gortiz commented on PR #12354:
URL: https://github.com/apache/pinot/pull/12354#issuecomment-1964199109
Extra investigations:
## Why performance in `normal` with `null handling` decreases so much when
going from 2 to 4 interval?
I've checked with perfnorm:
```
Benchmark (_impl)
(_nullHandlingEnabled) (_nullInterval) Mode Cnt Score Error
Units
BenchmarkSumAggregation2.test normal
true 2 thrpt 5 103.386 ± 3.852 ops/ms
BenchmarkSumAggregation2.test:CPI normal
true 2 thrpt 0.177 clks/insn
BenchmarkSumAggregation2.test:IPC normal
true 2 thrpt 5.660 insns/clk
BenchmarkSumAggregation2.test:L1-dcache-load-misses normal
true 2 thrpt 1321.767 #/op
BenchmarkSumAggregation2.test:L1-dcache-loads normal
true 2 thrpt 15314.726 #/op
BenchmarkSumAggregation2.test:L1-icache-load-misses normal
true 2 thrpt 1.106 #/op
BenchmarkSumAggregation2.test:L1-icache-loads normal
true 2 thrpt 329.697 #/op
BenchmarkSumAggregation2.test:branch-misses normal
true 2 thrpt 8.702 #/op
BenchmarkSumAggregation2.test:branches normal
true 2 thrpt 60461.749 #/op
BenchmarkSumAggregation2.test:cycles normal
true 2 thrpt 41090.010 #/op
BenchmarkSumAggregation2.test:dTLB-load-misses normal
true 2 thrpt 0.301 #/op
BenchmarkSumAggregation2.test:dTLB-loads normal
true 2 thrpt 2.005 #/op
BenchmarkSumAggregation2.test:iTLB-load-misses normal
true 2 thrpt 0.048 #/op
BenchmarkSumAggregation2.test:iTLB-loads normal
true 2 thrpt 0.219 #/op
BenchmarkSumAggregation2.test:instructions normal
true 2 thrpt 232553.341 #/op
BenchmarkSumAggregation2.test:stalled-cycles-backend normal
true 2 thrpt 25474.941 #/op
BenchmarkSumAggregation2.test:stalled-cycles-frontend normal
true 2 thrpt 57.691 #/op
BenchmarkSumAggregation2.test normal
true 4 thrpt 5 3.525 ± 0.129 ops/ms
BenchmarkSumAggregation2.test:CPI normal
true 4 thrpt 0.378 clks/insn
BenchmarkSumAggregation2.test:IPC normal
true 4 thrpt 2.645 insns/clk
BenchmarkSumAggregation2.test:L1-dcache-load-misses normal
true 4 thrpt 1689.903 #/op
BenchmarkSumAggregation2.test:L1-dcache-loads normal
true 4 thrpt 685807.174 #/op
BenchmarkSumAggregation2.test:L1-icache-load-misses normal
true 4 thrpt 21.785 #/op
BenchmarkSumAggregation2.test:L1-icache-loads normal
true 4 thrpt 10603.011 #/op
BenchmarkSumAggregation2.test:branch-misses normal
true 4 thrpt 15379.544 #/op
BenchmarkSumAggregation2.test:branches normal
true 4 thrpt 817946.383 #/op
BenchmarkSumAggregation2.test:cycles normal
true 4 thrpt 1223502.258 #/op
BenchmarkSumAggregation2.test:dTLB-load-misses normal
true 4 thrpt 2.123 #/op
BenchmarkSumAggregation2.test:dTLB-loads normal
true 4 thrpt 16.638 #/op
BenchmarkSumAggregation2.test:iTLB-load-misses normal
true 4 thrpt 0.639 #/op
BenchmarkSumAggregation2.test:iTLB-loads normal
true 4 thrpt 2.910 #/op
BenchmarkSumAggregation2.test:instructions normal
true 4 thrpt 3236651.002 #/op
BenchmarkSumAggregation2.test:stalled-cycles-backend normal
true 4 thrpt 116767.982 #/op
BenchmarkSumAggregation2.test:stalled-cycles-frontend normal
true 4 thrpt 17092.450 #/op
```
AFAIU there are far more cache failures and mispredictions.
## Why performance in `foldDouble` with `null handling` increases so much
when going from 4 to 8 interval?
