In proposed SortExec changes (apache/datafusion#5292), currently seeing negative performance impacts when sorting by single record batches (when multiple batches are sorted together, we are seeing the expected performance improvements).
This code lifts the sort_batch method from SortExec and runs various batch types
(e.g. "mixed-tuple") and sizes to examine the impact of batch size and workload type
on the two sort methods: DynComparator (main), and using the rows format (proposed).
The code then generates the following graphs comparing the sort methods runtimes as batch size increaes (each workload type in a separate graph).
New graphs can be created using cli:
> cargo run --release -- --help
Usage: bench-arrow-sort [OPTIONS] --step-size <STEP_SIZE> --begin <BEGIN> --end <END>
Options:
-c, --case <CASE> [default: mixed-tuple] [possible values: utf8-tuple, mixed-tuple, dictionary-tuple, mixed-dictionary-tuple]
-s, --step-size <STEP_SIZE> how many rows to increment batch size by every time
-b, --begin <BEGIN> how many rows to start with
-e, --end <END> end after we've tested batches with this number of rows
-i, --iters <ITERS> how many iterations to run. raise this to decrease variance [default: 50]
-h, --help Print help