layout | title | nav_order | parent |
---|---|---|---|
page |
Micro Benchmarks for Velox Backend |
3 |
Developer Overview |
This document explains how to use the existing micro benchmark template in Gluten Cpp.
A micro benchmark for Velox backend is provided in Gluten Cpp to simulate the execution of a first or middle stage in Spark. It serves as a more convenient alternative to debug in Gluten Cpp comparing with directly debugging in a Spark job. Developers can use it to create their own workloads, debug in native process, profile the hotspot and do optimizations.
To simulate a first stage, you need to dump the Substrait plan and input split info into two JSON files. The input URIs of the splits should be existing file locations, which can be either local or HDFS paths.
To simulate a middle stage, in addition to the JSON file, you also need to save the input data of this stage into Parquet files. The benchmark will load the data into Arrow format, then add Arrow2Velox to feed the data into Velox pipeline to reproduce the reducer stage. Shuffle exchange is not included.
Please refer to the sections below to learn how to dump the Substrait plan and create the input data files.
To run a micro benchmark, user should provide one file that contains the Substrait plan in JSON format, and optional one or more input data files in parquet format. The commands below help to generate example input files:
cd /path/to/gluten/
./dev/buildbundle-veloxbe.sh --build_tests=ON --build_benchmarks=ON
# Run test to generate input data files. If you are using spark 3.3, replace -Pspark-3.2 with -Pspark-3.3, If you are using uniffle, replace -Pceleborn with -Puniffle
mvn test -Pspark-3.2 -Pbackends-velox -Pcelenborn -pl backends-velox -am \
-DtagsToInclude="org.apache.gluten.tags.GenerateExample" -Dtest=none -DfailIfNoTests=false -Dexec.skip
The generated example files are placed in gluten/backends-velox:
$ tree gluten/backends-velox/generated-native-benchmark/
gluten/backends-velox/generated-native-benchmark/
├── example.json
├── example_lineitem
│ ├── part-00000-3ec19189-d20e-4240-85ae-88631d46b612-c000.snappy.parquet
│ └── _SUCCESS
└── example_orders
├── part-00000-1e66fb98-4dd6-47a6-8679-8625dbc437ee-c000.snappy.parquet
└── _SUCCESS
Run micro benchmark with the generated files as input. You need to specify the absolute path to the input files:
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /home/sparkuser/github/apache/incubator-gluten/backends-velox/generated-native-benchmark/example.json \
--data /home/sparkuser/github/apache/incubator-gluten/backends-velox/generated-native-benchmark/example_orders/part-00000-1e66fb98-4dd6-47a6-8679-8625dbc437ee-c000.snappy.parquet,\
/home/sparkuser/github/apache/incubator-gluten/backends-velox/generated-native-benchmark/example_lineitem/part-00000-3ec19189-d20e-4240-85ae-88631d46b612-c000.snappy.parquet \
--threads 1 --iterations 1 --noprint-result
The output should be like:
2022-11-18T16:49:56+08:00
Running ./generic_benchmark
Run on (192 X 3800 MHz CPU s)
CPU Caches:
L1 Data 48 KiB (x96)
L1 Instruction 32 KiB (x96)
L2 Unified 2048 KiB (x96)
L3 Unified 99840 KiB (x2)
