Query timing

Query timing mk Sun, 10/13/2013 - 14:28

Timing a query execution is supported in multiple ways, but largely depends on what you want to measure. Point-to-point wall clock time, or the actual behavior of kernel operations.

(1) The baseline is to use simple command line tools, such at TIME on Linux to assess the performance of running a script against mclient. Beware that /bin/time and /usr/bin/time are not the same, they mainly measure and report the wall-clock time spent by the given command/process. See their respective man pages for details.

(2) The next approach is to use the "--interactive" option of the mclient tool, which will report on the timing of each individual SQL query in a script in easy human consumable terms. It returns the wall-clock time between sending the query to the server and receiving the first block of answers. Its rendering can be controlled (see mclient).

(3) The query history can also be maintained in a separate log for post analysis. (see description)

Thus, (1) includes everything from loading the mclient binary and starting the mclient process, parsing the query in mclient, sending to the server, having the server execute the query and serialize the result, sending the result back to the client, to the client receiving, parsing and rendering the result, and sending the result to /dev/null ("for free"), to a file (I/O), or to a terminal (scrolling). (2) merely includes the time the server spends on receiving and executing the query and creating the result. The abovementioned costs on the client side to receive, parse, render, etc. the result are excluded. The same holds for (3)

A detailed time of an SQL query can be obtained with prepending the query with the modifier TRACE. It will produce a queryable table with a break down of all relational algebra operations (see TRACE command). The profiling tools stethoscope and tomograph provide further details for those interested in the inner working of the system. It provides a hook to many system parameters, e.g. input/output, CPU cycles, and threads' activities.

Timing a database query should be done with care. Often you will notice differences in response time for the same query ran multiple times. The underlying cause can be that the data itself resides on disk (slow) or is already avaiable in the memory caches (fast), a single user runs queries (fast) or has to compete with other users (slow), including competing with other processes on your box fighting over cpu, memory, and IO resources. As a precaution you might want to flush the system caches. The Windows tool flushes the cache. You'll need to press the "Flush Cache WS" and "Flush All Standby" buttons. On Linux you have to create a little job that consumes all memory.

For more general information on running experiments and measuring time, see our performance tutorial.

Query History

Query History mk Sat, 03/09/2013 - 17:36

The SQL implementation comes with a simple query profiler to detect expensive queries. It is centered around two predefined internal tables that store the definitions of all executed SQL queries and their execution time.

Query logging can be started by calling the procedure querylog_enable(), which saves some major compilation information of a query in the 'querylog_catalog' table:

name type references description
"id" OID    
"owner" STRING   The SQL user who has executed this query.
"defined" TIMESTAMP   Time when the query was started
"query" STRING   The query that has been executed
"pipe" STRING   The optimiser pipe line that has been used
"plan" STRING   Name of its MAL plan
"mal" INTEGER   Size of its MAL plan in the number of statements
"optimize" BIGINT   Time in microseconds for the optimiser pipeline

Query logging can be stoped by calling procedure querylog_disable().

The query performance is stored in the table  'querylog_calls'. The owner of the query definition is also the one who is referenced implicitly by the 'id' of a call event. The key timing attributes are 'run', i.e. the time to execute the query to produce the result set, and 'ship', i.e. the time to render the result set and sent it to the client. All times are in microseconds.

The remaining parameters illustrate resource claims. The 'tuples' attribute denotes the size of the result set in the number of rows. The 'cpu' load is derived from the operating system statistics (Linux only) and is given as a percentage. The same holds for the 'io' waiting time.

name type references description
"id" OID sys.querylog_catalog.id  
"start" TIMESTAMP   time the statement was started
"stop" TIMESTAMP   time the statement was completely finished
"arguments" STRING   actual call structure
"tuples" BIGINT   number of tuples in the result set
"run" BIGINT   time spent (in usec) until the result export
"ship" BIGINT   time spent (in usec) to ship the result set
"cpu" INT   average cpu load percentage during execution
"io" INT   percentage time waiting for IO to finish

The view 'sys.querylog_history' includes some useful information from both tables:

create view sys.querylog_history as
    select qd.*, ql."start",ql."stop", ql.arguments, ql.tuples, ql.run, ql.ship, ql.cpu, ql.io
    from sys.querylog_catalog() qd, sys.querylog_calls() ql
    where qd.id = ql.id and qd.owner = user;

The following code snippet illustrates its use:

sql>call sys.querylog_enable();
sql>select 1;
sql>select 1;
sql>select * from sys.querylog_catalog;
sql>select * from sys.querylog_calls;
sql>select * from sys.querylog_history;
sql>select id, query, avg(run) from sys.querylog_history group by id,query;
sql>call sys.querylog_disable();
sql>call sys.querylog_empty();

The 'sys.querylog_enable()' function also take a parameter, 'threshold', which is an integer in millisecond. When the query log is enabled with this parameter, it will only log those queries whose execution times are longer than the threshold.  This feature can be handy to prevent the database from being swarmed by too many short running queries, hence reduce the overhead incurred by the query log (see below), while helping the DBA detecting expensive queries.

Disabling the query log will not remove existing query logs; it only prevents subsequent queries to be logged. Once the query log is re-enabled, information of subsequently executed queries will be appended to the existing query logs.

Query logs are stored in persistent tables, ie they will survice a MonetDB server restart. They can only be removed 'sys.querylog_empty()'. A downside of this implementation is its relative high overhead because every read query will trigger a write transaction.