 616 | 621 | 621 | +| 1 | 130 | 66095 | 1.467585 | 681.3917 | 686 | 693 | 708 | 708 | +| 1 | 132 | 66975 | 1.448302 | 690.4639 | 691 | 694 | 725 | 725 | +| 1 | 134 | 67392 | 1.43934 | 694.7629 | 696 | 696 | 700 | 700 | +| 1 | 136 | 63605 | 1.525037 | 655.7217 | 657 | 658 | 661 | 661 | +| 1 | 138 | 69546 | 1.39476 | 716.9691 | 718 | 726 | 765 | 765 | +| 1 | 140 | 70188 | 1.382003 | 723.5876 | 725 | 725 | 728 | 728 | +| 1 | 142 | 70855 | 1.368993 | 730.4639 | 731 | 732 | 746 | 746 | +| 1 | 144 | 67093 | 1.445754 | 691.6804 | 693 | 694 | 703 | 703 | +| 1 | 146 | 72910 | 1.330407 | 751.6495 | 754 | 759 | 778 | 778 | +| 1 | 148 | 73901 | 1.312567 | 761.866 | 764 | 765 | 768 | 768 | +| 1 | 150 | 75287 | 1.288403 | 776.1547 | 776 | 779 | 835 | 835 | +| 1 | 152 | 72538 | 1.33723 | 747.8145 | 769 | 789 | 964 | 964 | +| 1 | 154 | 78081 | 1.2423 | 804.9587 | 800 | 837 | 1007 | 1007 | +| 1 | 156 | 78475 | 1.236062 | 809.0206 | 822 | 829 | 839 | 839 | +| 1 | 158 | 78480 | 1.235984 | 809.0721 | 809 | 814 | 852 | 852 | +| 1 | 160 | 74521 | 1.301647 | 768.2578 | 767 | 780 | 911 | 911 | +| 1 | 162 | 81131 | 1.195597 | 836.402 | 836 | 845 | 887 | 887 | +| 1 | 164 | 82448 | 1.176499 | 849.9794 | 859 | 867 | 902 | 902 | +| 1 | 168 | 78797 | 1.231011 | 812.3402 | 813 | 829 | 855 | 855 | +| 1 | 170 | 85168 | 1.138925 | 878.0206 | 885 | 890 | 932 | 932 | +| 1 | 172 | 86894 | 1.116303 | 895.8145 | 898 | 935 | 1045 | 1045 | +| 1 | 174 | 90203 | 1.075352 | 929.9279 | 966 | 1001 | 1147 | 1147 | +| 1 | 176 | 85560 | 1.133707 | 882.0618 | 915 | 972 | 1078 | 1078 | +| 1 | 178 | 88954 | 1.090451 | 917.0516 | 918 | 924 | 956 | 956 | +| 1 | 180 | 90373 | 1.073329 | 931.6804 | 930 | 955 | 1028 | 1028 | +| 1 | 182 | 91581 | 1.059172 | 944.134 | 941 | 949 | 1251 | 1251 | +| 1 | 184 | 86973 | 1.115289 | 896.6288 | 892 | 920 | 1364 | 1364 | +| 1 | 186 | 93538 | 1.037012 | 964.3093 | 970 | 989 | 1101 | 1101 | +| 1 | 188 | 94698 | 1.024309 | 976.2681 | 993 | 1015 | 1074 | 1074 | +| 1 | 190 | 95286 | 1.017988 | 982.3299 | 983 | 1018 | 1113 | 1113 | +| 1 | 192 | 91030 | 1.065583 | 938.4536 | 930 | 962 | 1441 | 1441 | +| 1 | 194 | 97956 | 0.990241 | 1009.856 | 1011 | 1032 | 1098 | 1098 | +| 1 | 196 | 99376 | 0.976091 | 1024.495 | 1030 | 1053 | 1128 | 1128 | +| 1 | 198 | 100879 | 0.961548 | 1039.99 | 1048 | 1078 | 1647 | 1647 | +| 1 | 200 | 94553 | 1.02588 | 974.7732 | 979 | 1005 | 1125 | 1125 | +| 1 | 202 | 101979 | 0.951176 | 1051.33 | 1060 | 1074 | 1231 | 1231 | +| 1 | 204 | 103626 | 0.936059 | 1068.309 | 1083 | 1106 | 1226 | 1226 | +| 1 | 206 | 104283 | 0.930161 | 1075.083 | 1083 | 1108 | 1440 | 1440 | +| 1 | 208 | 98584 | 0.983932 | 1016.33 | 1031 | 1041 | 1087 | 1087 | +| 1 | 210 | 106025 | 0.914879 | 1093.041 | 1110 | 1132 | 1197 | 1197 | +| 1 | 212 | 106613 | 0.909833 | 1099.103 | 1113 | 1131 | 1159 | 1159 | +| 1 | 214 | 107403 | 0.903141 | 1107.247 | 1117 | 1127 | 1209 | 1209 | +| 1 | 216 | 102126 | 0.949807 | 1052.845 | 1066 | 1085 | 1187 | 1187 | +| 1 | 218 | 109069 | 0.889345 | 1124.423 | 1137 | 1147 | 1173 | 1173 | +| 1 | 220 | 110427 | 0.878408 | 1138.423 | 1151 | 1167 | 1220 | 1220 | +| 1 | 222 | 111146 | 0.872726 | 1145.835 | 1164 | 1181 | 1216 | 1216 | +| 1 | 224 | 105529 | 0.919179 | 1087.928 | 1096 | 1107 | 1625 | 1625 | +| 1 | 226 | 114467 | 0.847406 | 1180.072 | 1190 | 1214 | 1258 | 1258 | +| 1 | 228 | 115350 | 0.840919 | 1189.175 | 1194 | 1217 | 1516 | 1516 | +| 1 | 230 | 115920 | 0.836784 | 1195.052 | 1211 | 1226 | 1262 | 1262 | +| 1 | 232 | 109363 | 0.886954 | 1127.454 | 1141 | 1157 | 1219 | 1219 | +| 1 | 234 | 118202 | 0.820629 | 1218.577 | 1232 | 1247 | 1300 | 1300 | +| 1 | 236 | 119203 | 0.813738 | 1228.897 | 1244 | 1260 | 1294 | 1294 | +| 1 | 238 | 119409 | 0.812334 | 1231.021 | 1238 | 1252 | 1295 | 1295 | +| 1 | 240 | 113127 | 0.857443 | 1166.258 | 1186 | 1192 | 1208 | 1208 | +| 1 | 242 | 121781 | 0.796512 | 1255.474 | 1267 | 1281 | 1542 | 1542 | +| 1 | 244 | 122538 | 0.791591 | 1263.278 | 1279 | 1291 | 1338 | 1338 | +| 1 | 246 | 123959 | 0.782517 | 1277.928 | 1298 | 1308 | 1407 | 1407 | +| 1 | 248 | 117470 | 0.825743 | 1211.031 | 1226 | 1259 | 1330 | 1330 | +| 1 | 250 | 126083 | 0.769334 | 1299.825 | 1307 | 1332 | 1704 | 1704 | +| 1 | 252 | 126090 | 0.769292 | 1299.897 | 1306 | 1342 | 1387 | 1387 | +| 1 | 254 | 127954 | 0.758085 | 1319.113 | 1343 | 1357 | 1507 | 1507 | +| 1 | 256 | 122189 | 0.793852 | 1259.68 | 1277 | 1293 | 1620 | 1620 | + + +响应时间随batch size的变化趋势如下： + +![](gpu-local-time-batchsize.png) + +结论：在batch size小于256前提下，随着batch size的增加，平均响应时间和各分位响应时间也随之增加，其中99.9分位的响应时间浮动较大，其它指标基本线性增长。 + + +qps随batch size的变化趋势如下： + +![](gpu-local-qps-batchsize.png) + +结论：在batch size小于256前提下，随着batch size的增加，qps逐渐下降。 + + + +## 2.2 单卡多并发 + +固定batch size（例如，可以固定为batch size= 10。以下所有场景中所有固定batch size的均以此为准），启动单个测试进程，绑定GPU0，concurrency = 1, 2, ... + +如本文开头所述，目前Paddle预测库GPU预测只用1个stream，并行的请求会被串行计算。本测试项是为了观察和验证这种场景下的性能表现。预期随着并发数增加，平响和99分位响应时间会急剧变大，但总qps略有上涨后保持稳定或开始下降（并发请求数将stream占满后，每秒内处理能力达到极限）。 + + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|----------|----------|----------|----------|----------|------------| +| 1 | 10 | 17.515349 | 5538.0 | 57.09278 | 57 | 58 | 60 | 60 | +| 2 | 10 | 28.65795 | 6769.5 | 69.78866 | 91 | 102 | 113 | 114 | +| 3 | 10 | 29.04675 | 10018.33 | 103.2818 | 131 | 135 | 159 | 167 | +| 4 | 10 | 26.34661 | 14726.75 | 151.8222 | 183 | 196 | 224 | 227 | +| 5 | 10 | 26.8101 | 18090.2 | 186.4969 | 236 | 251 | 279 | 294 | +| 6 | 10 | 24.38173 | 23870.33 | 246.0859 | 299 | 312 | 339 | 354 | +| 7 | 10 | 26.59259 | 25533.43 | 263.2312 | 328 | 341 | 379 | 387 | +| 8 | 10 | 27.77703 | 27936.75 | 288.0077 | 372 | 391 | 422 | 496 | +| 9 | 10 | 28.11504 | 31051 | 320.1134 | 410 | 436 | 480 | 504 | +| 10 | 10 | 25.72371 | 37708.4 | 388.7464 | 465 | 485 | 524 | 547 | +| 11 | 10 | 26.89542 | 39672.18 | 408.9916 | 513 | 537 | 600 | 659 | +| 12 | 10 | 26.08179 | 44628.83 | 460.0911 | 556 | 581 | 731 | 997 | +| 13 | 10 | 26.28853 | 47967.69 | 494.5123 | 601 | 624 | 672 | 727 | +| 14 | 10 | 25.59918 | 53048.57 | 546.8925 | 656 | 679 | 728 | 787 | +| 15 | 10 | 24.6679 | 58983.54 | 608.0776 | 699 | 727 | 808 | 938 | +| 16 | 10 | 25.56311 | 60712.5 | 625.902 | 746 | 775 | 842 | 932 | +| 17 | 10 | 23.53067 | 70078.77 | 722.4615 | 797 | 828 | 915 | 1090 | +| 18 | 10 | 22.67992 | 76984.39 | 793.6535 | 870 | 906 | 1093 | 1203 | +| 19 | 10 | 23.32594 | 79010.73 | 814.5437 | 900 | 938 | 1362 | 1755 | +| 20 | 10 | 22.71971 | 85388.4 | 880.2928 | 950 | 983 | 1132 | 1307 | +| 21 | 10 | 22.76425 | 89482.43 | 922.4993 | 1011 | 1057 | 1290 | 1475 | +| 22 | 10 | 22.47889 | 94933.5 | 978.6959 | 1070 | 1110 | 1228 | 1329 | +| 23 | 10 | 22.10953 | 100906.7 | 1040.275 | 1137 | 1186 | 1473 | 1761 | +| 24 | 10 | 21.92007 | 106204 | 1094.887 | 1206 | 1258 | 1533 | 1717 | +| 25 | 10 | 21.6662 | 111925.5 | 1153.871 | 1270 | 1338 | 1598 | 1768 | +| 26 | 10 | 21.88741 | 115226.1 | 1187.898 | 1316 | 1362 | 1498 | 1599 | +| 27 | 10 | 21.81319 | 120065 | 1237.783 | 1369 | 1426 | 1613 | 1760 | +| 28 | 10 | 21.92611 | 123870.6 | 1277.016 | 1409 | 1470 | 1632 | 1784 | +| 29 | 10 | 21.68458 | 129723.5 | 1337.356 | 1471 | 1531 | 1716 | 1972 | +| 30 | 10 | 21.39605 | 136006.4 | 1402.128 | 1528 | 1590 | 1803 | 1965 | +| 31 | 10 | 21.40258 | 140497.1 | 1448.424 | 1563 | 1630 | 1924 | 2040 | +| 32 | 10 | 21.47499 | 144540.3 | 1490.106 | 1617 | 1673 | 1852 | 1977 | +| 33 | 10 | 21.26626 | 150520.2 | 1551.754 | 1668 | 1732 | 2233 | 2928 | +| 34 | 10 | 21.70627 | 151937.6 | 1566.368 | 1705 | 1767 | 2008 | 2269 | +| 35 | 10 | 21.18346 | 160266.5 | 1652.232 | 1781 | 1852 | 2322 | 2536 | +| 36 | 10 | 21.69583 | 160952.6 | 1659.305 | 1780 | 1835 | 1986 | 2125 | +| 37 | 10 | 21.81777 | 164499 | 1695.866 | 1829 | 1886 | 2081 | 2256 | +| 38 | 10 | 20.93849 | 176039.4 | 1814.839 | 1938 | 2021 | 2461 | 2694 | +| 39 | 10 | 21.09936 | 179294.6 | 1848.398 | 1955 | 2019 | 2289 | 2621 | +| 40 | 10 | 21.73072 | 178549.1 | 1840.712 | 1981 | 2050 | 2487 | 2840 | +| 41 | 10 | 21.3417 | 186348.8 | 1921.121 | 2050 | 2110 | 2387 | 2584 | +| 42 | 10 | 21.29802 | 191285.4 | 1972.015 | 2085 | 2153 | 2393 | 2553 | +| 43 | 10 | 21.28315 | 195976.7 | 2020.378 | 2135 | 2211 | 2484 | 2736 | +| 44 | 10 | 21.20328 | 201289.7 | 2075.151 | 2199 | 2269 | 2735 | 2942 | +| 45 | 10 | 21.01285 | 207730 | 2141.547 | 2254 | 2310 | 2482 | 2612 | +| 46 | 10 | 20.90582 | 213433.4 | 2200.344 | 2309 | 2366 | 2511 | 2648 | +| 47 | 10 | 21.20298 | 215017 | 2216.67 | 2361 | 2433 | 2736 | 2983 | +| 48 | 10 | 20.65479 | 225419.9 | 2323.916 | 2430 | 2503 | 3150 | 3381 | +| 49 | 10 | 20.63753 | 230308.5 | 2374.315 | 2490 | 2559 | 2792 | 2958 | +| 50 | 10 | 20.99415 | 231016.7 | 2381.615 | 2501 | 2559 | 2709 | 2856 | +| 51 | 10 | 21.11356 | 234304.4 | 2415.509 | 2536 | 2600 | 2746 | 2853 | +| 52 | 10 | 20.68331 | 243868.1 | 2514.104 | 2645 | 2711 | 2913 | 3077 | +| 53 | 10 | 20.99738 | 244840.1 | 2524.125 | 2668 | 2744 | 3279 | 3478 | +| 54 | 10 | 20.8062 | 251751.9 | 2595.38 | 2721 | 2778 | 2926 | 3073 | +| 55 | 10 | 20.70358 | 257684.9 | 2656.546 | 2779 | 2856 | 3253 | 3628 | +| 56 | 10 | 20.72092 | 262150.6 | 2702.583 | 2821 | 2881 | 3036 | 3206 | +| 57 | 10 | 20.63749 | 267910.5 | 2761.964 | 2884 | 2956 | 3171 | 3374 | +| 58 | 10 | 20.72029 | 271521.3 | 2799.188 | 2927 | 2989 | 3170 | 3329 | +| 59 | 10 | 20.75952 | 275680.7 | 2842.069 | 2982 | 3052 | 3409 | 3645 | +| 60 | 10 | 20.66191 | 281677.7 | 2903.894 | 3038 | 3104 | 3260 | 3389 | +| 61 | 10 | 20.53766 | 288104.8 | 2970.153 | 3120 | 3208 | 3694 | 4031 | +| 62 | 10 | 20.74438 | 289909.9 | 2988.761 | 3140 | 3211 | 3386 | 3541 | +| 63 | 10 | 20.88214 | 292642.4 | 3016.932 | 3162 | 3237 | 3573 | 3948 | +| 64 | 10 | 20.75613 | 299092.4 | 3083.427 | 3223 | 3295 | 3515 | 3783 | +| 65 | 10 | 20.80666 | 303028 | 3124 | 3263 | 3344 | 3548 | 3767 | +| 66 | 10 | 20.61029 | 310621.5 | 3202.284 | 3347 | 3421 | 3759 | 4030 | +| 67 | 10 | 20.75842 | 313077.8 | 3227.607 | 3371 | 3443 | 3621 | 3743 | +| 68 | 10 | 20.78721 | 317310.5 | 3271.243 | 3410 | 3476 | 3659 | 3782 | +| 69 | 10 | 20.91105 | 320070 | 3299.691 | 3444 | 3519 | 3804 | 4088 | +| 70 | 10 | 21.06493 | 322336.8 | 3323.06 | 3461 | 3546 | 3853 | 4124 | +| 71 | 10 | 21.31435 | 323115.6 | 3331.089 | 3490 | 3563 | 3779 | 3965 | +| 72 | 10 | 21.23449 | 328898.9 | 3390.71 | 3540 | 3610 | 3790 | 3953 | +| 73 | 10 | 21.24925 | 333235.3 | 3435.416 | 3604 | 3681 | 3920 | 4102 | +| 74 | 10 | 21.04509 | 341077.2 | 3516.26 | 3679 | 3764 | 4249 | 4533 | +| 75 | 10 | 21.35814 | 340619.6 | 3511.542 | 3692 | 3771 | 4032 | 4263 | +| 76 | 10 | 21.17327 | 348174.9 | 3589.432 | 3742 | 3816 | 4035 | 4234 | +| 77 | 10 | 20.95005 | 356514.7 | 3675.409 | 3831 | 3923 | 4251 | 4521 | +| 78 | 10 | 20.97914 | 360644 | 3717.979 | 3870 | 3956 | 4253 | 4543 | +| 79 | 10 | 21.0908 | 363333.8 | 3745.709 | 3897 | 3976 | 4235 | 4580 | +| 80 | 10 | 21.12234 | 367383.6 | 3787.46 | 3938 | 4016 | 4209 | 4362 | +| 81 | 10 | 21.2174 | 370309.3 | 3817.622 | 4002 | 4079 | 4334 | 4574 | +| 82 | 10 | 20.90807 | 380427.3 | 3921.931 | 4086 | 4170 | 4378 | 4599 | +| 83 | 10 | 20.88443 | 385502.5 | 3974.253 | 4137 | 4227 | 4596 | 4932 | +| 84 | 10 | 21.00666 | 387877 | 3998.732 | 4167 | 4256 | 4480 | 4661 | +| 85 | 10 | 21.09852 | 390785.8 | 4028.719 | 4213 | 4300 | 4538 | 4739 | +| 86 | 10 | 20.93014 | 398563.9 | 4108.906 | 4281 | 4369 | 4614 | 4804 | +| 87 | 10 | 21.2002 | 398062.3 | 4103.735 | 4297 | 4376 | 4591 | 4781 | +| 88 | 10 | 20.92152 | 408001 | 4206.195 | 4362 | 4458 | 4857 | 5096 | +| 89 | 10 | 20.92099 | 412647.8 | 4254.101 | 4425 | 4513 | 4798 | 5019 | +| 90 | 10 | 20.91112 | 417481.3 | 4303.93 | 4484 | 4581 | 4901 | 5135 | +| 91 | 10 | 20.98747 | 420584.3 | 4335.921 | 4505 | 4594 | 4957 | 5358 | +| 92 | 10 | 20.97826 | 425392.7 | 4385.492 | 4557 | 4643 | 4901 | 5122 | +| 93 | 10 | 20.92941 | 431020.2 | 4443.507 | 4634 | 4727 | 5037 | 5358 | +| 94 | 10 | 20.99077 | 434381.4 | 4478.159 | 4648 | 4743 | 5014 | 5238 | +| 95 | 10 | 21.04644 | 437841.3 | 4513.828 | 4702 | 4791 | 5065 | 5303 | +| 96 | 10 | 20.90277 | 445491.1 | 4592.692 | 4773 | 4874 | 5447 | 5794 | + + +响应时间随并发数的变化趋势如下： + +![](gpu-local-time-concurrency.png) + + +结论：随着并发数的增加，平均响应时间和各分位响应时间也随之增加，其中99.9分位和99分位的响应时间有浮动，其它指标基本线性增长。 + + +qps随并发数的变化趋势如下： + +![](gpu-local-qps-concurrency.png) + +结论：随着并发数的增加，qps呈下降趋势，在并发数大于40后，qps相对稳定。 + +从并发数大于97个并发后，开始报错：Cannot malloc 47.7139 MB GPU memory + +## 2.3 单卡多进程 + +如果单个Serving进程不能将GPU卡的算力打满（例如，只有不到50%），那理论上多个Serving进程使用同1张卡，可以提高的GPU的使用率。本测试项就是为了观察和验证这种场景。 + +固定并发数和batch size（例如， 并发数为1，batch size=10），同时启动2个测试进程，都绑定GPU0。分别记录2个进程的qps和响应时间数据。 + + +| 并发数 | 进程id | batchsize | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------|-----------|-----------|---------|----------|----------|----------|----------|------------| +| 1 | 1 | 10 | 13.987022 | 6935 | 71.49484 | 85 | 101 | 127 | 127 | +| 1 | 2 | 10 | 13.610215 | 7127 | 73.47423 | 89 | 103 | 127 | 127 | +| 2 | 1 | 10 | 13.19728 | 14700 | 151.5464 | 176 | 179 | 195 | 197 | +| 2 | 2 | 10 | 13.83392 | 14023.5 | 144.5722 | 174 | 179 | 185 | 186 | + + +与单卡多并发情况相比： + +在并发为1时，总qps增加近50%，平均响应时间增加40%，90分位、99分位、99.9分位响应时间翻倍。 + +在并发为2时，总qps增略有下降，平均响应时间翻倍，90分位、99分位、99.9分位增加近70%。 + + + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|--------|----------|----------|----------|----------|------------| +| 1 | 10 | 17.515349 | 5538.0 | 57.09278 | 57 | 58 | 60 | 60 | +| 2 | 10 | 28.65795 | 6769.5 | 69.78866 | 91 | 102 | 113 | 114 | + + +## 2.4 多卡多并发 + +假设机器上有2张卡，同时启动2个测试进程，分别绑定GPU0和GPU1，固定并发数和batch size （例如，固定batch_size = 10）其他同2.1节 + +系统上安装有多张GPU卡的情况下，理想情况是各张卡各自工作，互不影响；本测试项是为了观察和验证这种场景下的性能表现。 + + +| 并发数 | 进程id | batchsize | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------|-----------|-----------|----------|-----------|----------|----------|----------|------------| +| 1 | 1 | 10 | 17.06545 | 5684 | 58.59794 | 59 | 60 | 62 | 62 | +| 1 | 2 | 10 | 17.15296 | 5655 | 58.29897 | 59 | 59 | 60 | 60 | +| 2 | 1 | 10 | 31.717485 | 6116.5 | 63.0567 | 65 | 76 | 110 | 112 | +| 2 | 2 | 10 | 31.084762 | 6241 | 64.340210 | 70 | 81 | 98 | 99 | +| 3 | 1 | 10 | 29.22763 | 9956.333 | 102.6426 | 132 | 140 | 162 | 172 | +| 3 | 2 | 10 | 27.58468 | 10549.33 | 108.756 | 138 | 148 | 159 | 160 | +| 4 | 1 | 10 | 26.07965 | 14877.5 | 153.3763 | 196 | 206 | 227 | 237 | +| 4 | 2 | 10 | 26.17907 | 14821 | 152.7938 | 195 | 204 | 224 | 228 | + + + +结论：与单卡多并发情况相比，各项性能指标基本持平。 + +| 并发数 | batchsize | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|-----------|-----------|----------|----------|----------|----------|----------|------------| +| 1 | 10 | 17.515349 | 5538.0 | 57.09278 | 57 | 58 | 60 | 60 | +| 2 | 10 | 28.65795 | 6769.5 | 69.78866 | 91 | 102 | 113 | 114 | +| 3 | 10 | 29.04675 | 10018.33 | 103.2818 | 131 | 135 | 159 | 167 | +| 4 | 10 | 26.34661 | 14726.75 | 151.8222 | 183 | 196 | 224 | 227 | + + +# 3. Serving性能测试 + +## 3.1 单卡单并发响应时间与batch size变化关系 + +本测试项是为了测试单个并发的情况下，响应时间随batch size变大而变化的情况： + +测试方法是：在GPU机器上启动1个Serving实例，然后用客户端压测工具以并发数1向Serving发压力，观察平响等指标随batch size增大的变化关系。随着batch size变大，平响会逐渐变大，直到忽然出现大幅增长，这时可认为batch数据用满了GPU kernel。 + +| 并发数 | batchsize | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|-----------|-----------|--------|-------------|----------|----------|----------|------------| +| 1 | 1 | 39.478879 | 2533 | 25.33 | 26 | 28 | 43 | 43 | +| 1 | 2 | 25.733402 | 3886 | 38.86 | 41 | 47 | 57 | 57 | +| 1 | 4 | 16.147263 | 6193 | 61.93 | 64 | 68 | 74 | 74 | +| 1 | 6 | 11.166946 | 8955 | 89.55 | 92 | 97 | 107 | 107 | +| 1 | 8 | 8.759636 | 11416 | 114.16 | 119 | 124 | 132 | 132 | +| 1 | 10 | 7.345919 | 13613 | 136.13 | 141 | 149 | 193 | 193 | +| 1 | 12 | 6.081985 | 16442 | 164.42 | 168 | 177 | 204 | 204 | +| 1 | 14 | 5.334187 | 18747 | 187.47 | 192 | 200 | 233 | 233 | +| 1 | 16 | 4.592633 | 21774 | 217.74 | 223 | 238 | 280 | 280 | +| 1 | 18 | 4.142159 | 24142 | 241.42 | 245 | 252 | 287 | 287 | +| 1 | 20 | 3.782864 | 26435 | 264.35 | 270 | 279 | 308 | 308 | +| 1 | 22 | 3.51062 | 28485 | 284.85 | 293 | 302 | 337 | 337 | +| 1 | 24 | 3.301856 | 30286 | 302.86 | 314 | 326 | 358 | 358 | +| 1 | 26 | 3.06551 | 32621 | 326.21 | 338 | 342 | 485 | 485 | +| 1 | 28 | 2.870182 | 34841 | 348.41 | 360 | 368 | 420 | 420 | +| 1 | 30 | 2.68442 | 37252 | 372.52 | 385 | 399 | 419 | 419 | +| 1 | 32 | 2.525954 | 39589 | 395.89 | 415 | 425 | 450 | 450 | +| 1 | 34 | 2.387091 | 41892 | 418.92 | 434 | 442 | 496 | 496 | +| 1 | 36 | 2.261471 | 44219 | 442.19 | 455 | 486 | 566 | 566 | +| 1 | 38 | 2.181453 | 45841 | 458.41 | 467 | 485 | 588 | 588 | +| 1 | 40 | 2.093276 | 47772 | 477.72 | 484 | 496 | 595 | 595 | +| 1 | 42 | 1.99005 | 50250 | 502.5 | 514 | 524 | 615 | 615 | +| 1 | 44 | 1.901538 | 52589 | 525.89 | 532 | 553 | 609 | 609 | +| 1 | 46 | 1.805087 | 55399 | 553.99 | 569 | 584 | 673 | 673 | +| 1 | 48 | 1.623034 | 61613 | 616.13 | 631 | 641 | 676 | 676 | +| 1 | 50 | 1.549355 | 64543 | 645.43 | 661 | 674 | 733 | 733 | +| 1 | 52 | 1.496491 | 66823 | 668.23 | 682 | 698 | 795 | 795 | +| 1 | 54 | 1.425537 | 70149 | 701.49 | 714 | 747 | 859 | 859 | +| 1 | 56 | 1.386059 | 72147 | 721.47 | 733 | 747 | 822 | 822 | +| 1 | 58 | 1.33672 | 74810 | 748.1 | 761 | 777 | 859 | 859 | +| 1 | 60 | 1.291673 | 77419 | 774.19 | 792 | 801 | 911 | 911 | +| 1 | 62 | 1.251361 | 79913 | 799.13 | 813 | 833 | 978 | 978 | +| 1 | 64 | 1.206607 | 82877 | 828.77 | 845 | 871 | 1034 | 1034 | +| 1 | 66 | 1.173213 | 85236 | 852.36 | 875 | 894 | 969 | 969 | +| 1 | 68 | 1.140914 | 87649 | 876.49 | 896 | 917 | 974 | 974 | +| 1 | 70 | 1.101576 | 90779 | 907.79 | 932 | 971 | 1042 | 1042 | +| 1 | 72 | 1.077284 | 92826 | 928.26 | 948 | 962 | 1078 | 1078 | +| 1 | 74 | 1.046616 | 95546 | 955.46 | 979 | 998 | 1076 | 1076 | +| 1 | 76 | 1.019368 | 98100 | 981 | 1011 | 1028 | 1063 | 1063 | +| 1 | 78 | 0.995877 | 100414 | 1004.14 | 1027 | 1049 | 1142 | 1142 | +| 1 | 80 | 0.970327 | 103058 | 1030.58 | 1046 | 1075 | 