Load Average: 0.28, 1.17, 1.59
***WARNING*** CPU scaling is enabled, the benchmark real time measurements may be noisy and will incur extra overhead.
-- Project[expressions: (n3_0:BIGINT, ROW["n1_0"]), (n3_1:VARCHAR, ROW["n1_1"])] -> n3_0:BIGINT, n3_1:VARCHAR
Output: 535 rows (65.81KB, 1 batches), Cpu time: 36.33us, Blocked wall time: 0ns, Peak memory: 1.00MB, Memory allocations: 3, Threads: 1
queuedWallNanos sum: 2.00us, count: 2, min: 0ns, max: 2.00us
-- HashJoin[RIGHT SEMI (FILTER) n0_0=n1_0] -> n1_0:BIGINT, n1_1:VARCHAR
Output: 535 rows (65.81KB, 1 batches), Cpu time: 191.56us, Blocked wall time: 0ns, Peak memory: 2.00MB, Memory allocations: 8
HashBuild: Input: 582 rows (16.45KB, 1 batches), Output: 0 rows (0B, 0 batches), Cpu time: 1.84us, Blocked wall time: 0ns, Peak memory: 1.00MB, Memory allocations: 3, Threads: 1
distinctKey0 sum: 583, count: 1, min: 583, max: 583
queuedWallNanos sum: 0ns, count: 1, min: 0ns, max: 0ns
rangeKey0 sum: 59748, count: 1, min: 59748, max: 59748
HashProbe: Input: 37897 rows (296.07KB, 1 batches), Output: 535 rows (65.81KB, 1 batches), Cpu time: 189.71us, Blocked wall time: 0ns, Peak memory: 1.00MB, Memory allocations: 5, Threads: 1
queuedWallNanos sum: 0ns, count: 1, min: 0ns, max: 0ns
-- ArrowStream[] -> n0_0:BIGINT
Input: 0 rows (0B, 0 batches), Output: 37897 rows (296.07KB, 1 batches), Cpu time: 1.29ms, Blocked wall time: 0ns, Peak memory: 0B, Memory allocations: 0, Threads: 1
-- ArrowStream[] -> n1_0:BIGINT, n1_1:VARCHAR
Input: 0 rows (0B, 0 batches), Output: 582 rows (16.45KB, 1 batches), Cpu time: 894.22us, Blocked wall time: 0ns, Peak memory: 0B, Memory allocations: 0, Threads: 1
-----------------------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
-----------------------------------------------------------------------------------------------------------------------------
InputFromBatchVector/iterations:1/process_time/real_time/threads:1 41304520 ns 23740340 ns 1 collect_batch_time=34.7812M elapsed_time=41.3113M
First, build Gluten with --build_benchmarks=ON
.
cd /path/to/gluten/
./dev/buildbundle-veloxbe.sh --build_benchmarks=ON
# For debugging purpose, rebuild Gluten with build type `Debug`.
./dev/buildbundle-veloxbe.sh --build_benchmarks=ON --build_type=Debug
First, get the Stage Id from spark UI for the stage you want to simulate. And then re-run the query with below configurations to dump the inputs to micro benchmark.
Parameters | Description | Recommend Setting |
---|---|---|
spark.gluten.sql.benchmark_task.taskId | Comma-separated string to specify the Task IDs to dump. If it's set, spark.gluten.sql.benchmark_task.stageId and spark.gluten.sql.benchmark_task.partitionId will be ignored. |
Comma-separated string of task IDs. Empty by default. |
spark.gluten.sql.benchmark_task.stageId | Spark stage ID. | Target stage ID |
spark.gluten.sql.benchmark_task.partitionId | Comma-separated string to specify the Partition IDs in a stage to dump. Must be specified together with spark.gluten.sql.benchmark_task.stageId . Empty by default, meaning all partitions of this stage will be dumped. To identify the partition ID, navigate to the Stage tab in the Spark UI and locate it under the Index column. |
Comma-separated string of partition IDs. Empty by default. |
spark.gluten.saveDir | Directory to save the inputs to micro benchmark, should exist and be empty. | /path/to/saveDir |
Check the files in spark.gluten.saveDir
. If the simulated stage is a first stage, you will get 3
or 4 types of dumped file:
- Configuration file: INI formatted, file name
conf_[stageId]_[partitionId].ini
. Contains the configurations to init Velox backend and runtime session. - Plan file: JSON formatted, file name
plan_[stageId]_[partitionId].json
. Contains the substrait plan to the stage, without input file splits. - Split file: JSON formatted, file name
split_[stageId]_[partitionId]_[splitIndex].json
. There can be more than one split file in a first stage task. Contains the substrait plan piece to the input file splits. - Data file(optional): Parquet formatted, file
name
data_[stageId]_[partitionId]_[iteratorIndex].parquet
. If the first stage contains one or more BHJ operators, there can be one or more input data files. The input data files of a first stage will be loaded as iterators to serve as the inputs for the pipeline:
"localFiles": {
"items": [
{
"uriFile": "iterator:0"
}
]
}
Run benchmark. By default, the result will be printed to stdout. You can use --noprint-result
to
suppress this output.
