Firebird OLTP-EMULator test

LTP-EMUL for Firebird Database  version >= 2.5   

Get actual version from here:

* if you are on WINDOWS:
    git clone --config core.autocrlf=true https://github.com/FirebirdSQL/oltp-emul

* if you are on POSIX:
    git clone --config core.autocrlf=false https://github.com/FirebirdSQL/oltp-emul

===============================================================================

This is a Quick-Start guide for the test that emulates OLTP workload on FB 2.5 and above.
Please READ all this file carefully before making any attempt to run the test.

In case of any questions feel free to contact: pz@ibase.ru

(C) Pavel Zotov, Moscow, Russia. 2014-2019.

===============================================================================

1. Ensure that you install Firebird client on the host from where you plan to run this test.
   Following binaries must also be on this machine:
  • isql
  • fbsvcmgr

2. Change to 'src' folder and find a test configuration file that appropriates to the version
   of Firebird and OS:
   -------------------------------------------------------------------------
   ! Major version of ! Operating System where ! Name of OLTP-EMUL         !
   ! Firebird server  ! ISQL sessions will run ! test config file          !
   !------------------!------------------------!---------------------------!
   !       2.5        !        Windows         ! oltp25_config.win.default !
   !       2.5        !        Linux           ! oltp25_config.nix.default !
   !       3.0        !        Windows         ! oltp30_config.win.default !
   !       3.0        !        Linux           ! oltp30_config.nix.default !
   !       4.0        !        Windows         ! oltp40_config.win.default !
   !       4.0        !        Linux           ! oltp40_config.nix.default !
   -------------------------------------------------------------------------


3. Make a copy of this file so that its extension will be without ".default", e.g.:
   copy .\oltp30_config.win.default .\oltp30_config.win
   cp ./oltp30_config.nix.default ./oltp30_config.nix


   Open just created a copy of the configuration file and change settings to be suitable for you.
   Pay attention to the following settings:

   * 'fbc' - path to isql executable on the host where you will launch ISQL sessions.

   *  'clu' - optional parameter that defines 'non-standard' name of binary ISQL: 'isql-fb' or so.
     (actual for installation from POSIX repository when ISQL is renamed to a different name).

   * 'dbnm' - path and name of the database file on the server host. Batch scenario (1run_oltp_emul.bat/.sh) can create the database file if you specify it using the full path and name of the file, but not as existing alias.
      Ensure that the path and name of the database file contain only ASCII characters.
   
   * 'host', 'port', 'usr' and 'pwd' - values for connecting to database, their meaning is obvious.

   * 'tmpdir' - path to the directory where the test will create temporary files for storing ISQL session logs, etc

   * 'init_docs' - how many documents should be created in the database before test start real workload.

   * 'warm_time' and 'test_time' - how long database should be warmed-up and duration of measured workload, in minutes.

   * 'sleep_min', 'sleep_max'  - how long every attachment should be PAUSED between transactions for making workload more realistic. Assigning zero to 'sleep_max' will suppress any pauses and a new transaction will be started immediately after the previous one finished.

   * 'sleep_ddl' - SQL script that contains a declaration of UDF that will be called for PAUSES. 
      This script is OPTIONAL: when commented then pauses will be implemented by SHELL call (from ISQL execution context).
      Name of SHELL command depends on OS: it will be cscript.exe + .vbs on Windows or 'sleep' command on POSIX but in any
      case this will add more workload on the operating system when the bulk of ISQL sessions are at work.
      It is recommended to use UDF invocation for making pauses, and you can use your own UDF or provide by this test.

      The latter is stored in a compressed binary:
  •         Windows: .\util\udf64\SleepUDF.dll.zip 
  •         POSIX: ./util/udf64/SleepUDF.so.bz2. 
      Extract these UDFs:
  •         7zip -x -tzip .\util\udf64\SleepUDF.dll.zip 
  •         bzip2 -dk ./util/udf64/SleepUDF.so.bz2
      - and put .dll /.so into UDF subfolder from your Firebird home directory.
     
   * 'wait_for_copy' - should test scenario make a pause after the test database will be filled up with the required number of documents. Value = 1 will save your time if you plan to launch a test again later: make a copy of the database that will be created and restore from it on 2nd, 3rd, etc launches.

