PERF PROFILING AND ROOFLINE MODELLING EXERCISE
• In this exercise we’re going to extend our profiling/analysis skills by using:
• perf for basic metric gathering
• Intelvtuneforrooflinemodelling
Copyright By PowCoder代写 加微信 powcoder
• We’re going to use another matrix multiplication kernel code to evaluate performance:
• Downloadmult.tar fromLearn
• Extract with tar xf mult.tar
• Load the compilers module load intel-compilers-19
• Build using make
• Aims are:
• Hardwarecounterprogramanalysis
• Lookatvariationbetweenalgorithmimplementations • Rooflineanalysisofapplications
• Lookatvariationbetweenalgorithmimplementations
• Linux based command line tool that provides access to profiling and performance counter/system information
• Gooddocumentationhere: https://perf.wiki.kernel.org/index.php/Main_Page
• Greatdocumentationhere:http://www.brendangregg.com/perf.html
• perf record ./application
• Profile an application, similar to gprof, prof, etc…
• Reliesonsystem/kernelprivilegestoprovideinformation
• ThesearedisabledonCirrus
• perf stat ./application
• Collectperformancecountersontheapplication
• Again,requiressystem/kernelprivilegeswhicharen’tyetavailableon Cirrus
• However, basic information is available
Trying out perf
• Compile the mult_orig and mult_opt applications using the makefile
• Remember, you will need the intel-compilers-19 module loaded to do this
• Submit the batch script run_perf.slurm to undertake the initial
• Batch script will need altered to use your Cirrus budget
• You can also run on the login node, although results might not be correct. You will need to load the perf module to do this.
• Readtheresultsproduced,lookatmult.htoseethedifferencebetween the two versions of the code
• Does the performance information perf provides make sense?
• Modify the makefile to change the level of compiler optimisation from
–O0 to –O3
• Re-make the applications
• Re-submit the applications to view the new output, what has changed?
Roofline modelling
Intel vtune for roofline modelling
• We will use Intel vtune to automatically collect results and produce a roofline model of the same applications
• Using the following commands we will collect various data sets
advixe-cl -collect survey -project-dir results — ./application advixe-cl -collect tripcounts -flop -project-dir results — ./application advixe-cl -collect roofline -project-dir results — ./application
• There is a batch script to do this automatically for you
• run_roofline.slurm
• Again, you can run this on the login node if you wish, but you will need to load the module in the batch script to get it to work
• Visualise and explore your results using: • advixe-gui
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com