Sample command:
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--conf /absolute_path/to/conf_[stageId]_[partitionId].ini \
--plan /absolute_path/to/plan_[stageId]_[partitionId].json \
--split /absolut_path/to/split_[stageId]_[partitionId]_0.json,/absolut_path/to/split_[stageId]_[partitionId]_1.json \
--threads 1 --noprint-result
# If the stage requires data files, use --data-file to specify the absolute path.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--conf /absolute_path/to/conf_[stageId]_[partitionId].ini \
--plan /absolute_path/to/plan_[stageId]_[partitionId].json \
--split /absolut_path/to/split_[stageId]_[partitionId]_0.json,/absolut_path/to/split_[stageId]_[partitionId]_1.json \
--data /absolut_path/to/data_[stageId]_[partitionId]_0.parquet,/absolut_path/to/data_[stageId]_[partitionId]_1.parquet \
--threads 1 --noprint-result
If the simulated stage is a middle stage, which means pure shuffle stage, you will get 3 types of dumped file:
- Configuration file: INI formatted, file name
conf_[stageId]_[partitionId].ini
. Contains the configurations to init Velox backend and runtime session. - Plan file: JSON formatted, file name
plan_[stageId]_[partitionId].json
. Contains the substrait plan to the stage. - Data file: Parquet formatted, file name
data_[stageId]_[partitionId]_[iteratorIndex].parquet
. There can be more than one input data file in a middle stage task. The input data files of a middle stage will be loaded as iterators to serve as the inputs for the pipeline:
"localFiles": {
"items": [
{
"uriFile": "iterator:0"
}
]
}
Sample command:
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--conf /absolute_path/to/conf_[stageId]_[partitionId].ini \
--plan /absolute_path/to/plan_[stageId]_[partitionId].json \
--data /absolut_path/to/data_[stageId]_[partitionId]_0.parquet,/absolut_path/to/data_[stageId]_[partitionId]_1.parquet \
--threads 1 --noprint-result
For some complex queries, stageId may cannot represent the Substrait plan input, please get the taskId from spark UI, and get your target parquet from saveDir.
In this example, only one partition input with partition id 2, taskId is 36, iterator length is 2.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /absolute_path/to/complex_plan.json \
--data /absolute_path/to/data_36_2_0.parquet,/absolute_path/to/data_36_2_1.parquet \
--threads 1 --noprint-result
You can save the output to a parquet file via --save-output <output>
Note: 1. This option cannot be used together with --with-shuffle
. 2. This option cannot be used
for write tasks. Please refer to section Simulate write tasks for more
details.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /absolute_path/to/plan.json \
--data /absolute_path/to/data.parquet
--threads 1 --noprint-result --save-output /absolute_path/to/result.parquet
You can add the shuffle write process at the end of the pipeline via --with-shuffle
Note: 1. This option cannot be used together with --save-output
. 2. This option cannot be used
for write tasks. Please refer to section Simulate write tasks for more
details.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /absolute_path/to/plan.json \
--split /absolute_path/to/split.json \
--threads 1 --noprint-result --with-shuffle
Developers can leverage the --with-shuffle
option to benchmark the shuffle-write process by creating
a simple pipeline of table scan + shuffle write
in Gluten. This can be done by dumping the micro benchmark
inputs from a first stage. The steps are demonstrated as below:
- Start spark-shell or pyspark
We need to set spark.gluten.sql.benchmark_task.stageId
and spark.gluten.saveDir
to dump the inputs.
Normally, the stage id should be greater than 0. You can run the command in step 2 in advance to get the
right stage id in your case. We shall set spark.default.parallelism
to 1 and spark.sql.files.maxPartitionBytes
large enough to make sure there will be only 1 task in the first stage.
# Start pyspark
./bin/pyspark --master local[*] \
--conf spark.gluten.sql.benchmark_task.stageId=1 \
--conf spark.gluten.saveDir=/path/to/saveDir \
--conf spark.default.parallelism=1 \
--conf spark.sql.files.maxPartitionBytes=10g
... # omit other spark & gluten config
- Run the table-scan command to dump the plan for the first stage
If simulating single or round-robin partitioning, the first stage can only have the table scan operator.
>>> spark.read.format("parquet").load("file:///example.parquet").show()
If simulating hash partitioning, there will be a projection for generating the hash partitioning key.