   For the first time, you can assign to 'init_docs' some small value, e.g. 3000 or 5000.
   The test will start with populating data up to value and after this number of documents will be reached, 
   two phases will begin to perform:
  1.    database warm-up during minutes;
  2.    measurement of further business actions during minutes.

4. Windows specifics
   Ensure that any protection software (antivirus or built-in Windows mechanism) does NOT check any type of files in the folder that you will define by 'tmpdir' config parameter. 
   Firebird can create temporary files:
   a) for storing data of GTT (fb_table_*) - in the folder that is defined by searching first non-empty env. variable from the following list: { FIREBIRD_TMP; TMP; TEMP }
   b) for sorting (fb_sort_*) and  storing monitoring snapshots (fb_recbuf_*, fb_blob_*) - in a directory that is defined by 'TempDirectories' parameter from firebird.conf.
   It is recommended that you will set the value of FIREBIRD_TMP variable equal to the value of 'TempDirectories' parameter from firebird.conf and remove any OS/antivirus protection from this folder. Otherwise, you can encounter extremely slow disk operations of all launched ISQL sessions.

5. It is highly recommended to increase values of the following parameters in firebird.conf:
 
  • DefaultDBCachePages
  • LockHashSlots and 
  • LockMemSize

  It is needed to EXPLICITLY SPECIFY parameter FileSystemCacheThreshold so that it will be greater than DefaultDBCachePages.
 The test does not accept firebird.conf with a missed or commented value of this parameter (this does not relate to FB instance which was originated from the POSIX repository and has an actual home folder differs from parent one for parameter 'fbc').

   Parameter DefaultDBCachePages must have a value that leads the total size of page cache to be equal to ~25% of total physical memory.
   Do NOT assign to DefaultDBCachePages values more than 2048 if you plan to test Classic Server or SuperClassic.

   The database always is created with page size 8192, this value is hardcoded.
   For medium workload (about 100 working sessions) following values can be set:

    +----------------------------------------------------------------------------------------------+
    |                           SuperServer 3.0+   |  [Super]Classic 3.0+   |   [Super]Classic 2.5 |
    |----------------------------------------------|------------------------|----------------------|
    |  DefaultDbCachePages      0.25 *
/ 8192      |                1024    |             1024     |
    |  LockHashSlots                  8191         |               16001    |            16001     |
    |  LockMemSize                 4194304         |            16777216    |         16777216     |
    +----------------------------------------------------------------------------------------------+

    = total physical memory


6. Open command interpreter (Windows: "Start/Run/cmd.exe"), change to 'src' directory and run:

   1run_oltp_emul [nostop]
   ###############################

   where:
        = 25, 30 or 40 - major version in simplified form for FB 2.5, 3.0 or 4.0 respectively;
        = number of ISQL sessions which should be launched;
       nostop = (optional) literal argument that forces script to skip any pauses, even if work
             will be impossible (useful when scenario is launched from scheduler)

   If the database does not exist, the script will attempt to create it for you but only if it's specified as a fully qualified filename (not an alias).
   If the database already exists but is empty or its creation was not completed during the previous run then the test will recreate all needed objects.


7. After + minutes test will stop itself, i.e. all ISQL sessions will terminate their job by raising an exception, issuing QUIT statement, and disconnect from the database.
   On Windows, every ISQL window will be closed, so you do not have to close them manually.

   ISQL session which was launched first among all others will make the final report in text format.

   If this ISQL session is running on Windows and setting 'make_html' has value 1, the final report will be created also in HTML format, but in that case, the time of this report creation will be increased on ~2x.

   Change to the folder with the name defined by parameter 'tmpdir' in your oltpNN_config file.
   Name of final report file depends on value of config parameter 'file_name_with_test_params':
   When parameter 'file_name_with_test_params':
   1. Is commented (default), the report will be created with the name: oltp**.report.txt
   2. Has value 'regular' - the report will be in a form that appropriates for 'accumulation' of files
      for further analysis by the simple look on the LIST of these files and found performance troubles;
      Sample of report name when this parameter = 'regular':
      20151102_1448_score_06543_build_31236_ss30__3h00m_100_att_fw__on.txt
   3. Has value 'benchmark' - the report will be in a form that appropriates for comparison of different
      test or database DDL settings.
      Sample of report name when this parameter = 'benchmark':
      ss30_fw_off_split_most__sel_1st_one_index_score_06543_build_31236__3h00m_100_att_20151102_1448.txt

   The final report contains:

