yordan-pavlov commented on a change in pull request #8630:
URL: https://github.com/apache/arrow/pull/8630#discussion_r520898403
##########
File path: rust/arrow/benches/filter_kernels.rs
##########
@@ -14,137 +14,136 @@
// KIND, either express or implied. See the License for the
// specific language governing permissions and limitations
// under the License.
+extern crate arrow;
+
+use rand::{
+ distributions::{Alphanumeric, Standard},
+ prelude::Distribution,
+ Rng,
+};
use arrow::array::*;
-use arrow::compute::{filter, FilterContext};
+use arrow::compute::{build_filter, filter};
use arrow::datatypes::ArrowNumericType;
+use arrow::datatypes::{Float32Type, UInt8Type};
+
use criterion::{criterion_group, criterion_main, Criterion};
-fn create_primitive_array<T, F>(size: usize, value_fn: F) -> PrimitiveArray<T>
+fn create_primitive_array<T>(size: usize, null_density: f32) ->
PrimitiveArray<T>
where
T: ArrowNumericType,
- F: Fn(usize) -> T::Native,
+ Standard: Distribution<T::Native>,
{
+ // use random numbers to avoid spurious compiler optimizations wrt to
branching
+ let mut rng = rand::thread_rng();
let mut builder = PrimitiveArray::<T>::builder(size);
- for i in 0..size {
- builder.append_value(value_fn(i)).unwrap();
+
+ for _ in 0..size {
+ if rng.gen::<f32>() < null_density {
+ builder.append_null().unwrap();
+ } else {
+ builder.append_value(rng.gen()).unwrap();
+ }
}
builder.finish()
}
-fn create_u8_array_with_nulls(size: usize) -> UInt8Array {
- let mut builder = UInt8Builder::new(size);
- for i in 0..size {
- if i % 2 == 0 {
- builder.append_value(1).unwrap();
- } else {
+fn create_string_array(size: usize, null_density: f32) -> StringArray {
+ // use random numbers to avoid spurious compiler optimizations wrt to
branching
+ let mut rng = rand::thread_rng();
+ let mut builder = StringBuilder::new(size);
+
+ for _ in 0..size {
+ if rng.gen::<f32>() < null_density {
builder.append_null().unwrap();
+ } else {
+ let value =
rng.sample_iter(&Alphanumeric).take(10).collect::<String>();
+ builder.append_value(&value).unwrap();
}
}
builder.finish()
}
-fn create_bool_array<F>(size: usize, value_fn: F) -> BooleanArray
-where
- F: Fn(usize) -> bool,
-{
+fn create_bool_array(size: usize, trues_density: f32) -> BooleanArray {
+ let mut rng = rand::thread_rng();
let mut builder = BooleanBuilder::new(size);
- for i in 0..size {
- builder.append_value(value_fn(i)).unwrap();
+ for _ in 0..size {
+ let value = rng.gen::<f32>() < trues_density;
Review comment:
I wonder if this is the reason for the inconsistency in the benchmark
results - usually a highly selective filter, where most filter bits are 0s and
only a small number of values are selected / copied to the output array will
always have the best performance because most batches of filter bits can be
skipped quickly and only a few values are copied to the output array;
this relationship can clearly be seen in the benchmark results listed in
this earlier PR https://github.com/apache/arrow/pull/7798;
however in the benchmark results listed in the description of this PR in
many cases the opposite is true - low selectivity filter benchmarks achieve
much better performance than their high selectivity counterparts; I wonder
what's the reason for that?
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