Same reason:
```
BenchmarkSumAggregation2.test foldDouble
true 4 thrpt 5 68.321 ± 0.344 ops/ms
BenchmarkSumAggregation2.test:CPI foldDouble
true 4 thrpt 0.234 clks/insn
BenchmarkSumAggregation2.test:IPC foldDouble
true 4 thrpt 4.264 insns/clk
BenchmarkSumAggregation2.test:L1-dcache-load-misses foldDouble
true 4 thrpt 2382.564 #/op
BenchmarkSumAggregation2.test:L1-dcache-loads foldDouble
true 4 thrpt 103018.730 #/op
BenchmarkSumAggregation2.test:L1-icache-load-misses foldDouble
true 4 thrpt 4.476 #/op
BenchmarkSumAggregation2.test:L1-icache-loads foldDouble
true 4 thrpt 402.647 #/op
BenchmarkSumAggregation2.test:branch-misses foldDouble
true 4 thrpt 5.758 #/op
BenchmarkSumAggregation2.test:branches foldDouble
true 4 thrpt 41678.023 #/op
BenchmarkSumAggregation2.test:cycles foldDouble
true 4 thrpt 64669.720 #/op
BenchmarkSumAggregation2.test:dTLB-load-misses foldDouble
true 4 thrpt 15.780 #/op
BenchmarkSumAggregation2.test:dTLB-loads foldDouble
true 4 thrpt 21.655 #/op
BenchmarkSumAggregation2.test:iTLB-load-misses foldDouble
true 4 thrpt 0.501 #/op
BenchmarkSumAggregation2.test:iTLB-loads foldDouble
true 4 thrpt 0.600 #/op
BenchmarkSumAggregation2.test:instructions foldDouble
true 4 thrpt 275779.218 #/op
BenchmarkSumAggregation2.test:stalled-cycles-backend foldDouble
true 4 thrpt 37349.094 #/op
BenchmarkSumAggregation2.test:stalled-cycles-frontend foldDouble
true 4 thrpt 700.245 #/op
BenchmarkSumAggregation2.test foldDouble
true 8 thrpt 5 120.523 ± 1.062 ops/ms
BenchmarkSumAggregation2.test:CPI foldDouble
true 8 thrpt 0.233 clks/insn
BenchmarkSumAggregation2.test:IPC foldDouble
true 8 thrpt 4.293 insns/clk
BenchmarkSumAggregation2.test:L1-dcache-load-misses foldDouble
true 8 thrpt 1838.574 #/op
BenchmarkSumAggregation2.test:L1-dcache-loads foldDouble
true 8 thrpt 56420.620 #/op
BenchmarkSumAggregation2.test:L1-icache-load-misses foldDouble
true 8 thrpt 1.978 #/op
BenchmarkSumAggregation2.test:L1-icache-loads foldDouble
true 8 thrpt 217.917 #/op
BenchmarkSumAggregation2.test:branch-misses foldDouble
true 8 thrpt 2.651 #/op
BenchmarkSumAggregation2.test:branches foldDouble
true 8 thrpt 21608.391 #/op
BenchmarkSumAggregation2.test:cycles foldDouble
true 8 thrpt 34548.798 #/op
BenchmarkSumAggregation2.test:dTLB-load-misses foldDouble
true 8 thrpt 8.186 #/op
BenchmarkSumAggregation2.test:dTLB-loads foldDouble
true 8 thrpt 10.305 #/op
BenchmarkSumAggregation2.test:iTLB-load-misses foldDouble
true 8 thrpt 0.277 #/op
BenchmarkSumAggregation2.test:iTLB-loads foldDouble
true 8 thrpt 0.263 #/op
BenchmarkSumAggregation2.test:instructions foldDouble
true 8 thrpt 148330.222 #/op
BenchmarkSumAggregation2.test:stalled-cycles-backend foldDouble
true 8 thrpt 21706.154 #/op
BenchmarkSumAggregation2.test:stalled-cycles-frontend foldDouble
true 8 thrpt 302.209 #/op
```
--
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.
To unsubscribe, e-mail: commits-unsubscr...@pinot.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@pinot.apache.org
For additional commands, e-mail: commits-h...@pinot.apache.org