1232 | 1232 | +| 1 | 82 | 0.949082 | 105365 | 1053.65 | 1080 | 1096 | 1188 | 1188 | +| 1 | 84 | 0.917339 | 109011 | 1090.11 | 1123 | 1160 | 1239 | 1239 | +| 1 | 86 | 0.888092 | 112601 | 1126.01 | 1153 | 1189 | 1237 | 1237 | +| 1 | 88 | 0.8726 | 114600 | 1146 | 1172 | 1193 | 1236 | 1236 | +| 1 | 90 | 0.838828 | 119214 | 1192.14 | 1202 | 1241 | 2258 | 2258 | +| 1 | 92 | 0.835122 | 119743 | 1197.43 | 1223 | 1244 | 1296 | 1296 | +| 1 | 94 | 0.816 | 122549 | 1225.49 | 1259 | 1282 | 1395 | 1395 | +| 1 | 96 | 0.794597 | 125850 | 1258.5 | 1295 | 1310 | 1477 | 1477 | +| 1 | 98 | 0.771575 | 129605 | 1296.05 | 1333 | 1365 | 1542 | 1542 | +| 1 | 100 | 0.76609 | 130533 | 1305.33 | 1336 | 1364 | 1428 | 1428 | +| 1 | 102 | 0.750103 | 133315 | 1333.15 | 1361 | 1398 | 1504 | 1504 | +| 1 | 104 | 0.736914 | 135701 | 1357.01 | 1396 | 1428 | 1484 | 1484 | +| 1 | 106 | 0.716009 | 139663 | 1396.63 | 1431 | 1483 | 1740 | 1740 | +| 1 | 108 | 0.70821 | 141201 | 1412.01 | 1444 | 1482 | 1637 | 1637 | +| 1 | 110 | 0.697015 | 143469 | 1434.69 | 1466 | 1493 | 1582 | 1582 | +| 1 | 112 | 0.679939 | 147072 | 1470.72 | 1510 | 1546 | 1668 | 1668 | +| 1 | 114 | 0.672233 | 148758 | 1487.58 | 1522 | 1544 | 1700 | 1700 | +| 1 | 116 | 0.651025 | 153604 | 1536.04 | 1588 | 1616 | 2524 | 2524 | +| 1 | 118 | 0.649401 | 153988 | 1539.88 | 1585 | 1598 | 1699 | 1699 | +| 1 | 120 | 0.639865 | 156283 | 1562.83 | 1603 | 1636 | 1784 | 1784 | +| 1 | 122 | 0.619725 | 161362 | 1613.62 | 1658 | 1690 | 2672 | 2672 | +| 1 | 124 | 0.610434 | 163818 | 1638.18 | 1660 | 1709 | 2942 | 2942 | +| 1 | 126 | 0.594015 | 168346 | 1683.46 | 1705 | 1741 | 3146 | 3146 | +| 1 | 128 | 0.58924 | 169710 | 1697.1 | 1725 | 1780 | 3013 | 3013 | +| 1 | 130 | 0.580801 | 172176 | 1721.76 | 1751 | 1798 | 3226 | 3226 | +| 1 | 132 | 0.575275 | 173830 | 1738.3 | 1779 | 1844 | 2168 | 2168 | +| 1 | 134 | 0.56464 | 177104 | 1771.04 | 1824 | 1920 | 2012 | 2012 | +| 1 | 136 | 0.55635 | 179743 | 1797.43 | 1843 | 1902 | 2272 | 2272 | +| 1 | 138 | 0.534536 | 187078 | 1870.78 | 1887 | 2040 | 3379 | 3379 | +| 1 | 140 | 0.5446 | 183621 | 1836.21 | 1882 | 1916 | 2229 | 2229 | +| 1 | 142 | 0.536121 | 186525 | 1865.25 | 1926 | 1965 | 2127 | 2127 | +| 1 | 144 | 0.530763 | 188408 | 1884.08 | 1937 | 1998 | 2082 | 2082 | +| 1 | 146 | 0.519983 | 192314 | 1923.14 | 1980 | 2052 | 2356 | 2356 | +| 1 | 148 | 0.510436 | 195911 | 1959.11 | 2019 | 2063 | 2738 | 2738 | +| 1 | 150 | 0.51067 | 195821 | 1958.21 | 2004 | 2047 | 2194 | 2194 | +| 1 | 152 | 0.499513 | 200195 | 2001.95 | 2038 | 2084 | 3105 | 3105 | +| 1 | 154 | 0.492126 | 203200 | 2032 | 2093 | 2180 | 2424 | 2424 | +| 1 | 156 | 0.48741 | 205166 | 2051.66 | 2114 | 2168 | 2424 | 2424 | +| 1 | 158 | 0.480137 | 208274 | 2082.74 | 2141 | 2185 | 2597 | 2597 | +| 1 | 160 | 0.478343 | 209055 | 2090.55 | 2150 | 2190 | 2338 | 2338 | +| 1 | 162 | 0.468615 | 213395 | 2133.95 | 2200 | 2261 | 2474 | 2474 | +| 1 | 164 | 0.466083 | 214554 | 2145.54 | 2211 | 2258 | 2615 | 2615 | +| 1 | 166 | 0.455612 | 219485 | 2194.85 | 2211 | 2254 | 4183 | 4183 | +| 1 | 168 | 0.450442 | 222004 | 2220.04 | 2279 | 2321 | 2877 | 2877 | +| 1 | 170 | 0.448893 | 222770 | 2227.7 | 2293 | 2339 | 2486 | 2486 | +| 1 | 172 | 0.433392 | 230738 | 2307.38 | 2368 | 2418 | 3391 | 3391 | +| 1 | 174 | 0.426087 | 234694 | 2346.94 | 2420 | 2510 | 3180 | 3180 | +| 1 | 176 | 0.427403 | 233971 | 2339.71 | 2403 | 2464 | 3321 | 3321 | +| 1 | 178 | 0.417155 | 239719 | 2397.19 | 2468 | 2507 | 4729 | 4729 | +| 1 | 180 | 0.417765 | 239369 | 2393.69 | 2455 | 2507 | 2827 | 2827 | +| 1 | 182 | 0.406866 | 245781 | 2457.81 | 2521 | 2623 | 4419 | 4419 | +| 1 | 184 | 0.396986 | 251898 | 2518.98 | 2517 | 2606 | 5000 | 5000 | +| 1 | 186 | 0.402329 | 248553 | 2485.53 | 2524 | 2556 | 4617 | 4617 | +| 1 | 188 | 0.394987 | 253173 | 2531.73 | 2567 | 2642 | 4964 | 4964 | +| 1 | 190 | 0.389011 | 257062 | 2570.62 | 2601 | 2684 | 4769 | 4769 | +| 1 | 192 | 0.394163 | 253702 | 2537.02 | 2583 | 2634 | 2781 | 2781 | +| 1 | 194 | 0.379815 | 263286 | 2632.86 | 2692 | 2775 | 5061 | 5061 | +| 1 | 196 | 0.377895 | 264624 | 2646.24 | 2709 | 2768 | 4873 | 4873 | +| 1 | 198 | 0.496347 | 201472 | 2616.519481 | 2703 | 2728 | 2785 | 2785 | + + +qps随batch size的变化趋势如下： + +![](gpu-serving-single-card-qps-batchsize.png) + +结论：在batch size小于200的前提下，随着batch size的增加，qps随之下降，并且qps与单机情况相比有大幅下降。 + + +![](gpu-serving-single-card-time-batchsize.png) + + +结论：随着batch size的增加，平均响应时间和各分位响应时间也随之增加，其中99.9分位的响应时间浮动较大，其它指标基本线性增长，各个响应时间指标较单机情况大幅增加。 + +## 3.2 单卡多并发 + +固定batch size（例如，可以固定为batch size= 10。以下所有场景中所有固定batch size的均以此为准），启动单个Serving进程，绑定GPU0；启动压力测试工具，以concurrency = 1, 2, ...向该Serving进程发压力，观察平响等指标随并发数增加的变化情况。 + +如本章开头所述，目前Paddle预测库GPU预测只用1个stream，并行的请求会被串行计算。本测试项是为了观察和验证这种场景下的性能表现。预期随着并发数增加，平响和99分位响应时间会急剧变大，但总qps略有上涨后保持稳定或开始下降（并发请求数将stream占满后，每秒内处理能力达到极限）。 + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|--------|-------------|----------|----------|----------|------------| +| 1 | 10 | 7.48559 | 13359 | 133.59 | 138 | 142 | 167 | 167 | +| 2 | 10 | 13.952839 | 14334 | 143.34 | 148 | 160 | 207 | 208 | +| 3 | 10 | 19.843895 | 15118 | 151.183333 | 149 | 177 | 249 | 261 | +| 4 | 10 | 16.42643 | 24351 | 243.51 | 283 | 292 | 324 | 327 | +| 5 | 10 | 17.525412 | 28530 | 285.302 | 330 | 337 | 361 | 377 | +| 6 | 10 | 18.325087 | 32742 | 327.42 | 381 | 390 | 423 | 443 | +| 7 | 10 | 19.130387 | 36591 | 365.918571 | 426 | 440 | 478 | 527 | +| 8 | 10 | 19.28129 | 41491 | 414.91125 | 472 | 487 | 519 | 539 | +| 9 | 10 | 19.857906 | 45322 | 453.227778 | 523 | 540 | 630 | 669 | +| 10 | 10 | 20.079918 | 49801 | 498.017 | 541 | 558 | 681 | 843 | +| 11 | 10 | 20.460929 | 53761 | 537.618182 | 579 | 672 | 885 | 988 | +| 12 | 10 | 20.565553 | 58350 | 583.505 | 667 | 749 | 943 | 998 | +| 13 | 10 | 20.541668 | 63286 | 632.863846 | 751 | 850 | 997 | 1057 | +| 14 | 10 | 20.426029 | 68540 | 685.400714 | 830 | 927 | 1029 | 1127 | +| 15 | 10 | 20.337328 | 73756 | 737.566667 | 934 | 974 | 1038 | 1154 | +| 16 | 10 | 20.386061 | 78485 | 784.858125 | 930 | 983 | 1044 | 1115 | +| 17 | 10 | 20.407428 | 83303 | 833.036471 | 964 | 995 | 1044 | 1088 | +| 18 | 10 | 20.28169 | 88750 | 887.503333 | 988 | 1011 | 1062 | 1118 | +| 19 | 10 | 20.26645 | 93751 | 937.511579 | 1009 | 1034 | 1180 | 1370 | +| 20 | 10 | 20.164136 | 99186 | 991.8675 | 1038 | 1086 | 1379 | 1514 | +| 21 | 10 | 20.25502 | 103678 | 1036.78 | 1087 | 1183 | 1464 | 1512 | +| 22 | 10 | 20.340801 | 108157 | 1081.574091 | 1187 | 1302 | 1485 | 1553 | +| 23 | 10 | 20.421028 | 112629 | 1126.297826 | 1266 | 1366 | 1505 | 1555 | +| 24 | 10 | 20.309208 | 118173 | 1181.737083 | 1358 | 1445 | 1529 | 1605 | +| 25 | 10 | 20.241276 | 123510 | 1235.104 | 1426 | 1481 | 1555 | 1633 | +| 26 | 10 | 20.213014 | 128630 | 1286.305 | 1468 | 1501 | 1560 | 1615 | +| 27 | 10 | 20.372131 | 132534 | 1325.342222 | 1464 | 1501 | 1572 | 1645 | +| 28 | 10 | 20.264891 | 138170 | 1381.702143 | 1497 | 1522 | 1599 | 1718 | +| 29 | 10 | 20.065733 | 144525 | 1445.254828 | 1531 | 1559 | 1695 | 1901 | +| 30 | 10 | 20.220675 | 148363 | 1483.635 | 1549 | 1601 | 1856 | 1958 | +| 31 | 10 | 20.2206 | 153309 | 1533.098387 | 1604 | 1687 | 1950 | 2006 | +| 32 | 10 | 20.255088 | 157985 | 1579.854063 | 1690 | 1807 | 1999 | 2094 | +| 33 | 10 | 20.111037 | 164089 | 1640.890909 | 1813 | 1939 | 2039 | 2094 | +| 34 | 10 | 20.289059 | 167578 | 1675.781765 | 1854 | 1949 | 2029 | 2072 | +| 35 | 10 | 20.167099 | 173550 | 1735.508857 | 1933 | 1987 | 2070 | 2122 | +| 36 | 10 | 20.371093 | 176721 | 1767.214444 | 1902 | 1977 | 2084 | 2130 | +| 37 | 10 | 20.316498 | 182118 | 1821.182162 | 1968 | 2002 | 2082 | 2130 | +| 38 | 10 | 20.108161 | 188978 | 1889.780789 | 2011 | 2055 | 2195 | 2349 | +| 39 | 10 | 20.355968 | 191590 | 1915.908462 | 2011 | 2049 | 2174 | 2384 | +| 40 | 10 | 20.266299 | 197372 | 1973.72725 | 2055 | 2105 | 2394 | 2502 | +| 41 | 10 | 20.187299 | 203098 | 2030.984146 | 2114 | 2211 | 2454 | 2513 | +| 42 | 10 | 20.180083 | 208126 | 2081.263571 | 2190 | 2328 | 2541 | 2621 | +| 43 | 10 | 20.171505 | 213172 | 2131.720465 | 2289 | 2410 | 2548 | 2611 | +| 44 | 10 | 20.145044 | 218416 | 2184.162955 | 2370 | 2470 | 2576 | 2655 | +| 45 | 10 | 20.128284 | 223566 | 2235.666 | 2421 | 2490 | 2566 | 2601 | +| 46 | 10 | 20.34966 | 226048 | 2260.487826 | 2427 | 2483 | 2576 | 3033 | +| 47 | 10 | 20.265085 | 231926 | 2319.266596 | 2471 | 2520 | 2628 | 3037 | +| 48 | 10 | 20.300103 | 236452 | 2364.526667 | 2497 | 2538 | 2634 | 3005 | +| 49 | 10 | 20.186124 | 242741 | 2427.414898 | 2543 | 2592 | 2748 | 2968 | +| 50 | 10 | 20.196145 | 247572 | 2475.7218 | 2562 | 2619 | 2850 | 2994 | +| 51 | 10 | 20.173412 | 252808 | 2528.083922 | 2630 | 2721 | 2974 | 3088 | +| 52 | 10 | 20.188843 | 257568 | 2575.686923 | 2715 | 2844 | 3033 | 3441 | +| 53 | 10 | 20.351271 | 260426 | 2604.267547 | 2728 | 2854 | 3026 | 3493 | +| 54 | 10 | 20.135204 | 268187 | 2681.871296 | 2879 | 2972 | 3082 | 3137 | +| 55 | 10 | 20.25917 | 271482 | 