Architecture Assessed Coursework 2
-
Create database:
cd db/ sqlite3 expirements.db < build-tables.sql
(NOTE: database filename is hardcoded in .py files)
- Create experiment:
./create-experiment.py C78DAB description about this experiment does
This will also print the experiment ID to output
- Add results for a run:
./add-data.py 44 file-containing-stdout file-containing-stderr
Note 44 here is the experiment id returned from create
Unfortunately lab machines only have CUDA compute capability = 1.1 so many of the advanced profiler features are not possible. Furthermore lab machines only have Nvidia's command line profiler installed and so we must make use of this.
In order to start profiling code
source profile.sh- this will export the relevant shell variables- Modify
config.txtto contain any counters etc, you wish. Beware only a few counters should be added For a full list of counters see https://svn.ece.lsu.edu/svn/gp/cuda/stream/.cuda-profile-config - Run code normally
- Find results in
profile.dat - Parse with
python python/profile_parser.py profile.datto see useful results. - Parse with
python python/profile_parser.py --latex profile.datfor tabulated results. - To turn profiling off
source profile_off.sh