   * FB architecture name, database and test settings;
   * overall performance results: total, dynamic for 30-time intervals, performance for each minute and detailed per each unit;
   * when test config setting 'mon_unit_perf' is 1: gathered monitoring data about the performance
     with details down to 1) for FB 2.5 - application units; for FB 3.0 and above - application
     units and tables;
   * exceptions that occurred during the test;
   * database statistics after test finish;
   * report about ratio "total versions" / "total records" for every table with number of records > 0.
   * database validation report (only for tables that are subject for modifications);
   * comparison of firebird.log that was before and after test finish, using standard console utilities
     which present on any version of the underlying OS: fc.exe (on Windows) and diff (on Linux).

   Pay note on performance reports.

   In the first report (which contain rows with text 'interval # NNN, overall') you can
   estimate performance for each of time interval in the column 'CNT_OK_PER_MINUTE': this is
   the number of business per one-minute actions that finished SUCCESSFULLY in bounds of each
   interval.

   In the second report (which contains a row with text "*** OVERALL *** for N minutes") you can estimate the aggregated value of performance for the last three hours or time of actual work. Value in the column "AVG_TIMES_PER_MINUTE" has the same sense: an average number of business actions per minute which finished SUCCESSFULLY, but not split into intervals.

   You can also get these data if connect to the test database and run:

   SQL> select * from srv_mon_perf_dynamic;
   SQL> select * from srv_mon_perf_total;

8. The test will stop itself automatically on every run, you have to specify only two time-based
   values: duration of database warm-up and how long the measure phase should last, in minutes.

   However, if you'll encounter that database become unavailable because of too heavy workload
   you may stop all ISQL sessions almost immediately.
   Default way for this - running batch scenario with name = '1stoptest.tmp' (.bat/.sh) which
   will be created every time on test launch in temporary folder which is defined by test config
   parameter 'tmpdir'.

   *** NOTES ABOUT STOPPING TEST WHEN ISQL SESSIONS ARE LAUNCHED FROM SEVERAL MACHINES ***

   If your planning to launch test from SEVERAL MACHINES, one may to consider another way to 
   premature stop test:
   1) edit your firebird.conf and uncomment setting ExternalFileAccess;
   2) set its value:
      2.1) either to 'Restrict, where is the directory where Firebird process will be able to create a file for an external table,

      2.2) or 'Full'.
   3) open oltp**_config file and find parameter 'use_external_to_stop'. It is COMMENTED by default.
      Uncomment it and set its value to:
      3.1) either ONLY name of external file (without path) - if ExternalFileAccess = Restrict

,
      3.2) or full path and name of file - if ExternalFileAccess = Full.

   Restart FB service. Ensure that the stop-file file specified in oltp**_config is EMPTY before *each*  time you are going to launch this test (you have to manually clear content of this file).
   In case when this file is on Windows host, run Notepad.exe, type any ASCII character, and press LF
   than choose File / Save As. Specify the *same* name, i.e. overwrite this file.


9. If you plan to run this test several times to estimate the effect of changing some of its settings it is recommended that you will do this by starting workload from the same 'point' each time. This means that you might want to create a test database, wait until the test will finish the process of adding the initial number of documents, and then make a test scenario to be PAUSED until you make a copy to that database.
   If this is what you want, change the value of config parameter 'wait_for_copy' to 1.
   After the test will finish, save its report somewhere and restore the test database from the previously created copy.
   If parameter 'use_external_to_stop' is defined, do not forget to make server-side file 'stoptest.txt'
   empty before running new test session!

The short description of the test logic

The idea of the test is based on real business processes of a car-service enterprise:
1. A customer does order (wants we supply him some set of parts); we also can make an order for our purposes;
2. We gather several customer orders and unite them into a single document and send it to the supplier;
3. The supplier sends us an invoice that can fully or partially satisfy our demands.
4. When we receive invoice its content is added to stock remainders. 
   Parts that were ordered by the customer will be immediately reserved for selling.
5. Further, we sell ordered parts and make appropriate write-offs from stock.
6. All previously mentioned operations can be canceled in any step.
7. Each client order and each invoice to the supplier can be paid for 100% or less, and this must be reflected in customer or supplier balance (separately). Each payment operation can be canceled.
8. After the test finish first of the running session invokes reports about performance (total and detailed), list of occurred exceptions, database statistics, the result of DB validation et al.