2714.824182 | 2885 | 2971 | 3121 | 3234 | +| 56 | 10 | 20.141784 | 278029 | 2780.295179 | 2968 | 3017 | 3096 | 3167 | +| 57 | 10 | 20.150386 | 282873 | 2828.730351 | 2989 | 3049 | 3163 | 3291 | +| 58 | 10 | 20.092982 | 288658 | 2886.581552 | 3022 | 3067 | 3193 | 3341 | +| 59 | 10 | 20.252296 | 291325 | 2913.255085 | 3027 | 3066 | 3236 | 3461 | +| 60 | 10 | 20.250157 | 296294 | 2962.946667 | 3073 | 3124 | 3350 | 3521 | +| 61 | 10 | 20.2362 | 301440 | 3014.409344 | 3115 | 3208 | 3475 | 3562 | +| 62 | 10 | 20.34074 | 304807 | 3048.075 | 3164 | 3249 | 3496 | 3637 | +| 63 | 10 | 20.111668 | 313251 | 3132.516508 | 3300 | 3423 | 3568 | 3635 | +| 64 | 10 | 20.073897 | 318822 | 3188.223125 | 3387 | 3490 | 3602 | 3773 | +| 65 | 10 | 20.182073 | 322068 | 3220.688154 | 3415 | 3493 | 3615 | 3675 | +| 66 | 10 | 20.152487 | 327503 | 3275.038333 | 3461 | 3517 | 3652 | 4039 | +| 67 | 10 | 20.201777 | 331654 | 3316.54403 | 3477 | 3534 | 3671 | 3741 | +| 68 | 10 | 20.172475 | 337093 | 3370.938824 | 3512 | 3566 | 3705 | 3834 | +| 69 | 10 | 20.121487 | 342917 | 3429.171159 | 3564 | 3619 | 3860 | 4056 | +| 70 | 10 | 20.212228 | 346325 | 3463.258714 | 3567 | 3635 | 3907 | 4078 | +| 71 | 10 | 20.316537 | 349469 | 3494.692535 | 3618 | 3702 | 3953 | 4447 | +| 72 | 10 | 20.244395 | 355654 | 3556.545556 | 3683 | 3779 | 4000 | 4096 | +| 73 | 10 | 20.38514 | 358104 | 3581.045068 | 3708 | 3799 | 4024 | 4171 | +| 74 | 10 | 20.234612 | 365710 | 3657.103378 | 3823 | 3925 | 4076 | 4199 | +| 75 | 10 | 20.12526 | 372666 | 3726.664133 | 3936 | 4010 | 4162 | 4701 | +| 76 | 10 | 20.187692 | 376467 | 3764.679605 | 3945 | 4015 | 4142 | 4253 | +| 77 | 10 | 20.114365 | 382811 | 3828.118182 | 4005 | 4056 | 4167 | 4400 | +| 78 | 10 | 20.213957 | 385872 | 3858.725128 | 4016 | 4069 | 4224 | 4485 | +| 79 | 10 | 20.029207 | 394424 | 3944.242532 | 4081 | 4142 | 4445 | 4582 | +| 80 | 10 | 20.239586 | 395265 | 3952.658 | 4076 | 4148 | 4422 | 4552 | +| 81 | 10 | 20.141438 | 402156 | 4021.566049 | 4144 | 4255 | 4521 | 4608 | +| 82 | 10 | 20.049096 | 408996 | 4089.960854 | 4238 | 4372 | 4607 | 4752 | +| 83 | 10 | 20.136491 | 412187 | 4121.876867 | 4284 | 4395 | 4612 | 4721 | +| 84 | 10 | 20.142629 | 417026 | 4170.2625 | 4365 | 4473 | 4639 | 4771 | +| 85 | 10 | 20.307964 | 418555 | 4185.555765 | 4347 | 4427 | 4588 | 4648 | +| 86 | 10 | 20.176947 | 426229 | 4262.294884 | 4454 | 4532 | 4708 | 4768 | +| 87 | 10 | 20.174942 | 431228 | 4312.282759 | 4491 | 4551 | 4700 | 4874 | +| 88 | 10 | 20.161706 | 436471 | 4364.713409 | 4523 | 4580 | 4708 | 4820 | +| 89 | 10 | 20.200645 | 440580 | 4405.808876 | 4550 | 4610 | 4800 | 4963 | +| 90 | 10 | 20.230811 | 444866 | 4448.661 | 4587 | 4658 | 4906 | 5096 | +| 91 | 10 | 20.209243 | 450289 | 4502.895714 | 4645 | 4733 | 4957 | 5071 | +| 92 | 10 | 20.284691 | 453544 | 4535.442826 | 4670 | 4769 | 4975 | 5095 | +| 93 | 10 | 20.212294 | 460116 | 4601.16129 | 4772 | 4890 | 5075 | 5180 | +| 94 | 10 | 20.153639 | 466417 | 4664.177766 | 4858 | 4965 | 5136 | 5332 | +| 95 | 10 | 20.196866 | 470370 | 4703.701158 | 4893 | 4988 | 5178 | 5293 | +| 96 | 10 | 20.27545 | 473479 | 4734.793438 | 4900 | 5001 | 5200 | 5272 | +| 97 | 10 | 20.22852 | 479521 | 4795.211959 | 4970 | 5034 | 5176 | 5393 | +| 98 | 10 | 20.275331 | 483346 | 4833.468061 | 4987 | 5046 | 5253 | 5751 | +| 99 | 10 | 20.198682 | 490131 | 4901.316364 | 5057 | 5127 | 5331 | 5528 | +| 100 | 10 | 20.139608 | 496534 | 4965.3454 | 5111 | 5187 | 5391 | 5487 | +| 101 | 10 | 20.236101 | 499108 | 4991.083465 | 5122 | 5205 | 5491 | 5576 | +| 102 | 10 | 20.1116 | 507170 | 5071.703922 | 5227 | 5339 | 5616 | 5785 | +| 103 | 10 | 20.19889 | 509929 | 5099.292621 | 5253 | 5362 | 5628 | 5742 | +| 104 | 10 | 20.205082 | 514722 | 5147.221538 | 5338 | 5434 | 5602 | 5715 | +| 105 | 10 | 20.178183 | 520364 | 5203.646476 | 5388 | 5492 | 5667 | 5767 | +| 106 | 10 | 20.171611 | 525491 | 5254.914245 | 5441 | 5519 | 5673 | 5785 | +| 107 | 10 | 20.175128 | 530356 | 5303.56486 | 5485 | 5561 | 5731 | 5965 | +| 108 | 10 | 20.188765 | 534951 | 5349.517685 | 5522 | 5591 | 5792 | 6052 | +| 109 | 10 | 20.232432 | 538739 | 5387.395872 | 5540 | 5615 | 5799 | 5998 | +| 110 | 10 | 20.151835 | 545856 | 5458.563182 | 5602 | 5700 | 6072 | 6166 | +| 111 | 10 | 20.223147 | 548876 | 5488.760721 | 5639 | 5721 | 6001 | 6089 | +| 112 | 10 | 20.140263 | 556100 | 5561.009554 | 5736 | 5848 | 6093 | 6264 | +| 113 | 10 | 20.145799 | 560911 | 5609.11115 | 5790 | 5909 | 6164 | 6377 | +| 114 | 10 | 20.184636 | 564786 | 5647.865439 | 5820 | 5930 | 6207 | 6283 | +| 115 | 10 | 20.176288 | 569976 | 5699.765391 | 5893 | 5984 | 6153 | 6262 | +| 116 | 10 | 20.186271 | 574648 | 5746.488017 | 5940 | 6036 | 6225 | 6424 | +| 117 | 10 | 20.182992 | 579696 | 5796.960769 | 5977 | 6056 | 6262 | 6410 | +| 118 | 10 | 20.152992 | 585521 | 5855.219576 | 6025 | 6109 | 6302 | 6527 | +| 119 | 10 | 20.193861 | 589288 | 5892.880168 | 6062 | 6126 | 6322 | 6534 | +| 120 | 10 | 20.192366 | 594284 | 5942.845917 | 6104 | 6194 | 6420 | 6533 | +| 121 | 10 | 20.184227 | 599478 | 5994.788182 | 6167 | 6259 | 6547 | 6693 | +| 122 | 10 | 20.165256 | 605001 | 6050.019836 | 6206 | 6304 | 6553 | 6636 | +| 123 | 10 | 20.20202 | 608850 | 6088.503008 | 6264 | 6367 | 6584 | 6727 | +| 124 | 10 | 20.123141 | 616206 | 6162.064435 | 6357 | 6468 | 6687 | 6820 | +| 125 | 10 | 20.110625 | 621562 | 6215.62888 | 6433 | 6518 | 6656 | 6742 | +| 126 | 10 | 20.142082 | 625556 | 6255.560476 | 6446 | 6535 | 6757 | 6845 | +| 127 | 10 | 20.162026 | 629897 | 6298.97378 | 6497 | 6591 | 6785 | 6888 | +| 128 | 10 | 20.106217 | 636619 | 6366.198594 | 6557 | 6626 | 6804 | 7097 | + +![](gpu-serving-single-card-qps-concurrency.png) + +结论：在并发数小于10时，随着并发数的增加，qps也随之增加，大于10并发后，qps基本不变，与单机情况持平。 + +![](gpu-serving-single-card-time-concurrency.png) + +结论：随着并发数的增加，平均响应时间和各分位响应时间也随之增加，与单机相比基本持平，在并发数大于96时，没有发生显存不足的问题。 + +## 3.3 单卡多进程 + +如果单个Serving进程不能将GPU卡的算力打满（例如，只有不到50%），那理论上多个Serving进程使用同1张卡，可以提高的GPU的使用率。本测试项就是为了观察和验证这种场景。 + +固定并发数和batch size（例如， 并发数为1，batch size=10），同时启动2个Serving进程，他们都绑定GPU0。分别启动2个压测工具实例，分别向这个进程以一定并发数发压力，记录2个进程的qps和响应时间数据。 + +| 并发数 | 进程id | batch size | qps | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------|------------|----------|--------|----------|----------|----------|------------| +| 1 | 1 | 10 | 5.499945 | 181.82 | 198 | 202 | 315 | 315 | +| 1 | 2 | 10 | 5.465974 | 182.95 | 198 | 203 | 361 | 361 | +| 2 | 1 | 10 | 9.639937 | 207.47 | 208 | 214 | 314 | 314 | +| 2 | 2 | 10 | 9.662303 | 206.99 | 209 | 212 | 282 | 311 | + + +结论：与单机多并发情况相比较，单进程qps下降，总qps提升，平均和各个分位的响应时间都增加。 + +该测试项的测试数据，应该与3.2节中数据做对比。 + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|-------|--------|----------|----------|----------|------------| +| 1 | 10 | 7.48559 | 13359 | 133.59 | 138 | 142 | 167 | 167 | +| 2 | 10 | 13.952839 | 14334 | 143.34 | 148 | 160 | 207 | 208 | + + +## 3.4 多卡单并发响应时间与batch_size的变化关系 + +由于PaddlePaddle预测库只能做单卡预测，间接使得Paddle Serving实例也只能绑定1张卡；在多卡机器上，可以启动多个Serving实例，各自绑定1张GPU卡，并行提供预测服务。本测试项是为了观察与验证，在启动多个Serving实例，每个实例绑定1张独立的GPU卡的情况下，单并发响应时间与batch_size的变化关系。 + +启动2个Serving实例，分别绑定GPU0和GPU1；再启动2个client端，分别向2个Serving进程以单并发发送请求，不断增大batch_size。 + +client 1： + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|--------|---------|----------|----------|----------|------------| +| 1 | 2 | 27.419797 | 3647 | 36.47 | 37 | 39 | 64 | 64 | +| 1 | 4 | 16.509823 | 6057 | 60.57 | 62 | 65 | 85 | 85 | +| 1 | 6 | 11.837121 | 8448 | 84.48 | 86 | 89 | 144 | 144 | +| 1 | 8 | 9.099181 | 10990 | 109.9 | 113 | 117 | 204 | 204 | +| 1 | 10 | 7.753741 | 12897 | 128.97 | 132 | 139 | 173 | 173 | +| 1 | 12 | 6.393044 | 15642 | 156.42 | 160 | 164 | 182 | 182 | +| 1 | 14 | 5.747457 | 17399 | 173.99 | 178 | 180 | 234 | 234 | +| 1 | 16 | 4.870683 | 20531 | 205.31 | 210 | 220 | 278 | 278 | +| 1 | 18 | 4.398698 | 22734 | 227.34 | 232 | 239 | 285 | 285 | +| 1 | 20 | 4.037794 | 24766 | 247.66 | 255 | 261 | 291 | 291 | +| 1 | 22 | 3.659251 | 27328 | 273.28 | 279 | 284 | 353 | 353 | +| 1 | 24 | 3.336559 | 29971 | 299.71 | 307 | 320 | 509 | 509 | +| 1 | 26 | 3.098469 | 32274 | 322.74 | 331 | 340 | 423 | 423 | +| 1 | 28 | 2.837362 | 35244 | 352.44 | 368 | 379 | 419 | 419 | +| 1 | 30 | 2.674655 | 37388 | 373.88 | 388 | 399 | 533 | 533 | +| 1 | 32 | 2.554539 | 39146 | 391.46 | 401 | 406 | 443 | 443 | +| 1 | 34 | 2.338525 | 42762 | 427.62 | 436 | 458 | 661 | 661 | +| 1 | 36 | 2.205899 | 45333 | 453.33 | 465 | 483 | 727 | 727 | +| 1 | 38 | 2.116223 | 47254 | 472.54 | 490 | 500 | 571 | 571 | +| 1 | 40 | 2.029921 | 49263 | 492.63 | 508 | 524 | 716 | 716 | +| 1 | 42 | 1.918208 | 52132 | 521.32 | 531 | 553 | 750 | 750 | +| 1 | 44 | 1.82156 | 54898 | 548.98 | 568 | 588 | 900 | 900 | +| 1 | 46 | 1.701143 | 58784 | 587.84 | 597 | 612 | 764 | 764 | +| 1 | 48 | 1.655711 | 60397 | 603.97 | 610 | 629 | 867 | 867 | +| 1 | 50 | 1.550171 | 64509 | 645.09 | 659 | 679 | 924 | 924 | +| 1 | 52 | 1.529637 | 65375 | 653.75 | 673 | 691 | 792 | 792 | +| 1 | 54 | 1.461198 | 68437 | 684.37 | 700 | 721 | 894 | 894 | +| 1 | 56 | 1.385349 | 72184 | 721.84 | 741 | 772 | 892 | 892 | +| 1 | 58 | 1.333671 | 74981 | 749.81 | 779 | 798 | 1053 | 1053 | +| 1 | 60 | 1.293393 | 77316 | 773.16 | 790 | 806 | 1049 | 1049 | +| 1 | 62 | 1.247676 | 80149 | 801.49 | 826 | 847 | 1003 | 1003 | +| 1 | 64 | 1.215067 | 82300 | 823 | 849 | 886 | 984 | 984 | +| 1 | 66 | 1.175392 | 85078 | 850.78 | 874 | 891 | 1036 | 1036 | +| 1 | 68 | 1.129854 | 88507 | 885.07 | 911 | 947 | 1105 | 1105 | +| 1 | 70 | 1.0915 | 91617 | 916.17 | 942 | 972 | 1099 | 1099 | +| 1 | 72 | 1.079389 | 92645 | 926.45 | 957 | 990 | 1059 | 1059 | +| 1 | 74 | 1.050431 | 95199 | 951.99 | 986 | 1014 | 1154 | 1154 | +| 1 | 76 | 1.012556 | 98760 | 987.6 | 1021 | 1081 | 1244 | 1244 | +| 1 | 78 | 0.980825 | 101955 | 1019.55 | 1049 | 1097 | 1459 | 1459 | +| 1 | 80 | 0.967754 | 103332 | 1033.32 | 1060 | 1092 | 1252 | 1252 | +| 1 | 82 | 0.944412 | 105886 | 1058.86 | 1084 | 1134 | 1528 | 1528 | +| 1 | 84 | 0.923651 | 108266 | 1082.66 | 1115 | 1142 | 1494 | 1494 | +| 1 | 86 | 0.89844 | 111304 | 1113.04 | 1143 | 1173 | 1314 | 1314 | +| 1 | 88 | 0.874195 | 114391 | 1143.91 | 1183 | 1213 | 1432 | 1432 | +| 1 | 90 | 0.857552 | 116611 | 1166.11 | 1202 | 1242 | 1338 | 1338 | +| 1 | 92 | 0.809894 | 123473 | 1234.73 | 1246 | 1283 | 1678 | 1678 | +| 1 | 94 | 0.798441 | 125244 | 1252.44 | 1274 | 1303 | 1964 | 1964 | +| 1 | 96 | 0.781409 | 127974 | 1279.74 | 1299 | 1329 | 1531 | 1531 | +| 1 | 98 | 0.761771 | 131273 | 1312.73 | 1334 | 1363 | 1583 | 1583 | +| 1 | 100 | 0.745107 | 134209 | 1342.09 | 1374 | 1402 | 1653 | 1653 | +| 1 | 102 | 0.730268 | 136936 | 1369.36 | 1390 | 1420 | 1893 | 1893 | +| 1 | 104 | 0.718649 | 139150 | 1391.5 | 1427 | 1462 | 1592 | 1592 | +| 1 | 106 | 0.707899 | 141263 | 1412.63 | 1432 | 1481 | 2048 | 2048 | +| 1 | 108 | 0.698753 | 143112 | 1431.12 | 1459 | 1492 | 1673 | 1673 | +| 1 | 110 | 0.66929 | 149412 | 1494.12 | 1514 | 1607 | 2330 | 2330 | +| 1 | 112 | 0.666858 | 149957 | 1499.57 | 1539 | 1569 | 1706 | 1706 | +| 1 | 114 | 0.650889 | 153636 | 1536.36 | 1580 | 1618 | 1795 | 1795 | +| 1 | 116 | 0.644883 | 155067 | 1550.67 | 1595 | 1635 | 1799 | 1799 | +| 1 | 118 | 0.626378 | 159648 | 1596.48 | 1637 | 1676 | 1832 | 1832 | +| 1 | 120 | 0.617166 | 162031 | 1620.31 | 1679 | 1747 | 2233 | 2233 | +| 1 | 122 | 0.606516 | 164876 | 1648.76 | 1693 | 1722 | 2266 | 2266 | +| 1 | 124 | 0.610452 | 163813 | 1638.13 | 1670 | 1702 | 1806 | 1806 | +| 1 | 126 | 0.585943 | 170665 | 1706.65 | 1753 | 1819 | 2163 | 2163 | +| 1 | 128 | 0.578921 | 172735 | 1727.35 | 1766 | 1837 | 2062 | 2062 | +| 1 | 130 | 0.567424 | 176235 | 1762.35 | 1801 | 1853 | 2147 | 2147 | +| 1 | 132 | 0.56117 | 178199 | 1781.99 | 1804 | 1902 | 2648 | 2648 | +| 1 | 134 | 0.5539 | 180538 | 1805.38 | 1832 | 1916 | 2937 | 2937 | +| 1 | 136 | 0.544642 | 183607 | 1836.07 | 1910 | 1986 | 2210 | 2210 | +| 1 | 138 | 0.537912 | 185904 | 1859.04 | 1903 | 1946 | 2316 | 2316 | +| 1 | 140 | 0.52977 | 188761 | 1887.61 | 1914 | 2001 | 2617 | 2617 | +| 1 | 142 | 0.523747 | 190932 | 1909.32 | 1959 | 2006 | 2231 | 2231 | +| 1 | 144 | 0.515634 | 193936 | 1939.36 | 1962 | 2084 | 2457 | 2457 | +| 1 | 146 | 0.509507 | 196268 | 1962.68 | 2004 | 2063 | 2265 | 2265 | +| 1 | 148 | 0.502841 | 198870 | 1988.7 | 2049 | 2144 | 2413 | 2413 | +| 1 | 150 | 0.502134 | 199150 | 1991.5 | 2037 | 2086 | 2281 | 2281 | +| 1 | 152 | 0.491792 | 203338 | 2033.38 | 2075 | 2094 | 2700 | 2700 | +| 1 | 154 | 0.486351 | 205613 | 2056.13 | 2102 | 2152 | 2604 | 2604 | +| 1 | 156 | 0.475545 | 210285 | 2102.85 | 2168 | 2201 | 2920 | 2920 | +| 1 | 158 | 0.474496 | 210750 | 2107.5 | 2150 | 2227 | 2365 | 2365 | +| 1 | 160 | 0.467611 | 213853 | 2138.53 | 2176 | 2255 | 3022 | 3022 | +| 1 | 162 | 0.460397 | 217204 | 2172.04 | 2194 | 2331 | 2988 | 2988 | +| 1 | 164 | 0.456759 | 218934 | 2189.34 | 2230 | 2348 | 2636 | 2636 | +| 1 | 166 | 0.449879 | 222282 | 2222.82 | 2249 | 2322 | 3306 | 3306 | +| 1 | 168 | 0.440624 | 226951 | 2269.51 | 2299 | 2383 | 3632 | 3632 | +| 1 | 170 | 0.44145 | 226526 | 2265.26 | 2305 | 2426 | 2585 | 2585 | +| 1 | 172 | 0.431635 | 231677 | 2316.77 | 2378 | 2417 | 2714 | 2714 | +| 1 | 174 | 0.42495 | 235322 | 2353.22 | 2391 | 2480 | 2831 | 2831 | +| 1 | 176 | 0.409281 | 244331 | 2443.31 | 2563 | 2624 | 3402 | 3402 | +| 1 | 178 | 0.41879 | 238783 | 2387.83 | 2444 | 2525 | 2649 | 2649 | +| 1 | 180 | 0.412865 | 242210 | 2422.1 | 2454 | 2564 | 2992 | 2992 | +| 1 | 182 | 0.40871 | 244672 | 2446.72 | 2511 | 2568 | 2683 | 2683 | +| 1 | 184 | 0.403302 | 247953 | 2479.53 | 2531 | 2613 | 3200 | 3200 | +| 1 | 186 | 0.396192 | 252403 | 2524.03 | 2583 | 2622 | 3824 | 3824 | +| 1 | 188 | 0.391426 | 255476 | 2554.76 | 2602 | 2646 | 3317 | 3317 | +| 1 | 190 | 0.389557 | 256702 | 2567.02 | 2640 | 2700 | 3580 | 3580 | +| 1 | 192 | 0.385269 | 259559 | 2595.59 | 2659 | 2749 | 3585 | 3585 | +| 1 | 194 | 0.379858 | 263256 | 2632.56 | 2716 | 2816 | 3394 | 3394 | +| 1 | 196 | 0.375418 | 266370 | 2663.7 | 2740 | 2831 | 3553 | 3553 | +| 1 | 198 | 0.372534 | 268432 | 2684.32 | 2741 | 2817 | 3939 | 3939 | +| 1 | 200 | 0.367207 | 272326 | 2723.26 | 2759 | 2841 | 4114 | 4114 | +| 1 | 202 | 0.367204 | 272328 | 2723.28 | 2800 | 2861 | 4085 | 4085 | +| 1 | 204 | 0.361073 | 276952 | 2769.52 | 2830 | 2932 | 4057 | 4057 | +| 1 | 206 | 0.356544 | 280470 | 2804.7 | 2868 | 3070 | 3483 | 3483 | +| 1 | 208 | 0.355707 | 281130 | 2811.3 | 2895 | 3001 | 3093 | 3093 | +| 1 | 210 | 0.355328 | 281430 | 2814.3 | 2888 | 2995 | 3500 | 3500 | +| 1 | 212 | 0.350765 | 285091 | 2850.91 | 2927 | 3016 | 3267 | 3267 | +| 1 | 214 | 0.338437 | 295476 | 2954.76 | 3040 | 3173 | 4770 | 4770 | +| 1 | 216 | 0.342413 | 292045 | 2920.45 | 2980 | 3085 | 3951 | 3951 | +| 1 | 218 | 0.341919 | 292467 | 2924.67 | 2985 | 3078 | 3437 | 3437 | +| 1 | 220 | 0.336953 | 296777 | 2967.77 | 3048 | 3119 | 3429 | 3429 | +| 1 | 222 | 0.333076 | 300232 | 3002.32 | 3042 | 3144 | 4037 | 4037 | +| 1 | 224 | 0.331037 | 302081 | 3020.81 | 3102 | 3221 | 3512 | 3512 | +| 1 | 226 | 0.327405 | 305432 | 3054.32 | 3127 | 3274 | 3755 | 3755 | +| 1 | 228 | 0.324315 | 308342 | 3083.42 | 3187 | 3277 | 3471 | 3471 | +| 1 | 230 | 0.324681 | 307995 | 3079.95 | 3159 | 3232 | 3551 | 3551 | +| 1 | 232 | 0.321199 | 311333 | 3113.33 | 3161 | 3285 | 3684 | 3684 | +| 1 | 234 | 0.314788 | 317674 | 3176.74 | 3281 | 3324 | 3742 | 3742 | +| 1 | 236 | 0.316314 | 316142 | 3161.42 | 3256 | 3342 | 3585 | 3585 | +| 1 | 238 | 0.311316 | 321217 | 3212.17 | 3337 | 3412 | 4320 | 4320 | +| 1 | 240 | 0.306876 | 325865 | 3258.65 | 3319 | 3395 | 4779 | 4779 | +| 1 | 242 | 0.30333 | 329674 | 3296.74 | 3414 | 3480 | 3951 | 3951 | + +![](gpu-serving-multi-card-single-concurrency-qps-batchsize-client1.png) + +![](gpu-serving-multi-card-single-concurrency-time-batchsize-client1.png) + + +client 2： + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|----------|--------|---------|----------|----------|----------|------------| +| 1 | 2 | 26.71654 | 3743 | 37.43 | 39 | 41 | 65 | 65 | +| 1 | 4 | 16.34521 | 6118 | 61.18 | 63 | 65 | 114 | 114 | +| 1 | 6 | 11.58883 | 8629 | 86.29 | 89 | 94 | 125 | 125 | +| 1 | 8 | 9.084302 | 11008 | 110.08 | 116 | 119 | 166 | 166 | +| 1 | 10 | 7.747134 | 12908 | 129.08 | 132 | 140 | 163 | 163 | +| 1 | 12 | 6.392636 | 15643 | 156.43 | 159 | 164 | 191 | 191 | +| 1 | 14 | 5.747126 | 17400 | 174 | 178 | 183 | 246 | 246 | +| 1 | 16 | 4.916904 | 20338 | 203.38 | 210 | 216 | 260 | 260 | +| 1 | 18 | 4.441878 | 22513 | 225.13 | 232 | 237 | 262 | 262 | +| 1 | 20 | 3.993132 | 25043 | 250.43 | 256 | 265 | 317 | 317 | +| 1 | 22 | 3.653502 | 27371 | 273.71 | 280 | 288 | 326 | 326 | +| 1 | 24 | 3.328119 | 30047 | 300.47 | 307 | 321 | 468 | 468 | +| 1 | 26 | 3.171985 | 31526 | 315.26 | 320 | 329 | 353 | 353 | +| 1 | 28 | 2.894021 | 34554 | 345.54 | 346 | 362 | 682 | 682 | +| 1 | 30 | 2.768319 | 36123 | 361.23 | 370 | 379 | 414 | 414 | +| 1 | 32 | 2.588193 | 38637 | 386.37 | 389 | 399 | 522 | 522 | +| 1 | 34 | 2.421483 | 41297 | 412.97 | 420 | 448 | 696 | 696 | +| 1 | 36 | 2.280762 | 43845 | 438.45 | 447 | 461 | 565 | 565 | +| 1 | 38 | 2.182548 | 45818 | 458.18 | 464 | 484 | 556 | 556 | +| 1 | 40 | 2.04499 | 48900 | 489 | 501 | 518 | 698 | 698 | +| 1 | 42 | 1.977535 | 50568 | 505.68 | 515 | 525 | 619 | 619 | +| 1 | 44 | 1.879028 | 53219 | 532.19 | 540 | 565 | 649 | 649 | +| 1 | 46 | 1.717416 | 58227 | 582.27 | 595 | 615 | 679 | 679 | +| 1 | 48 | 1.668669 | 59928 | 599.28 | 605 | 623 | 671 | 671 | +| 1 | 50 | 1.571907 | 63617 | 636.17 | 657 | 674 | 722 | 722 | +| 1 | 52 | 1.515726 | 65975 | 659.75 | 674 | 691 | 853 | 853 | +| 1 | 54 | 1.460323 | 68478 | 684.78 | 701 | 719 | 843 | 843 | +| 1 | 56 | 1.424075 | 70221 | 702.21 | 712 | 739 | 779 | 779 | +| 1 | 58 | 1.340824 | 74581 | 745.81 | 754 | 807 | 1337 | 1337 | +| 1 | 60 | 1.344411 | 74382 | 743.82 | 752 | 772 | 856 | 856 | +| 1 | 62 | 1.278805 | 78198 | 781.98 | 793 | 818 | 992 | 992 | +| 