The CalQL query language

The Caliper Query Language (CalQL) is used to filter, aggregate, and create reports from Caliper data in Caliper’s data processing tools (cali-query) and report services. Its syntax is very similar to SQL.

For example, filtering and aggregating a trace into a function time profile for an application’s main loop and printing results in a table, sorted by time, is accomplished with the following statement:

SELECT *, count(), sum(time.duration)
WHERE loop=mainloop
GROUP BY function
FORMAT table
ORDER BY time.duration

Quick reference

This table contains a quick reference of all CalQL statements:

LET <list>                   # Compute additional attributes in input records
  <a>=ratio(<b>,<c>,<S>)     # computes a = <b>/<c>*S
  <a>=scale(<b>,<S>)         # computes <a> = <b>*S
  <a>=truncate(<b>,<S>)      # computes <a> = <b> - mod(<b>, S)
  <a>=first(<a0>,<a1>, ...)  # <a> is the first of <a0>, <a1>, ... found in the input record
  <a>=sum(<a0>,<a1>,...)     # computes sum of <a0>, <a1>, ... in the input record
  ... IF <condition>         # apply only if input record meets condition

SELECT <list>                # Select attributes and define aggregations (i.e., select columns)
  *                          # select all attributes
  path                       # select the nested region hierarchy
  <attribute>                # select <attribute>
  count()                    # number of input records in grouping
  sum(<attribute>)           # compute sum of <attribute> for grouping
  min(<attribute>)           # compute min of <attribute> for grouping
  max(<attribute>)           # compute max of <attribute> for grouping
  avg(<attribute>)           # compute average of <attribute> for grouping
  percent_total(<attribute>) # compute percent of total sum for <attribute> in grouping
  ratio(<a>,<b>,<S>)         # compute sum(<a>)/sum(<b>)*S for grouping
  inclusive_sum(<a>)         # computes inclusive sum of <a> in a nested hierarchy
  scale(<a>,<S>)             # computes sum of <a> multiplied by scaling factor S
  inclusive_scale(<a>,<S>)   # computes inclusive scaled sum of <a> in a nested hierarchy
  inclusive_percent_total(<a>) # computes inclusive percent-of-total of <a> in a nested hierarchy
  any(<attribute>)           # pick one value for <attribute> out of all records in grouping
  scale_count(<S>)           # count number of input records for grouping and multiplied by <S>
  inclusive_ratio(<a>,<b>,<S>) # computes inclusive_sum(<a>)/inclusive_sum(<b>)*<S>
  inclusive_min(<a>)         # compute inclusive min of <a>
  inclusive_max(<a>)         # compute inclusive max of <a>
  ... AS <name>              # use <name> as column header in tree or table formatter
  ... UNIT <unit>            # use <unit> as unit name

GROUP BY <list>              # Define aggregation grouping (what to aggregate over, e.g. "function,mpi.rank")
  <attribute>                # include <attribute> in grouping
  path                       # include the nested region hierarchy in grouping

WHERE <list>                 # define filter (i.e., select records/rows)
  <attribute>                # records where <attribute> exists
  <attribute> = <value>      # records where <attribute> = <value>
  <attribute> < <value>      # records where <attribute> > <value>
  <attribute> > <value>      # records where <attribute> < <value>
  NOT <clause>               # negate the filter clause

FORMAT <formatter>           # Define output format
  cali                       # .cali format
  expand                     # “<attribute1>=<value1>,<attibute2>=<value2>,...”
  json                       # write json records { “attribute1”: “value1”, “attribute2”: “value2” }
    (pretty)                 #   ... in a more human-readable format
    (object)                 #   ... in a json object that includes attribute info and globals
  json-split                 # json format w/ separate node hierarchy for hatchet library
  table                      # human-readable text table
  tree                       # human-readable text tree output

ORDER BY <list>              # Sort records in output (table and tree formatters)
  <attribute>                # order by <attribute>
  ... ASC                    # sort in ascending order
  ... DESC                   # sort in descending order

LET

The LET statement applies computation operators on input record entries and adds the results as a new entries in the input records. The new entries can then be used in subsequent SELECT, GROUP BY, or FORMAT statements. For example, we can use the scale() operator to scale a value before subsequent aggregations:

LET
  sec=scale(time.duration,1e-6)
SELECT
  path,sum(sec)

We can use the truncate() operator on an iteration counter to aggregate blocks of 10 iterations in a time-series profile:

LET
  block=truncate(iteration#mainloop,10)
SELECT
  block,sum(time.duration)
GROUP BY
  block

The first() operator returns the first attribute out of a list of attribute names found in an input record. It can also be used to rename attributes:

LET
  time=first(time.duration,sum#time.duration)
SELECT
  sum(time) AS Time
GROUP BY
  path

LET terms have the general form

a = f(…) [IF <condition>]

where f is one of the operators and a is the name of the result attribute. The result is added to the input records before the record is processed further. Result entries are only added to a record if all required input operands are present.