Therefore we need to explicitly run the repartition
to generate the scan + project
pipeline for the first stage.
Note that using different number of shuffle partitions here doesn't change the generated pipeline.
>>> spark.read.format("parquet").load("file:///example.parquet").repartition(10, "key1", "key2").show()
Simuating range partitioning is not supported.
- Run the micro benchmark with dumped inputs
General configurations for shuffle write:
-
--with-shuffle
: Add shuffle write process at the end of the pipeline -
--shuffle-writer
: Specify shuffle writer type. Valid options are sort and hash. Default is hash. -
--partitioning
: Specify partitioning type. Valid options are rr, hash and single. Defualt is rr. The partitioning type should match the command in step 2. -
--shuffle-partitions
: Specify number of shuffle partitions. -
--compression
: By default, the compression codec for shuffle outputs is lz4. You can switch to other compression codecs or use hardware accelerators Valid options are: lz4, zstd, qat-gzip, qat-zstd and iaa-gzip. The compression levels are fixed (use default compression level 1).Note using QAT or IAA codec requires Gluten cpp is built with these features. Please check the corresponding section in Velox document first for how to setup, build and enable these features in Gluten. For QAT support, please check Intel® QuickAssist Technology (QAT) support. For IAA support, please check Intel® In-memory Analytics Accelerator (IAA/IAX) support
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /path/to/saveDir/plan_1_0.json \
--conf /path/to/saveDir/conf_1_0.ini \
--split /path/to/saveDir/split_1_0_0.json \
--with-shuffle \
--shuffle-writer sort \
--partitioning hash \
--threads 1
Developers can only run shuffle write task via specifying --run-shuffle
and --data
options.
The parquet format input will be read from arrow-parquet reader and sent to shuffle writer.
The --run-shuffle
option is similar to the --with-shuffle
option, but it doesn't require the plan and split files.
The round-robin partitioner is used by default. Besides, random partitioning can be used for testing purpose.
By specifying option --partitioning random
, the partitioner will generate a random partition id for each row.
To evaluate the shuffle reader performance, developers can set --run-shuffle-read
option to add read process after the write task finishes.
The below command will run shuffle write/read in single thread, using sort shuffle writer with 40000 partitions and random partition id.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--run-shuffle \
--run-shuffle-read \
--data /path/to/input_for_shuffle_write.parquet
--shuffle-writer sort \
--partitioning random \
--shuffle-partitions 40000 \
--threads 1
The output should be like:
-------------------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
-------------------------------------------------------------------------------------------------------------------------
ShuffleWriteRead/iterations:1/process_time/real_time/threads:1 121637629714 ns 121309450910 ns 1 elapsed_time=121.638G read_input_time=25.2637G shuffle_compress_time=10.8311G shuffle_decompress_time=4.04055G shuffle_deserialize_time=7.24289G shuffle_spill_time=0 shuffle_split_time=69.9098G shuffle_write_time=2.03274G
spark.gluten.sql.debug
(debug mode) is set to false by default thereby the google glog levels are limited to only print WARNING
or higher severity logs.
Unless spark.gluten.sql.debug
is set in the INI file via --conf
, the logging behavior is same as debug mode off.
Developers can use --debug-mode
command line flag to turn on debug mode when needed, and set verbosity/severity level via command line flags --v
and --minloglevel
. Note that constructing and deconstructing log strings can be very time-consuming, which may cause benchmark times to be inaccurate.
After enabling the dynamic loading of libhdfs.so at runtime to support HDFS, if you run the benchmark with an HDFS file, you need to set the classpath for Hadoop. You can do this by running
export CLASSPATH=`$HADOOP_HOME/bin/hdfs classpath --glob`
Otherwise, the HDFS connection will fail. If you have replaced ${HADOOP_HOME}/lib/native/libhdfs.so with libhdfs3.so, there is no need to set the CLASSPATH
.
The last operator for a write task is a file write operator, and the output from Velox pipeline only
contains several columns of statistics data. Therefore, specifying
options --with-shuffle
and --save-output
does not take effect. You can specify the output path
for the writer via --write-path
option. Default is /tmp.
cd /path/to/gluten/cpp/build/velox/benchmarks
./generic_benchmark \
--plan /absolute_path/to/plan.json \
--split /absolute_path/to/split.json \
--write-path /absolute_path/<dir>
You can simulate task spilling by specify a memory hard limit from --memory_limit
. By default, spilled files are written to the /tmp
directory.
