Github user Dooyoung-Hwang commented on the issue: https://github.com/apache/spark/pull/22347 I tested in my local PC. 3.3 GHz Intel Core i5, and selected 400,000 rows x 25 times. I took a total execution time between decodeUnsafeRows. My tested data is skewed, so gathered rows from executor are distributed between 400000 and 800000. The average execution time decreased from 175.92ms to 93.52ms. Memory usage also improved, and total GC Time is decreased from 13.883 sec to 10.764 sec. ## Before Patch ### GC statics S0 | S1 | E | O | M | CCS | YGC | YGCT | FGC | FGCT | GCT -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- 0 | 100 | 24.92 | 48.66 | 96.92 | 88.25 | 150 | 13.883 | 0 | 0 | 13.883 ### Wall time : AVG 175.92 ms Row Count | Decode Time(ms) -- | -- 428942 | 73 473726 | 106 476322 | 78 509996 | 83 510590 | 124 556896 | 94 556896 | 362 595272 | 193 595272 | 175 642478 | 120 644970 | 279 679544 | 269 693354 | 116 723532 | 124 729912 | 136 730218 | 120 730246 | 184 773640 | 183 774148 | 380 810198 | 128 811606 | 131 859090 | 138 895474 | 314 895954 | 339 939636 | 149 ## After Patch : 93.52ms ### GC statics S0 | S1 | E | O | M | CCS | YGC | YGCT | FGC | FGCT | GCT -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- 0 | 100 | 81.37 | 33.34 | 97.21 | 88.35 | 127 | 10.764 | 0 | 0 | 10.764 ### Wall time : AVG 93.52 ms Row Count | Decode time (ms) -- | -- 421922 | 61 422516 | 180 422850 | 110 473218 | 62 473218 | 103 473438 | 115 507198 | 60 554606 | 144 557202 | 119 601392 | 71 642652 | 61 645276 | 64 679036 | 64 679036 | 63 729624 | 242 729652 | 62 729912 | 131 773814 | 122 774234 | 62 807908 | 59 810198 | 64 814900 | 72 844772 | 59 858582 | 127 858582 | 61
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org