1388722b81
This PR adds basic MLlib infrastructure to run some benchmarks against ML pipelines. There are 2 ways to describe and run ML pipelines: - programatically, in scala (see MLBenchmarks.scala) - using a simple YAML file (see mllib-small.yaml for an example) The YAML approach is preferred because it generates programmatically the cartesian product of all the experiments to run and validates the types of the objects in the yaml file. In both cases, all the ML experiments are standard benchmarks. This PR also moves some code in `Benchmark.scala` : the current code generates path-dependent structural signatures and confuses intellij. It does not include tests, but some small benchmarks can be run locally against a spark 2 installation: ``` $SPARK_HOME/bin/spark-shell --jars $PWD/target/scala-2.10/spark-sql-perf-assembly-0.4.9-SNAPSHOT.jar ``` and then: ```scala com.databricks.spark.sql.perf.mllib.MLLib.run(yamlFile="src/main/scala/configs/mllib-small.yaml") ``` Author: Timothy Hunter <timhunter@databricks.com> Closes #69 from thunterdb/1605-mllib2. |
||
|---|---|---|
| bin | ||
| build | ||
| dev | ||
| project | ||
| src | ||
| .gitignore | ||
| .travis.yml | ||
| build.sbt | ||
| LICENSE | ||
| README.md | ||
| version.sbt | ||
Spark SQL Performance Tests
This is a performance testing framework for Spark SQL in Apache Spark 1.6+.
Note: This README is still under development. Please also check our source code for more information.
Quick Start
$ bin/run --help
spark-sql-perf 0.2.0
Usage: spark-sql-perf [options]
-b <value> | --benchmark <value>
the name of the benchmark to run
-f <value> | --filter <value>
a filter on the name of the queries to run
-i <value> | --iterations <value>
the number of iterations to run
--help
prints this usage text
$ bin/run --benchmark DatasetPerformance
TPC-DS
How to use it
The rest of document will use TPC-DS benchmark as an example. We will add contents to explain how to use other benchmarks add the support of a new benchmark dataset in future.
Setup a benchmark
Before running any query, a dataset needs to be setup by creating a Benchmark object. Generating
the TPCDS data requires dsdgen built and available on the machines. We have a fork of dsdgen that
you will need. It can be found here.
import com.databricks.spark.sql.perf.tpcds.Tables
// Tables in TPC-DS benchmark used by experiments.
// dsdgenDir is the location of dsdgen tool installed in your machines.
val tables = new Tables(sqlContext, dsdgenDir, scaleFactor)
// Generate data.
tables.genData(location, format, overwrite, partitionTables, useDoubleForDecimal, clusterByPartitionColumns, filterOutNullPartitionValues)
// Create metastore tables in a specified database for your data.
// Once tables are created, the current database will be switched to the specified database.
tables.createExternalTables(location, format, databaseName, overwrite)
// Or, if you want to create temporary tables
tables.createTemporaryTables(location, format)
// Setup TPC-DS experiment
import com.databricks.spark.sql.perf.tpcds.TPCDS
val tpcds = new TPCDS (sqlContext = sqlContext)
Run benchmarking queries
After setup, users can use runExperiment function to run benchmarking queries and record query execution time. Taking TPC-DS as an example, you can start an experiment by using
val experiment = tpcds.runExperiment(tpcds.interactiveQueries)
For every experiment run (i.e. every call of runExperiment), Spark SQL Perf will use the timestamp of the start time to identify this experiment. Performance results will be stored in the sub-dir named by the timestamp in the given resultsLocation (for example results/1429213883272). The performance results are stored in the JSON format.
Retrieve results
While the experiment is running you can use experiment.html to list the status. Once the experiment is complete, the results will be saved to the table sqlPerformance in json.
// Get all experiments results.
tpcds.createResultsTable()
sqlContext.table("sqlPerformance")
// Get the result of a particular run by specifying the timestamp of that run.
sqlContext.table("sqlPerformance").filter("timestamp = 1429132621024")