zhli1142015 commented on code in PR #12101:
URL: https://github.com/apache/gluten/pull/12101#discussion_r3265402305


##########
cpp/velox/benchmarks/VeloxBatchResizerBenchmark.cc:
##########
@@ -0,0 +1,568 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements.  See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License.  You may obtain a copy of the License at
+ *
+ *    http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <limits>
+#include <memory>
+#include <optional>
+#include <string>
+#include <vector>
+
+#include <arrow/util/bit_util.h>
+#include <arrow/util/bitmap_ops.h>
+#include <benchmark/benchmark.h>
+
+#include "memory/ColumnarBatchIterator.h"
+#include "memory/VeloxColumnarBatch.h"
+#include "shuffle/Payload.h"
+#include "utils/Exception.h"
+#include "utils/VeloxBatchResizer.h"
+#include "velox/common/memory/Memory.h"
+#include "velox/vector/ComplexVector.h"
+#include "velox/vector/FlatVector.h"
+
+using namespace facebook::velox;
+
+namespace gluten {
+namespace {
+
+constexpr int32_t kInputBatches = 64;
+constexpr int32_t kRowsPerBatch = 64;
+constexpr int32_t kTotalRows = kInputBatches * kRowsPerBatch;
+constexpr int64_t kPreferredBatchBytes = std::numeric_limits<int64_t>::max();

Review Comment:
   
      I added two fallback-focused benchmark scenarios for the copyRanges 
default-on path:
   
      - `DictionaryHeavy_64x64`
      - `ConstantHeavy_64x64`
   
      Each scenario compares:
   
      1. `BM_VeloxBatchResizerFallbackAppendOptOutBaseline`: 
`enableCopyRanges=false`, the old incremental append path.
      2. `BM_VeloxBatchResizerDefaultCopyRangesFallback`: default 
`enableCopyRanges=true`; dictionary/constant inputs are unsupported by child 
`copyRanges`, so   they fall back to `RowVector::copy` inside the new 
collect-and-copy path.
   
      Benchmark command:
   
      ```bash
      /tmp/gluten-cpp-ut-build/velox/benchmarks/velox_batch_resizer_benchmark \
        --benchmark_min_time=0.05s \
        --benchmark_repetitions=3 \
        --benchmark_format=json \
        > /tmp/velox_batch_resizer_perf.json
   
     Environment:
   
      - CPU: 32 logical CPUs @ ~2995 MHz
      - Caches: L1D 48 KiB x16, L1I 32 KiB x16, L2 2 MiB x16, L3 36 MiB x1
      - Note: ASLR was enabled, so the numbers may contain normal benchmark 
noise.
   
     Dense flat scenarios, CPU time averaged over 3 repetitions:
   
     
┌──────────────────┬─────────────────────────┬────────────────────┬─────────┬─────────────────────────┬─────────────────────────────┐
     │ Scenario         │ Append opt-out baseline │ Default copyRanges │ 
Speedup │ Direct child copyRanges │ Raw payload bulk-copy model │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ Mixed_64x64      │ 94.3 us                 │ 11.3 us            │ 8.35x  
 │ 14.4 us                 │ 34.8 us                     │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ Mixed_16x256     │ 31.7 us                 │ 6.3 us             │ 5.01x  
 │ 5.3 us                  │ 11.2 us                     │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ Mixed_256x16     │ 234.2 us                │ 48.7 us            │ 4.81x  
 │ 30.6 us                 │ 119.2 us                    │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ Fixed2_64x64     │ 29.3 us                 │ 5.4 us             │ 5.43x  
 │ 2.1 us                  │ 13.4 us                     │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ Fixed16_64x64    │ 123.5 us                │ 26.7 us            │ 4.62x  
 │ 18.6 us                 │ 95.7 us                     │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ LongString_64x64 │ 33.3 us                 │ 7.2 us             │ 4.61x  
 │ 4.1 us                  │ 15.7 us                     │
     
├──────────────────┼─────────────────────────┼────────────────────┼─────────┼─────────────────────────┼─────────────────────────────┤
     │ BoolHeavy_64x64  │ 73.4 us                 │ 10.8 us            │ 6.78x  
 │ 5.4 us                  │ 38.1 us                     │
     
└──────────────────┴─────────────────────────┴────────────────────┴─────────┴─────────────────────────┴─────────────────────────────┘
   
     Fallback scenarios requested by the review comment, CPU time averaged over 
3 repetitions:
   
     
┌───────────────────────┬─────────────────────────┬─────────────────────────────┬──────────────┐
     │ Scenario              │ Append opt-out baseline │ Default copyRanges 
fallback │ Result       │
     
├───────────────────────┼─────────────────────────┼─────────────────────────────┼──────────────┤
     │ DictionaryHeavy_64x64 │ 102.7 us                │ 77.7 us                
     │ 1.32x faster │
     
├───────────────────────┼─────────────────────────┼─────────────────────────────┼──────────────┤
     │ ConstantHeavy_64x64   │ 67.8 us                 │ 42.0 us                
     │ 1.61x faster │
     
└───────────────────────┴─────────────────────────┴─────────────────────────────┴──────────────┘
   
     The fallback numbers are not regressions. The reason is that 
enableCopyRanges=true fallback is not the same as the old 
enableCopyRanges=false append path.
   
     With enableCopyRanges=false, VeloxBatchResizer creates an empty output 
RowVector and incrementally appends each small input batch via  
buffer->append(input). That path grows and updates the output vector 
batch-by-batch.
   
     With enableCopyRanges=true, even when a batch is not eligible for 
child-vector copyRanges because it contains dictionary or constant children, 
the resizer   first collects the small input batches, computes the final 
totalRows, allocates/resizes the output RowVector once, and then copies each 
unsupported input   into the pre-sized output using:
   
      buffer->copy(input.get(), offset, 0, input->size());
   
     So these fallback benchmarks still benefit from the new collect-and-copy 
structure:
   
      - one output allocation / resize for the final row count;
      - fixed-offset RowVector::copy into a pre-sized output;
      - less incremental append bookkeeping;
      - Velox copy still has encoded-vector handling for dictionary and 
constant inputs, even though child copyRanges is not used.
   
     This explains why the default copyRanges-enabled fallback path is faster 
than the append opt-out baseline in these two scenarios. The important 
validation   point for the review is that dictionary-heavy and constant-heavy 
inputs do not show a perf regression after enabling copyRanges by default.



-- 
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: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to