With the optional IF condition, the operation is only applied for input records that meet a condition. One can use this to compute values for a specific subset of records. The condition clauses use the same syntax as WHERE filter clauses. The example below defines a “work” attribute with the time in records that contain “omp.work” regions, and then uses that to compute efficiency from the total and “work” time:

LET

work=first(time.duration) IF omp.work

SELECT

sum(time.duration) AS Total, sum(work) AS Work, ratio(work,time.duration) AS Efficiency

GROUP BY

path

SELECT

The SELECT statement selects which attributes in snapshot records should be printed, and defines aggregations. SELECT * selects all attributes. SELECT op(arg) enables aggregation operation op with argument arg. Generally, attributes will be printed in the order given in the SELECT statement.

An example to print all attributes and enable visit count aggregation:

SELECT *, count()

Aggregation operations create a new output attribute. The name is typically derived from the input attribute(s). For example, the result of sum(attr) is stored in sum#attr. All selection attributes and aggregation arguments must come from the input data; recursive aggregations (e.g., min(count())) within a query are not supported.

The AS keyword assigns a user-defined name to a selected attribute or aggregation result. Example:

SELECT sum(time.duration) AS "Time (usec)" FORMAT table

Here, the table formatter uses “Time (usec)” instead of “sum#time.duration” as column name for the sum(time.duration) column. Only some formatters (table, tree, json, and json-split) support AS.

Inclusive aggregation operations (inclusive_sum, inclusive_scale, and inclusive_percent_total) compute inclusive values (value for a tree node plus all of its children) for datasets with hierarchical regions. This applies to the hierarchy defined by attributes with the CALI_ATTR_NESTED property, including the default function, annotation, and loop attributes from Caliper’s high-level annotation macros.

A more complex example:

SELECT
  *,
  scale(time.duration,1e-6) AS Time,
  inclusive_percent_total(time.duration) AS "Time %"
GROUP BY
  path
FORMAT
  tree

The computes the (exclusive) sum of time.duration divided by 100000 and the inclusive percent-of-total for time.duration. Example output:

Path      Time  Time %
main         5     100
  foo       35      90
    bar     10      20

WHERE

The WHERE statement can be used to filter the records to aggregate/print. The statement takes a comma-separated list of clauses. Records that don’t match all of the clauses are filtered out. Filters can only be defined on input attributes, i.e. it is not possible to filter on aggregation results.

Currently, there are clauses to test for existance of an attribute label in a record, and to filter for specific attribute values. Clauses can be negated with NOT. Example:

WHERE loop,function=foo

Select records that contain loop attributes where function equals foo. Note: for nested attributes, the match may occur on any level. The above example matches the following:

{ loop=mainloop, iteration=5, function=main/foo }      Matches
{ loop=mainloop, iteration=5, function=main/foo/bar }  Matches ('function=foo' will be matched on any nesting level)
{ function=main/foo }                                  No match: 'loop' attribute missing

The NOT keyword negates clauses:

WHERE NOT iteration#mainloop=4

matches every record where iteration#mainloop is not 4 (including records without iteration#mainloop attributes.

GROUP BY

The GROUP BY statement defines the aggregation key for aggregation operations. The aggregation key defines for which attributes separate (aggregate) records will be kept. That is, the aggregator will generate an aggregate record for each unique combination of key values found in the input records. The values of the aggregation attributes in the input records will be accumulated and appended to the aggregate record.

Consider the following table of input records:

function loop     iteration
         mainloop
foo      mainloop         0
bar      mainloop         0
foo      mainloop         1
bar      mainloop         1
foo      mainloop         2
bar      mainloop         2

With this input, the following GROUP BY statement will create a function profile:

SELECT *, count() GROUP BY function

function count
foo          3
bar          3

FORMAT

The FORMAT statement selects the output format option. Caliper can produce machine-readable (e.g., json or Caliper’s own csv-style) or human-readable output (text tables or a tree representation).

See Output formatting for a list of available formatters.

ORDER BY

Sort output records by the given sort criteria. The statement defines a list of attributes to sort output records by. Records can be sorted ascending (using the ASC keyword) or descending (using DESC). Sorting is available in the “tree” and “table” formatters.

The following example prints a iteration/function profile ordered by time and iteration number. Note that one must use the original attribute name and not an alias assigned with AS:

SELECT
  *,
  sum(time.inclusive.duration) AS Time
FORMAT
  table
ORDER BY
  sum#time.inclusive.duration DESC,
  iteration#mainloop

function loop     iteration#mainloop     Time
main                                   100000
main     mainloop                       80000
main/foo mainloop                  0     2500
main     mainloop                  0     1500
main/foo mainloop                  1     3500
main     mainloop                  1     2000
main     mainloop                  2     1000
main/foo mainloop                  2      600
...