1 | 64 | 1.24148 | 80549 | 805.49 | 818 | 840 | 1194 | 1194 | +| 1 | 66 | 1.203847 | 83067 | 830.67 | 838 | 874 | 978 | 978 | +| 1 | 68 | 1.168757 | 85561 | 855.61 | 866 | 901 | 1256 | 1256 | +| 1 | 70 | 1.134391 | 88153 | 881.53 | 886 | 911 | 1259 | 1259 | +| 1 | 72 | 1.09523 | 91305 | 913.05 | 941 | 964 | 1021 | 1021 | +| 1 | 74 | 1.071949 | 93288 | 932.88 | 950 | 975 | 1083 | 1083 | +| 1 | 76 | 1.046518 | 95555 | 955.55 | 968 | 989 | 1107 | 1107 | +| 1 | 78 | 1.020387 | 98002 | 980.02 | 995 | 1003 | 1238 | 1238 | +| 1 | 80 | 0.994649 | 100538 | 1005.38 | 1021 | 1043 | 1214 | 1214 | +| 1 | 82 | 0.967146 | 103397 | 1033.97 | 1052 | 1072 | 1311 | 1311 | +| 1 | 84 | 0.934239 | 107039 | 1070.39 | 1083 | 1131 | 1568 | 1568 | +| 1 | 86 | 0.921353 | 108536 | 1085.36 | 1107 | 1129 | 1197 | 1197 | +| 1 | 88 | 0.893479 | 111922 | 1119.22 | 1141 | 1206 | 1486 | 1486 | +| 1 | 90 | 0.873393 | 114496 | 1144.96 | 1167 | 1211 | 1806 | 1806 | +| 1 | 92 | 0.809422 | 123545 | 1235.45 | 1274 | 1337 | 1591 | 1591 | +| 1 | 94 | 0.795513 | 125705 | 1257.05 | 1284 | 1325 | 1588 | 1588 | +| 1 | 96 | 0.786015 | 127224 | 1272.24 | 1296 | 1324 | 1456 | 1456 | +| 1 | 98 | 0.764941 | 130729 | 1307.29 | 1337 | 1377 | 1495 | 1495 | +| 1 | 100 | 0.741983 | 134774 | 1347.74 | 1377 | 1415 | 1646 | 1646 | +| 1 | 102 | 0.730028 | 136981 | 1369.81 | 1400 | 1450 | 1685 | 1685 | +| 1 | 104 | 0.717875 | 139300 | 1393 | 1431 | 1457 | 1769 | 1769 | +| 1 | 106 | 0.699222 | 143016 | 1430.16 | 1471 | 1512 | 1892 | 1892 | +| 1 | 108 | 0.692761 | 144350 | 1443.5 | 1475 | 1542 | 2001 | 2001 | +| 1 | 110 | 0.668802 | 149521 | 1495.21 | 1529 | 1632 | 2161 | 2161 | +| 1 | 112 | 0.666693 | 149994 | 1499.94 | 1539 | 1583 | 1717 | 1717 | +| 1 | 114 | 0.651861 | 153407 | 1534.07 | 1572 | 1597 | 1803 | 1803 | +| 1 | 116 | 0.64229 | 155693 | 1556.93 | 1593 | 1642 | 2031 | 2031 | +| 1 | 118 | 0.627388 | 159391 | 1593.91 | 1632 | 1671 | 1898 | 1898 | +| 1 | 120 | 0.617501 | 161943 | 1619.43 | 1669 | 1724 | 1928 | 1928 | +| 1 | 122 | 0.608665 | 164294 | 1642.94 | 1680 | 1722 | 2035 | 2035 | +| 1 | 124 | 0.607083 | 164722 | 1647.22 | 1681 | 1719 | 1947 | 1947 | +| 1 | 126 | 0.587865 | 170107 | 1701.07 | 1757 | 1811 | 1921 | 1921 | +| 1 | 128 | 0.576735 | 173390 | 1733.9 | 1778 | 1859 | 2104 | 2104 | +| 1 | 130 | 0.567524 | 176204 | 1762.04 | 1792 | 1880 | 2367 | 2367 | +| 1 | 132 | 0.558822 | 178948 | 1789.48 | 1840 | 1936 | 2407 | 2407 | +| 1 | 134 | 0.551128 | 181446 | 1814.46 | 1851 | 1899 | 3198 | 3198 | +| 1 | 136 | 0.541213 | 184770 | 1847.7 | 1913 | 1981 | 2777 | 2777 | +| 1 | 138 | 0.53749 | 186050 | 1860.5 | 1896 | 1958 | 2078 | 2078 | +| 1 | 140 | 0.530211 | 188604 | 1886.04 | 1928 | 1970 | 2521 | 2521 | +| 1 | 142 | 0.523437 | 191045 | 1910.45 | 1959 | 2005 | 2312 | 2312 | +| 1 | 144 | 0.519891 | 192348 | 1923.48 | 1965 | 2016 | 2317 | 2317 | +| 1 | 146 | 0.509674 | 196204 | 1962.04 | 2016 | 2043 | 2179 | 2179 | +| 1 | 148 | 0.502854 | 198865 | 1988.65 | 2029 | 2125 | 2544 | 2544 | +| 1 | 150 | 0.49988 | 200048 | 2000.48 | 2043 | 2091 | 2495 | 2495 | +| 1 | 152 | 0.491123 | 203615 | 2036.15 | 2105 | 2166 | 2599 | 2599 | +| 1 | 154 | 0.484921 | 206219 | 2062.19 | 2102 | 2163 | 3197 | 3197 | +| 1 | 156 | 0.474877 | 210581 | 2105.81 | 2167 | 2222 | 2584 | 2584 | +| 1 | 158 | 0.474994 | 210529 | 2105.29 | 2168 | 2222 | 2327 | 2327 | +| 1 | 160 | 0.468325 | 213527 | 2135.27 | 2159 | 2272 | 2638 | 2638 | +| 1 | 162 | 0.457815 | 218429 | 2184.29 | 2235 | 2319 | 2986 | 2986 | +| 1 | 164 | 0.457032 | 218803 | 2188.03 | 2226 | 2309 | 2924 | 2924 | +| 1 | 166 | 0.450455 | 221998 | 2219.98 | 2263 | 2340 | 3110 | 3110 | +| 1 | 168 | 0.440455 | 227038 | 2270.38 | 2307 | 2408 | 4202 | 4202 | +| 1 | 170 | 0.441143 | 226684 | 2266.84 | 2322 | 2375 | 2650 | 2650 | +| 1 | 172 | 0.436178 | 229264 | 2292.64 | 2327 | 2396 | 2974 | 2974 | +| 1 | 174 | 0.424744 | 235436 | 2354.36 | 2404 | 2503 | 2818 | 2818 | +| 1 | 176 | 0.409717 | 244071 | 2440.71 | 2542 | 2635 | 3947 | 3947 | +| 1 | 178 | 0.418648 | 238864 | 2388.64 | 2470 | 2561 | 2776 | 2776 | +| 1 | 180 | 0.412746 | 242280 | 2422.8 | 2479 | 2528 | 3451 | 3451 | +| 1 | 182 | 0.408542 | 244773 | 2447.73 | 2512 | 2549 | 2716 | 2716 | +| 1 | 184 | 0.404973 | 246930 | 2469.3 | 2500 | 2641 | 3004 | 3004 | +| 1 | 186 | 0.396032 | 252505 | 2525.05 | 2603 | 2682 | 3701 | 3701 | +| 1 | 188 | 0.391526 | 255411 | 2554.11 | 2586 | 2697 | 4194 | 4194 | +| 1 | 190 | 0.392437 | 254818 | 2548.18 | 2596 | 2676 | 3244 | 3244 | +| 1 | 192 | 0.384134 | 260326 | 2603.26 | 2648 | 2729 | 3479 | 3479 | +| 1 | 194 | 0.378748 | 264028 | 2640.28 | 2719 | 2838 | 4297 | 4297 | +| 1 | 196 | 0.375518 | 266299 | 2662.99 | 2721 | 2858 | 3788 | 3788 | +| 1 | 198 | 0.372256 | 268632 | 2686.32 | 2766 | 2869 | 4055 | 4055 | +| 1 | 200 | 0.366987 | 272489 | 2724.89 | 2767 | 2851 | 4497 | 4497 | +| 1 | 202 | 0.36685 | 272591 | 2725.91 | 2789 | 2888 | 3836 | 3836 | +| 1 | 204 | 0.36094 | 277054 | 2770.54 | 2829 | 2881 | 4396 | 4396 | +| 1 | 206 | 0.356806 | 280264 | 2802.64 | 2916 | 3006 | 4142 | 4142 | +| 1 | 208 | 0.355916 | 280965 | 2809.65 | 2878 | 2953 | 3317 | 3317 | +| 1 | 210 | 0.354379 | 282184 | 2821.84 | 2914 | 3009 | 3225 | 3225 | +| 1 | 212 | 0.350447 | 285350 | 2853.5 | 2940 | 2996 | 3622 | 3622 | +| 1 | 214 | 0.33866 | 295281 | 2952.81 | 3045 | 3121 | 4354 | 4354 | +| 1 | 216 | 0.3423 | 292141 | 2921.41 | 3046 | 3153 | 3806 | 3806 | +| 1 | 218 | 0.341059 | 293204 | 2932.04 | 3014 | 3102 | 3269 | 3269 | +| 1 | 220 | 0.336905 | 296820 | 2968.2 | 3063 | 3124 | 3548 | 3548 | +| 1 | 222 | 0.333258 | 300068 | 3000.68 | 3102 | 3193 | 3682 | 3682 | +| 1 | 224 | 0.331324 | 301819 | 3018.19 | 3118 | 3220 | 3743 | 3743 | +| 1 | 226 | 0.327409 | 305428 | 3054.28 | 3146 | 3278 | 3798 | 3798 | +| 1 | 228 | 0.325404 | 307310 | 3073.1 | 3195 | 3300 | 3628 | 3628 | +| 1 | 230 | 0.3249 | 307787 | 3077.87 | 3174 | 3238 | 3361 | 3361 | +| 1 | 232 | 0.321356 | 311181 | 3111.81 | 3179 | 3325 | 3691 | 3691 | +| 1 | 234 | 0.315266 | 317192 | 3171.92 | 3277 | 3393 | 3858 | 3858 | +| 1 | 236 | 0.318074 | 314392 | 3143.92 | 3243 | 3295 | 3645 | 3645 | +| 1 | 238 | 0.309675 | 322919 | 3229.19 | 3350 | 3487 | 4056 | 4056 | +| 1 | 240 | 0.303451 | 329542 | 3295.42 | 3384 | 3590 | 5369 | 5369 | +| 1 | 242 | 0.303498 | 329492 | 3294.92 | 3396 | 3439 | 4505 | 4505 | + +![](gpu-serving-multi-card-single-concurrency-qps-batchsize-client2.png) + +![](gpu-serving-multi-card-single-concurrency-time-batchsize-client2.png) + + +结论：与单卡单进程相比，qps基本翻倍，各个响应时间基本不变，可以认为各个卡之间无影响。 + +## 3.5 多卡多并发响应时间与batch_size的变化关系 + + +启动2个Serving实例，分别绑定GPU0和GPU1；再启动2个client端，分别向2个Serving进程以并发数=10发送请求，不断增大batch_size。 + +client 1： + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|---------|-------------|----------|----------|----------|------------| +| 10 | 10 | 18.008284 | 55530 | 555.303 | 642 | 709 | 769 | 785 | +| 10 | 20 | 9.865825 | 101360 | 1013.604 | 1062 | 1179 | 1421 | 1466 | +| 10 | 30 | 6.661737 | 150111 | 1501.116 | 1573 | 1726 | 2034 | 2185 | +| 10 | 40 | 5.164756 | 193620 | 1936.207 | 2049 | 2149 | 2601 | 2690 | +| 10 | 50 | 4.221155 | 236902 | 2369.029 | 2557 | 2642 | 3257 | 3890 | +| 10 | 60 | 2.741273 | 364794 | 3647.942 | 4711 | 5351 | 11657 | 15556 | +| 10 | 70 | 2.845889 | 351384 | 3513.843 | 3711 | 4003 | 4999 | 5554 | +| 10 | 80 | 2.516749 | 397338 | 3973.387 | 4225 | 4400 | 5316 | 6568 | +| 10 | 90 | 2.279223 | 438746 | 4387.463 | 4725 | 4990 | 6146 | 7834 | +| 10 | 100 | 1.945752 | 513940 | 5139.407 | 5466 | 5762 | 7093 | 8963 | +| 10 | 110 | 1.851197 | 540191 | 5401.912 | 5805 | 6024 | 7253 | 10026 | +| 20 | 10 | 17.879972 | 111857 | 1118.578 | 1246 | 1270 | 1302 | 1342 | +| 20 | 20 | 9.912473 | 201766 | 2017.661 | 2126 | 2280 | 2444 | 2518 | +| 20 | 30 | 6.731015 | 297132 | 2971.3235 | 3139 | 3318 | 3575 | 3819 | +| 20 | 40 | 5.093271 | 392675 | 3926.757 | 4130 | 4345 | 4770 | 5046 | +| 20 | 50 | 4.167448 | 479910 | 4799.108 | 5079 | 5230 | 5864 | 6170 | +| 20 | 60 | 3.514994 | 568991 | 5689.9185 | 6073 | 6245 | 6839 | 7535 | +| 20 | 70 | 2.961598 | 675311 | 6753.1105 | 7163 | 7433 | 8218 | 9018 | +| 20 | 80 | 2.536989 | 788336 | 7883.3685 | 8355 | 8719 | 9897 | 10623 | +| 20 | 90 | 2.20668 | 906339 | 9063.399 | 9624 | 10006 | 11071 | 14247 | +| 20 | 100 | 1.937449 | 1032285 | 10322.854 | 10837 | 11318 | 12615 | 16793 | +| 20 | 110 | 1.804763 | 1108179 | 11081.7915 | 11695 | 12083 | 13537 | 15422 | +| 30 | 10 | 17.801513 | 168525 | 1685.253333 | 1785 | 1835 | 2030 | 2102 | +| 30 | 20 | 9.891197 | 303300 | 3033.001333 | 3175 | 3312 | 3494 | 3575 | +| 30 | 30 | 6.739619 | 445129 | 4451.299667 | 4721 | 4911 | 5226 | 5617 | +| 30 | 40 | 5.315031 | 564437 | 5644.378 | 5952 | 6081 | 6631 | 8711 | +| 30 | 50 | 4.123065 | 727614 | 7276.141 | 7696 | 7912 | 8735 | 9337 | +| 30 | 60 | 3.406521 | 880664 | 8806.644333 | 9236 | 9578 | 10875 | 12432 | +| 30 | 70 | 2.880179 | 1041602 | 10416.025 | 10948 | 