To simulate real Gluten workloads, which utilize multiple spill directories, set the environment variable GLUTEN_SPARK_LOCAL_DIRS to a comma-separated string.
Please check Simulate Gluten workload with multiple processes and threads for more details.
You can use below command to launch several processes and threads to simulate parallel execution on
Spark. Each thread in the same process will be pinned to the core number starting from --cpu
.
Suppose running on a bare-metal machine with 48C, 2-socket, HT-on, launching below command will utilize all vcores.
processes=24 # Same value of spark.executor.instances
threads=8 # Same value of spark.executor.cores
for ((i=0; i<${processes}; i++)); do
./generic_benchmark --plan /path/to/plan.json --split /path/to/split.json --noprint-result --threads $threads --cpu $((i*threads)) &
done
To include the shuffle write process or trigger spilling via --memory-limit
,
you can specify multiple directories by setting the GLUTEN_SPARK_LOCAL_DIRS
environment variable
to a comma-separated string. This will distribute the I/O load across multiple disks, similar to how it works for Gluten workloads.
Temporary subdirectories will be created under each specified directory at runtime and will be automatically deleted if the process completes normally.
mkdir -p {/data1,/data2,/data3}/tmp # Make sure each directory has been already created.
export GLUTEN_SPARK_LOCAL_DIRS=/data1/tmp,/data2/tmp,/data3/tmp
processes=24 # Same value of spark.executor.instances
threads=8 # Same value of spark.executor.cores
for ((i=0; i<${processes}; i++)); do
./generic_benchmark --plan /path/to/plan.json --split /path/to/split.json --noprint-result --with-shuffle --threads $threads --cpu $((i*threads)) &
done
We also provide some example inputs in cpp/velox/benchmarks/data. E.g. Files under generic_q5 simulates a first-stage in TPCH Q5, which has a heavy table scan. You can follow below steps to run this example.
Open generic_q5/q5_first_stage_0_split.json
with file editor. Search for "uriFile": "LINEITEM"
and replace LINEITEM
with the URI to one
partition file in lineitem. In the next line, replace the number in "length": "..."
with the
actual file length. Suppose you are using the provided small TPCH table
in cpp/velox/benchmarks/data/tpch_sf10m, the replaced
JSON should be like:
{
"items": [
{
"uriFile": "file:///path/to/gluten/cpp/velox/benchmarks/data/tpch_sf10m/lineitem/part-00000-6c374e0a-7d76-401b-8458-a8e31f8ab704-c000.snappy.parquet",
"length": "1863237",
"parquet": {}
}
]
}
- Launch multiple processes and multiple threads. Set
GLUTEN_SPARK_LOCAL_DIRS
and add--with-shuffle
to the command.
mkdir -p {/data1,/data2,/data3}/tmp # Make sure each directory has been already created.
export GLUTEN_SPARK_LOCAL_DIRS=/data1/tmp,/data2/tmp,/data3/tmp
processes=24 # Same value of spark.executor.instances
threads=8 # Same value of spark.executor.cores
for ((i=0; i<${processes}; i++)); do
./generic_benchmark --plan /path/to/gluten/cpp/velox/benchmarks/data/generic_q5/q5_first_stage_0.json --split /path/to/gluten/cpp/velox/benchmarks/data/generic_q5/q5_first_stage_0_split.json --noprint-result --with-shuffle --threads $threads --cpu $((i*threads)) &
done >stdout.log 2>stderr.log
You can find the "elapsed_time" and other metrics in stdout.log. In below output, the "elapsed_time" is ~10.75s. If you run TPCH Q5 with Gluten on Spark, a single task in the same Spark stage should take about the same time.
------------------------------------------------------------------------------------------------------------------
Benchmark Time CPU Iterations UserCounters...
------------------------------------------------------------------------------------------------------------------
SkipInput/iterations:1/process_time/real_time/threads:8 1317255379 ns 10061941861 ns 8 collect_batch_time=0 elapsed_time=10.7563G shuffle_compress_time=4.19964G shuffle_spill_time=0 shuffle_split_time=0 shuffle_write_time=1.91651G