jonatan1626 commented on issue #17159: Performance regression from 1.4.1 to 1.5.1 URL: https://github.com/apache/incubator-mxnet/issues/17159#issuecomment-568603588 These were the initial numbers, which is why this investigation is going on: Now we are trying to find the culprit for the regression from 1.4.1 -> 1.5.1 There was a gain from 1.5.1 -> 1.6 from the mkldnn upgrade [#15422](https://github.com/apache/incubator-mxnet/pull/15422) =========================== MXNet 1.4.1 Pip ============================== Model: resnext-50 ----Latency---- Model Partition Time: 0 Model Load Time: 84.53512191772461 First Inference: 116.00756645202637 p99: 19.801616668701172 p90: 19.596172332763672 p50: 15.244483947753906 Avg: 15.836643512970845 StdDev: 3.903761278317364 Iterations: 100000 Context: cpu =========================== MXNet 1.5.1 Pip ============================== Model: resnext-50 ----Latency---- Model Partition Time: 0 Model Load Time: 84.19561386108398 First Inference: 121.11115455627441 p99: 34.098148345947266 p90: 30.607223510742188 p50: 27.56810188293457 Avg: 27.974184123915748 StdDev: 7.246931590666572 Iterations: 100000 Context: cpu =========================== MXNet 1.6 Pip ============================== Model: resnext-50 ----Latency---- Model Partition Time: 0 Model Load Time: 71.62761688232422 First Inference: 104.33459281921387 p99: 21.777629852294922 p90: 20.840167999267578 p50: 17.279386520385742 Avg: 18.166795754279853 StdDev: 4.648682601625848 Iterations: 100000 Context: cpu
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
