2.5. Benchmark Driver

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

Download presto-benchmark-driver-0.220-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 fileformats and legacy_orc. Thefile_formats suite will run queries with names matching the regular expressionsingle. or tpch_. in all schemas matching the regular expressiontpchsf.*._.?. The legacyorc suite adds a session property todisable 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 filewithout the extension.

Output

The benchmark driver will measure the wall time, total CPU time used byall Presto processes and the CPU time used by the query. For each timing, thedriver reports median, mean and standard deviation of the query runs. Thedifference between process and query CPU times is the query overhead, whichis normally from garbage collections. The following is the output from thefile_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 thestandard TSV that the driver outputs.

The driver can add additional columns to the output by extracting values fromthe schema name or SQL files. In the suite file above, the schema namescontain named regular expression capturing groups for compression,format, and scale, so if we ran the queries in a catalog containing theschemas tpch_sf100_orc_none, tpch_sf100_orc_snappy, andtpch_sf100_orc_zlib, we get the above output.

Another way to create additional output columns is by adding tags to theSQL 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 controlwhich suites and queries to run, the number of warm-up runs and the numberof measurement runs. All of the command line arguments can be seen with the—help option.