[PERF] Performance impact of mixed_type_as_string JSON reader option in reading JSON lines

[shrshi]

This report presents some preliminary findings on the normalization, mixed types handling, byte range reading, and error recovery handling options for JSON lines input. Given a valid JSON input string i.e. with no modifications to the data generation and reading a single chunk i.e. the byte range consists of all records, we expect to see no significant impact of enabling these options.

Benchmarks were run on A100 80GB GPU, with all combinations of the above options being enabled/disabled, and a performance degradation of 98% was observed on enabling mixed_type_as_string (keeping normalize_single_quotes=NO row_selection=ALL recovery_mode=RECOVER_WITH_NULL constant between the two experiments). Refer to figure for performance comparison.

To investigate the impact of mixed_type_as_string being enabled, the benchmark was profiled with --axis normalize_single_quotes=NO --axis row_selection=ALL --axis mixed_types_as_string=YES --axis recovery_mode=RECOVER_WITH_NULL.
The infer_column_type_kernel appears to be the bottleneck due to stalled warps resulting in achieved occupancy of 9.3% (nsys and ncu profiles below).

Next steps

• More investigation required to improve performance of reader with mixed types option

Related information

• Feature request to improve JSON reader benchmarks
• Benchmarking PR

[GregoryKimball]

@karthikeyann, would you please work with @shrshi on this performance hotspot?

[karthikeyann]

Provided @shrshi a minor patch to benchmark. This kernel could be skipped entirely for those cases.

[GregoryKimball]

Closed by #15236