11238 | 12129 | 20313 | +| 30 | 80 | 2.532526 | 1184588 | 11845.88433 | 12463 | 12872 | 13801 | 23180 | +| 30 | 90 | 2.297083 | 1306004 | 13060.04433 | 13904 | 14173 | 15172 | 28654 | +| 30 | 100 | 2.002666 | 1498003 | 14980.03733 | 15855 | 16244 | 17360 | 18525 | +| 30 | 110 | 1.83226 | 1637322 | 16373.22567 | 17140 | 17532 | 28046 | 62853 | +| 40 | 10 | 18.22257 | 219508 | 2195.08175 | 2308 | 2339 | 2528 | 2571 | +| 40 | 20 | 9.752268 | 410161 | 4101.6195 | 4299 | 4434 | 4657 | 6085 | +| 40 | 30 | 6.688695 | 598024 | 5980.24125 | 6253 | 6419 | 6636 | 6843 | +| 40 | 40 | 5.04902 | 792233 | 7922.33075 | 8286 | 8515 | 8911 | 9145 | +| 40 | 50 | 4.096338 | 976482 | 9764.82875 | 10252 | 10485 | 11262 | 12351 | +| 40 | 60 | 3.41772 | 1170371 | 11703.7165 | 12202 | 12530 | 13455 | 14044 | +| 40 | 70 | 2.893795 | 1382268 | 13822.68025 | 14445 | 14746 | 15817 | 47842 | +| 40 | 80 | 2.560123 | 1562425 | 15624.2525 | 16406 | 16682 | 17744 | 59196 | +| 40 | 90 | 2.322667 | 1722158 | 17448.41793 | 17162 | 17661 | 60713 | 64893 | +| 40 | 100 | 2.389919 | 1673697 | 18162.75529 | 17667 | 18244 | 62601 | 100313 | +| 40 | 110 | 2.595806 | 1540947 | 18241.46316 | 17599 | 18057 | 63818 | 106641 | +| 50 | 10 | 17.804936 | 280821 | 2808.2188 | 2940 | 3016 | 3109 | 3155 | +| 50 | 20 | 9.779722 | 511262 | 5112.6264 | 5383 | 5476 | 5709 | 5863 | +| 50 | 30 | 6.704192 | 745802 | 7458.0222 | 7769 | 7919 | 8199 | 8953 | +| 50 | 40 | 5.080263 | 984201 | 9842.0198 | 10239 | 10424 | 10954 | 13369 | +| 50 | 50 | 4.056525 | 1232582 | 12325.8268 | 12860 | 13162 | 13952 | 14435 | +| 50 | 60 | 3.391622 | 1474221 | 14742.2166 | 15457 | 15893 | 16684 | 18819 | +| 50 | 70 | 3.091244 | 1617472 | 16771.80071 | 16865 | 17324 | 60101 | 100898 | +| 50 | 80 | 2.926027 | 1708802 | 18240.84799 | 17097 | 17623 | 63539 | 104849 | +| 50 | 90 | 3.19614 | 1564387 | 18535.391 | 17315 | 17946 | 63391 | 104642 | +| 50 | 100 | 3.428791 | 1458240 | 18972.68983 | 17907 | 18387 | 65215 | 104798 | +| 50 | 110 | 3.98105 | 1255950 | 18584.65197 | 17727 | 18299 | 64607 | 101528 | + +![](gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client1.png) + +![](gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client1.png) + +client 2： + + +| 并发数 | batch size | qps | 总时间 | 平均响应时间 | 80分位响应时间 | 90分位响应时间 | 99分位响应时间 | 99.9分位响应时间 | +|-----|------------|-----------|---------|-------------|----------|----------|----------|------------| +| 10 | 10 | 17.911838 | 55829 | 558.29 | 660 | 716 | 767 | 829 | +| 10 | 20 | 10.071102 | 99294 | 992.946 | 1044 | 1094 | 1334 | 1425 | +| 10 | 30 | 6.773232 | 147640 | 1476.406 | 1550 | 1718 | 2069 | 2180 | +| 10 | 40 | 5.068835 | 197284 | 1972.843 | 2072 | 2269 | 2654 | 2867 | +| 10 | 50 | 4.232141 | 236287 | 2362.873 | 2512 | 2610 | 3202 | 3518 | +| 10 | 60 | 2.868971 | 348557 | 3485.575 | 4417 | 4864 | 7832 | 9154 | +| 10 | 70 | 2.965414 | 337221 | 3372.218 | 3590 | 3757 | 4693 | 5634 | +| 10 | 80 | 2.596715 | 385102 | 3851.022 | 4128 | 4308 | 5044 | 5858 | +| 10 | 90 | 2.277536 | 439071 | 4390.714 | 4688 | 4895 | 6029 | 6622 | +| 10 | 100 | 2.022523 | 494432 | 4944.321 | 5316 | 5494 | 6473 | 7203 | +| 10 | 110 | 1.800653 | 555354 | 5553.541 | 5907 | 6197 | 7703 | 10422 | +| 20 | 10 | 18.120374 | 110373 | 1103.734 | 1234 | 1266 | 1317 | 1367 | +| 20 | 20 | 10.012265 | 199755 | 1997.559 | 2089 | 2223 | 2447 | 2508 | +| 20 | 30 | 6.703311 | 298360 | 2983.6015 | 3171 | 3342 | 3553 | 3685 | +| 20 | 40 | 5.136305 | 389385 | 3893.8545 | 4075 | 4245 | 4679 | 4864 | +| 20 | 50 | 4.268515 | 468547 | 4685.4785 | 4961 | 5109 | 5688 | 6049 | +| 20 | 60 | 3.517219 | 568631 | 5686.314 | 6020 | 6194 | 6970 | 7398 | +| 20 | 70 | 2.985079 | 669999 | 6699.9995 | 7089 | 7351 | 8276 | 10734 | +| 20 | 80 | 2.609539 | 766419 | 7664.193 | 8103 | 8315 | 9341 | 12035 | +| 20 | 90 | 2.239466 | 893070 | 8930.704 | 9496 | 9785 | 10967 | 12314 | +| 20 | 100 | 1.968684 | 1015907 | 10159.0795 | 10745 | 11239 | 12502 | 16077 | +| 20 | 110 | 1.812463 | 1103471 | 11034.7155 | 11667 | 11964 | 13870 | 16550 | +| 30 | 10 | 17.881517 | 167771 | 1677.713333 | 1773 | 1806 | 1951 | 2527 | +| 30 | 20 | 9.881716 | 303591 | 3035.910667 | 3228 | 3359 | 3529 | 3695 | +| 30 | 30 | 6.745059 | 444770 | 4447.703667 | 4719 | 4844 | 5062 | 5247 | +| 30 | 40 | 5.123948 | 585486 | 5854.866333 | 6136 | 6345 | 6689 | 6899 | +| 30 | 50 | 4.243005 | 707046 | 7070.469 | 7432 | 7597 | 8180 | 11107 | +| 30 | 60 | 3.460866 | 866835 | 8668.355 | 9115 | 9464 | 10176 | 10713 | +| 30 | 70 | 2.937211 | 1021377 | 10213.77733 | 10771 | 11023 | 12116 | 14869 | +| 30 | 80 | 2.524868 | 1188181 | 11881.811 | 12412 | 12744 | 13858 | 19586 | +| 30 | 90 | 2.271227 | 1320872 | 13208.728 | 13898 | 14260 | 15619 | 25786 | +| 30 | 100 | 2.028742 | 1478749 | 14787.49667 | 15657 | 16033 | 17395 | 27276 | +| 30 | 110 | 1.830449 | 1638942 | 16389.42833 | 17089 | 17440 | 31311 | 62343 | +| 40 | 10 | 18.006176 | 222146 | 2221.46775 | 2315 | 2359 | 2547 | 2572 | +| 40 | 20 | 9.891686 | 404380 | 4043.8095 | 4248 | 4363 | 4542 | 4658 | +| 40 | 30 | 6.720927 | 595156 | 5951.56925 | 6257 | 6394 | 6610 | 7016 | +| 40 | 40 | 5.10964 | 782834 | 7828.34 | 8168 | 8395 | 8722 | 11487 | +| 40 | 50 | 4.119032 | 971102 | 9711.0215 | 10159 | 10433 | 10948 | 11252 | +| 40 | 60 | 3.4089 | 1173399 | 11733.992 | 12281 | 12624 | 13369 | 14668 | +| 40 | 70 | 2.844483 | 1406231 | 14062.3195 | 14664 | 15082 | 15904 | 18239 | +| 40 | 80 | 2.574988 | 1553405 | 15534.0575 | 16255 | 16566 | 17966 | 58696 | +| 40 | 90 | 2.37624 | 1683332 | 17128.79954 | 17148 | 17729 | 59420 | 97044 | +| 40 | 100 | 2.319443 | 1724552 | 18578.53057 | 17896 | 18326 | 64014 | 108697 | +| 40 | 110 | 2.368145 | 1689086 | 19074.94777 | 17729 | 18423 | 66734 | 110545 | +| 50 | 10 | 17.97727 | 278129 | 2781.292 | 2918 | 2998 | 3091 | 3145 | +| 50 | 20 | 9.961767 | 501919 | 5019.1924 | 5286 | 5405 | 5858 | 6067 | +| 50 | 30 | 6.723476 | 743663 | 7436.6352 | 7767 | 7887 | 8251 | 8688 | +| 50 | 40 | 5.139671 | 972825 | 9728.2504 | 10131 | 10350 | 10733 | 13665 | +| 50 | 50 | 4.104245 | 1218251 | 12182.5152 | 12709 | 13032 | 13693 | 14109 | +| 50 | 60 | 3.439755 | 1453592 | 14535.9218 | 15261 | 15646 | 16469 | 16898 | +| 50 | 70 | 2.907504 | 1719688 | 17196.8836 | 17057 | 17457 | 61265 | 94428 | +| 50 | 80 | 2.982454 | 1676472 | 18026.58559 | 17221 | 17728 | 62984 | 105593 | +| 50 | 90 | 3.265575 | 1531124 | 18253.74416 | 17210 | 17818 | 63490 | 108811 | +| 50 | 100 | 3.805989 | 1313719 | 18015.89441 | 17708 | 18188 | 61707 | 105320 | +| 50 | 110 | 4.218359 | 1185295 | 17926.42862 | 17490 | 17925 | 63049 | 100128 | + +![](gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client2.png) + +![](gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client2.png) + +结论：多卡情况，随着并发数的增加，qps基本一致，响应时间有所增加，与单卡多并发相比，响应时间增加10%左右。 diff --git a/doc/INDEX.md b/doc/INDEX.md index c9399b2f5cf50cda10062023f009b92bfc0c8158..11f330a10eefa83522631e9f630fc27da93dabfe 100644 --- a/doc/INDEX.md +++ b/doc/INDEX.md @@ -16,4 +16,6 @@ [Benchmarking](BENCHMARKING.md) +[GPU Benchmarking](GPU_BENCHMARKING.md) + [FAQ](FAQ.md) diff --git a/doc/SERVING_CONFIGURE.md b/doc/SERVING_CONFIGURE.md index f5887f5cd139b1e70b49f4eee2e2552658692103..f34436a064963ae25411af53b3116f00b9a493de 100644 --- a/doc/SERVING_CONFIGURE.md +++ b/doc/SERVING_CONFIGURE.md @@ -165,7 +165,7 @@ reloadable_type: 检查reload条件：timestamp_ne/timestamp_gt/md5sum/revision/ model_data_path: 模型文件路径 -runtime_thread_num: 若大于0， 则启用bsf多线程调度框架，在每个预测bthread worker内启动多线程预测。 +runtime_thread_num: 若大于0， 则启用bsf多线程调度框架，在每个预测bthread worker内启动多线程预测。要注意的是，当启用worker内多线程预测，workflow中OP需要用Serving框架的BatchTensor类做预测的输入和输出 (predictor/framework/infer_data.h, `class BatchTensor`)。 