Performance measurements: operf, gprof, gcov and valgrind

1. You need to do this lab on a computer with operf enabled, which means either power.cs.lth.se or your own Linux machine. See page 221 for how to enable operf on Linux. See Appendix A for how to work remotely with a Unix machine (Linux or macOS) and login to power.cs.lth.se.
   
   
2. You will use but don't have to implement isfinite from the pseudo code since it is part of the Standard C library.
   
   
3. When you have implemented these functions your integer program solver can find the optimal solution for our example input --- since all variables happen to be integers.
   
   
4. (this is in progress) Now run with larger input so that your program takes for example 10 seconds to complete. (Jonas will upload more input cases)
   
   You will next use four different program for performance measurements.
   
   
5. First type

   operf -e CYCLES:100000:0:0:1 ./a.out < input
   

   See page 220. Your program is run with hardware counters enabled which count the number of clock cycles.
   
   Use opreport and opannotate to figure out which functions take most of the time in your program.
   
   
6. Next compile with

   gcc -pg -g intopt.c
   

   and run your program. Then give the command

   gprof -T a.out
   

   What does the output mean and how can you use it to better understand your program?
   
   
7. Next compile with

   gcc -fprofile-arcs -ftest-coverage -g intopt.c
   

   and run your program again. Then give the command

   gcov intopt.c
   

   This will create a file intopt.c.gcov. What does it contain how and can you use this information?
   
   To get even more detailed information about which cases in your program are most common, you can use

   gcov -b intopt.c
   

   What does it say?
   
   
8. In addition to finding memory errors, Valgrind can also be useful when doing performance measurements. Valgrind is slow than the other tools due to it simulates the computer so use the small input, at least initially. Recompile with -O3 and give the command

   valgrind --tool=cachegrind --I1=65536,1,128 --D1=32768,2,128 \
   > --LL=1048576,8,128 ./a.out < i
   

   These options specify the cache parameters to match our POWER8 machine.
   
   How can you see how many instructions in total, load instructions, and store instructions are executed?
   
   What are the cache miss rates and are the cache likely to be a performance problem for this input?