Tuning Performance Bottlenecks

Procedure

1. Modify the source file, as shown in Figure 1. After modification, rename the file to cpu_branch_prediction_after.cpp and upload it to the /home/demo directory.
   The modified source code reorders data to reduce the CPU branch misprediction rate.

   Figure 1 Modified source file
   
2. Compile the source file.

   g++ -o /home/demo/cpu_branch_prediction_after /home/demo/cpu_branch_prediction_after.cpp

3. Switch to the installation directory of the Kunpeng Performance Boundary Analyzer. Replace xxx in the command with the actual version.

   cd /home/ksys-x.x.x-Linux-aarch64

4. Collect the application performance data after optimization.

   ./ksys collect /home/demo/cpu_branch_prediction_after

   Figure 2 Microarchitecture statistics
   

   According to the microarchitecture statistics, the proportion of Branch Mispredicts(%) under Bad Speculation(%) decreases from 54% to less than 1%, resulting in improved application performance.

5. Switch to the demo directory and check the runtime of the application after the optimization.
   1. Switch to the demo directory.

      cd /home/demo

   2. Check the application runtime after the optimization.

      time /home/demo/cpu_branch_prediction_after

      After the command is executed, it is found that the application runtime decreases from 61 seconds to 21 seconds. As a result of the optimization, the application computation performance improves.

      Figure 3 Runtime