Closed GregoryKimball closed 1 year ago
My first thought was this was a locale issue. but the scale of the scientific notation is wrong for that.
Since we see the correct values being printed to the screen, it stands that it is an issue in fmt
or csv_printer
. But I am have been unable to reproduce this locally with fmt 7 or fmt 9.
I have been able to reproduce this issue by building the container locally that @GregoryKimball was running.
Still investigating what the differences are between the two builds to cause the failure.
Some of the libcudf benchmarks using nvbench generate invalid JSON and CSV files, even though the console output looks normal. One example is JOIN_NVBENCH, which shows the correct "Samples" value in the JSON and CSV output, but all of the timing data is set to invalid values like
0e-20
,0e-18
etc. I'm not certain that benchmark skipping is the trigger for this behavior and I will update this issue when I collect more evidence.Repro: Build libcudf benchmarks and run the following command.
console:
"output.csv"