Benchmark Driver

The benchmark driver can be used to measure the performance of queries in a Presto cluster. We use it to continuously measure the performance of trunk.

Download presto-benchmark-driver-0.268-executable.jar, rename it to presto-benchmark-driver, then make it executable with chmod +x.

Suites

Create a suite.json file:

  1. {
  2. "file_formats": {
  3. "query": ["single_.*", "tpch_.*"],
  4. "schema": [ "tpch_sf(?<scale>.*)_(?<format>.*)_(?<compression>.*?)" ],
  5. "session": {}
  6. },
  7. "legacy_orc": {
  8. "query": ["single_.*", "tpch_.*"],
  9. "schema": [ "tpch_sf(?<scale>.*)_(?<format>orc)_(?<compression>.*?)" ],
  10. "session": {
  11. "hive.optimized_reader_enabled": "false"
  12. }
  13. }
  14. }

This example contains two suites file_formats and legacy_orc. The file_formats suite will run queries with names matching the regular expression single_.* or tpch_.* in all schemas matching the regular expression tpch_sf.*_.*_.*?. The legacy_orc suite adds a session property to disable the optimized ORC reader and only runs in the tpch_sf.*_orc_.*? schema.

Queries

The SQL files are contained in a directory named sql and must have the .sql file extension. The name of the query is the name of the file without the extension.

Output

The benchmark driver will measure the wall time, total CPU time used by all Presto processes and the CPU time used by the query. For each timing, the driver reports median, mean and standard deviation of the query runs. The difference between process and query CPU times is the query overhead, which is normally from garbage collections. The following is the output from the file_formats suite above:

  1. suite query compression format scale wallTimeP50 wallTimeMean wallTimeStd processCpuTimeP50 processCpuTimeMean processCpuTimeStd queryCpuTimeP50 queryCpuTimeMean queryCpuTimeStd
  2. ============ ============== =========== ====== ===== =========== ============ =========== ================= ================== ================= =============== ================ ===============
  3. file_formats single_varchar none orc 100 597 642 101 100840 97180 6373 98296 94610 6628
  4. file_formats single_bigint none orc 100 238 242 12 33930 34050 697 32452 32417 460
  5. file_formats single_varchar snappy orc 100 530 525 14 99440 101320 7713 97317 99139 7682
  6. file_formats single_bigint snappy orc 100 218 238 35 34650 34606 83 33198 33188 83
  7. file_formats single_varchar zlib orc 100 547 543 38 105680 103373 4038 103029 101021 3773
  8. file_formats single_bigint zlib orc 100 282 269 23 38990 39030 282 37574 37496 156

Note that the above output has been reformatted for readability from the standard TSV that the driver outputs.

The driver can add additional columns to the output by extracting values from the schema name or SQL files. In the suite file above, the schema names contain named regular expression capturing groups for compression, format, and scale, so if we ran the queries in a catalog containing the schemas tpch_sf100_orc_none, tpch_sf100_orc_snappy, and tpch_sf100_orc_zlib, we get the above output.

Another way to create additional output columns is by adding tags to the SQL files. For example, the following SQL file declares two tags, projection and filter:

  1. projection=true
  2. filter=false
  3. =================
  4. SELECT SUM(LENGTH(comment))
  5. FROM lineitem

This will cause the driver to output these values for each run of this query.

CLI Arguments

The presto-benchmark-driver program contains many CLI arguments to control which suites and queries to run, the number of warm-up runs and the number of measurement runs. All of the command line arguments can be seen with the --help option.