Analyzing Data with cali-query
March 19, 2025 ยท View on GitHub
When not using Hatchet, you can use Caliper's cali-query tool or Python reader API to convert, analyze, and query Caliper performance data.
In addition to the Hatchet .json format, the hatchet-region-profile config recipe has two other output format options, which you can select with the output.format parameter.
The json output format creates a simple JSON format (different from the JSON format that Hatchet uses) with easy-to-parse dictionary records:
$ basic_example hatchet-region-profile,output.format=json,output=stdout
[
{"time":0.000750},
{"path":"main","time":0.000023},
{"path":"main/setup","time":0.000074},
{"path":"main/main loop","time":0.000024},
{"path":"main/main loop/foo","time":0.000005},
{"path":"main/main loop/bar","time":0.000008}
]
The cali output option creates a .cali file in Caliper's "native" data format. We can read and display this data with cali-query. With the --tree option, cali-query prints a table with hierarchical region. With hatchet-region-profile data, we see times for each MPI rank:
$ mpirun -n 8 lulesh2.0 -i 4 -P hatchet-region-profile,output.format=cali
$ cali-query --tree region_profile.cali
Path mpi.rank time
main
|- 0 0.051542
|- 1 0.032454
|- 2 0.036181
|- 3 0.037724
|- 4 0.033592
|- 5 0.036821
|- 6 0.037938
|- 7 0.056281
lulesh.cycle
|- 0 0.000044
|- 1 0.000052
|- 2 0.000053
|- 3 0.000065
|- 4 0.000055
|- 5 0.000059
|- 6 0.000060
|- 7 0.000065
[...]
Complex queries with CalQL
The event-trace recipe records a complete event trace, i.e. a record for each region begin and end event. With cali-query, we can run custom queries in Caliper's CalQL query language on this data. As an example, here is how we compute and print the time in each loop iteration in the basic_example program:
$ basic_example event-trace,output=trace.cali
$ cali-query -q "select iteration#main\ loop as iteration,sum(time.duration) as time group by iteration#main\ loop format table" trace.cali
iteration time
0.000168
0 0.000012
1 0.000006
2 0.000006
3 0.000006
This example just gives you an idea of what the query system can do. It is most useful when creating custom configuration recipes - however, these are outside of the scope of this tutorial.