batch_infer_size: 启用bsf多线程预测时，每个预测线程的batch size diff --git a/doc/gpu-local-qps-batchsize.png b/doc/gpu-local-qps-batchsize.png new file mode 100755 index 0000000000000000000000000000000000000000..97db37518e2afd3cacccb82b0be967d605655bc9 Binary files /dev/null and b/doc/gpu-local-qps-batchsize.png differ diff --git a/doc/gpu-local-qps-concurrency.png b/doc/gpu-local-qps-concurrency.png new file mode 100755 index 0000000000000000000000000000000000000000..cd4c3dd8f672c3620e452213ab363b1b7b00905e Binary files /dev/null and b/doc/gpu-local-qps-concurrency.png differ diff --git a/doc/gpu-local-time-batchsize.png b/doc/gpu-local-time-batchsize.png new file mode 100755 index 0000000000000000000000000000000000000000..35af4a6b1e6cea4793861c6dc34b52ebddd38495 Binary files /dev/null and b/doc/gpu-local-time-batchsize.png differ diff --git a/doc/gpu-local-time-concurrency.png b/doc/gpu-local-time-concurrency.png new file mode 100755 index 0000000000000000000000000000000000000000..28626f3470fae72685300d8a6aed05c23da9593f Binary files /dev/null and b/doc/gpu-local-time-concurrency.png differ diff --git a/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client1.png b/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client1.png new file mode 100755 index 0000000000000000000000000000000000000000..4ce901823695184deb10be4b56b2c8e9eb6d5ad7 Binary files /dev/null and b/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client1.png differ diff --git a/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client2.png b/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client2.png new file mode 100755 index 0000000000000000000000000000000000000000..b1535b95e112307c4ce917b1bdad5f7b34e24ad3 Binary files /dev/null and b/doc/gpu-serving-multi-card-multi-concurrency-qps-batchsize-concurrency-client2.png differ diff --git a/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client1.png b/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client1.png new file mode 100755 index 0000000000000000000000000000000000000000..dd651e94f250a649978f0445962723aeb6ef3071 Binary files /dev/null and b/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client1.png differ diff --git a/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client2.png b/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client2.png new file mode 100755 index 0000000000000000000000000000000000000000..048c095fc0607e80f3b7e51e6e2930978e20dcd0 Binary files /dev/null and b/doc/gpu-serving-multi-card-multi-concurrency-time-batchsize-concurrency-client2.png differ diff --git a/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client1.png b/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client1.png new file mode 100755 index 0000000000000000000000000000000000000000..9823b85055e897ea496d8ae557c9fe84f6bfedd4 Binary files /dev/null and b/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client1.png differ diff --git a/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client2.png b/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client2.png new file mode 100755 index 0000000000000000000000000000000000000000..8a3723ed217eca900e561d35dfa01fdf916c57a6 Binary files /dev/null and b/doc/gpu-serving-multi-card-single-concurrency-qps-batchsize-client2.png differ diff --git a/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client1.png b/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client1.png new file mode 100755 index 0000000000000000000000000000000000000000..40504bb13f15feee0ebdb90f3490a57d13cd53d8 Binary files /dev/null and b/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client1.png differ diff --git a/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client2.png b/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client2.png new file mode 100755 index 0000000000000000000000000000000000000000..d8a90087444d85eccf7be8dd34ae1a6e5d4eaa4e Binary files /dev/null and b/doc/gpu-serving-multi-card-single-concurrency-time-batchsize-client2.png differ diff --git a/doc/gpu-serving-single-card-qps-batchsize.png b/doc/gpu-serving-single-card-qps-batchsize.png new file mode 100755 index 0000000000000000000000000000000000000000..35e0a5b60ca5659cc8e0a6e2e66e90c26a385d2a Binary files /dev/null and b/doc/gpu-serving-single-card-qps-batchsize.png differ diff --git a/doc/gpu-serving-single-card-qps-concurrency.png b/doc/gpu-serving-single-card-qps-concurrency.png new file mode 100755 index 0000000000000000000000000000000000000000..025a3908d1b172b22344a8950f4d9f0bbc6c246d Binary files /dev/null and b/doc/gpu-serving-single-card-qps-concurrency.png differ diff --git a/doc/gpu-serving-single-card-time-batchsize.png b/doc/gpu-serving-single-card-time-batchsize.png new file mode 100755 index 0000000000000000000000000000000000000000..4c7498ae22c9c5febc6d7a6566a839ded6ec6cf2 Binary files /dev/null and b/doc/gpu-serving-single-card-time-batchsize.png differ diff --git a/doc/gpu-serving-single-card-time-concurrency.png b/doc/gpu-serving-single-card-time-concurrency.png new file mode 100755 index 0000000000000000000000000000000000000000..fc4332536bb145846439b2e5693d1da5c3feaac3 Binary files /dev/null and b/doc/gpu-serving-single-card-time-concurrency.png differ diff --git a/doc/qps-threads-bow.png b/doc/qps-threads-bow.png new file mode 100755 index 0000000000000000000000000000000000000000..8123b71ee154ca86ea8b028fba4bceecac363007 Binary files /dev/null and b/doc/qps-threads-bow.png differ diff --git a/doc/qps-threads-cnn.png b/doc/qps-threads-cnn.png new file mode 100755 index 0000000000000000000000000000000000000000..f983d4882be0a923c0bbfbb41bbca56d684d4035 Binary files /dev/null and b/doc/qps-threads-cnn.png differ diff --git a/doc/qps-threads-lstm.png b/doc/qps-threads-lstm.png new file mode 100755 index 0000000000000000000000000000000000000000..d474b8423f6fc7f49f817181fce48adfa3e15df3 Binary files /dev/null and b/doc/qps-threads-lstm.png differ diff --git a/doc/serving-timings.png b/doc/serving-timings.png new file mode 100755 index 0000000000000000000000000000000000000000..32bab31a57799f7f1d4b80f7e809e284df98d332 Binary files /dev/null and b/doc/serving-timings.png differ diff --git a/predictor/framework/memory.cpp b/predictor/framework/memory.cpp index 32f2547817983d8c6d280269b6003de117247c6c..78a4643468eeca9f1da925a67f1e1330b89927a3 100644 --- a/predictor/framework/memory.cpp +++ b/predictor/framework/memory.cpp @@ -19,6 +19,27 @@ namespace baidu { namespace paddle_serving { namespace predictor { +struct MempoolRegion { + MempoolRegion(im::fugue::memory::Region *region, + im::Mempool *mempool) : + _region(region), _mempool(mempool){} + im::fugue::memory::Region *region() {return _region;} + im::Mempool *mempool() {return _mempool;} + + im::fugue::memory::Region* _region; + im::Mempool* _mempool; + ~MempoolRegion() { + if (_region) { + delete _region; + _region = NULL; + } + if (_mempool) { + delete _mempool; + _mempool = NULL; + } + } +}; + int MempoolWrapper::initialize() { if (THREAD_KEY_CREATE(&_bspec_key, NULL) != 0) { LOG(ERROR) << "unable to create thread_key of thrd_data"; @@ -33,16 +54,20 @@ int MempoolWrapper::initialize() { } int MempoolWrapper::thread_initialize() { - _region.init(); - im::Mempool* p_mempool = new (std::nothrow) im::Mempool(&_region); - if (p_mempool == NULL) { + im::fugue::memory::Region *region = new im::fugue::memory::Region(); + region->init(); + im::Mempool* mempool = new (std::nothrow) im::Mempool(region); + MempoolRegion *mempool_region = new MempoolRegion(region, mempool); + if (mempool == NULL) { LOG(ERROR) << "Failed create thread mempool"; return -1; } - if (THREAD_SETSPECIFIC(_bspec_key, p_mempool) != 0) { + if (THREAD_SETSPECIFIC(_bspec_key, mempool_region) != 0) { LOG(ERROR) << "unable to set the thrd_data"; - delete p_mempool; + delete region; + delete mempool; + delete mempool_region; return -1; } @@ -51,23 +76,34 @@ int MempoolWrapper::thread_initialize() { } int MempoolWrapper::thread_clear() { - im::Mempool* p_mempool = (im::Mempool*)THREAD_GETSPECIFIC(_bspec_key); - if (p_mempool) { - p_mempool->release_block(); - _region.reset(); + MempoolRegion* mempool_region = (MempoolRegion*)THREAD_GETSPECIFIC(_bspec_key); + if (mempool_region == NULL) { + LOG(WARNING) << "THREAD_GETSPECIFIC() returned NULL"; + return -1; + } + im::Mempool* mempool = mempool_region->mempool(); + im::fugue::memory::Region* region = mempool_region->region(); + if (mempool) { + mempool->release_block(); + region->reset(); } - return 0; } void* MempoolWrapper::malloc(size_t size) { - im::Mempool* p_mempool = (im::Mempool*)THREAD_GETSPECIFIC(_bspec_key); - if (!p_mempool) { + MempoolRegion* mempool_region = (MempoolRegion*)THREAD_GETSPECIFIC(_bspec_key); + if (mempool_region == NULL) { + LOG(WARNING) << "THREAD_GETSPECIFIC() returned NULL"; + return NULL; + } + + im::Mempool* mempool = mempool_region->mempool(); + if (!mempool) { LOG(WARNING) << "Cannot malloc memory:" << size << ", since mempool is not thread initialized"; return NULL; } - return p_mempool->malloc(size); + return mempool->malloc(size); } } // namespace predictor diff --git a/predictor/framework/memory.h b/predictor/framework/memory.h index 9cc12e8f5c3b89c07578cb7807b03b4df26d75c5..c096bc2c6b2d96ff2bd93daa804bac007cef6295 100644 --- a/predictor/framework/memory.h +++ b/predictor/framework/memory.h @@ -39,7 +39,7 @@ class MempoolWrapper { void* malloc(size_t size); private: - im::fugue::memory::Region _region; + //im::fugue::memory::Region _region; THREAD_KEY_T _bspec_key; }; diff --git a/predictor/src/pdserving.cpp b/predictor/src/pdserving.cpp index be7f988744b6ef0530c8b725cb3d6275725831ec..3fc1eed7b632b68be1c5c9f39db63f260319d48c 100644 --- a/predictor/src/pdserving.cpp +++ b/predictor/src/pdserving.cpp @@ -143,7 +143,7 @@ int main(int argc, char** argv) { std::string filename(argv[0]); filename = filename.substr(filename.find_last_of('/') + 1); - settings.log_file = (std::string("./log/") + filename + ".log").c_str(); + settings.log_file = strdup((std::string("./log/") + filename + ".log").c_str()); settings.delete_old = logging::DELETE_OLD_LOG_FILE; logging::InitLogging(settings);