1. Intel® performance analyze tools Nikita Panov (nikita.v.panov@intel.com)nikita.v.panov@intel.com Idrisov Renat

2. Intel VTune™ Amplifier XE Performance Profiler provides information on code performance for users developing serial and multithreaded applications on Windows* and Linux* operating systems on Windows systems, the VTune Amplifier XE integrates into Microsoft Visual Studio* software and is also available as a standalone GUI client on Linux systems, VTune Amplifier XE works only as a standalone GUI client on both Windows and Linux systems, you can benefit from using the command-line interface for collecting data remotely or for performing regression testing 2

3. Intel VTune™ Amplifier XE Performance Profiler Use the VTune Amplifier XE to locate or determine the following: The most time-consuming (hot) functions in your application and/or on the whole system Sections of code that do not effectively utilize available processor time The best sections of code to optimize for sequential performance and for threaded performance Synchronization objects that affect the application performance Whether, where, and why your application spends time on input/output operations The performance impact of different synchronization methods, different numbers of threads, or different algorithms Thread activity and transitions Hardware-related bottlenecks in your code 3

4. Hotspot analysis Choose an analysis target. Choose the Hotspots analysis type. Run the Hotspots analysis to locate most time- consuming functions in an application. Analyze the function call flow and threads. Analyze the source code to locate the most time- critical code lines. Compare results before and after optimization. 4

5. Creating project If symbolic debug information is compiled into the executable it will help to find right lines of the code. But to analyze real application workflow it is recommended to compile with normal options 5

6. Choose the hotspots analysis type 6 On the left pane of the Analysis Type window, locate the analysis tree and select Algorithm Analysis > Hotspots.

7. Analysis results Note that CPU Time for the sample application is equal to 64.907 seconds. It is the sum of CPU time for all application threads. Total Thread Count is 3, so the sample application is multi- threaded. The Top Hotspots section provides data on the most time- consuming functions (hotspot functions) sorted by CPU time spent on their execution. 7

8. Call stack Select the initialize_2D_buffer function in the grid and explore the data provided in the Call Stack pane on the right. 8

9. Analyzing the results 9

10. Analyzizng the results 1. Timeline area. When you hover over the graph element, the timeline tooltip displays the time passed since the application has been launched. 2. Threads area that shows the distribution of CPU time utilization per thread. Hover over a bar to see the CPU time utilization in percent for this thread at each moment of time. Green zones show the time threads are active. 3. CPU Usage area that shows the distribution of CPU time utilization for the whole application. Hover over a bar to see the application-level CPU time utilization in percent at each moment of time. 10

11. Analyzing the code 1 – source code, 2 – assembler, 3 – processor time, 4 и 5 – useful markers and scroll controls to identify problem code 11

12. Comparing the results Specify the Hotspots analysis results you want to compare and click the Compare Results button 12

13. Comparing the results 1 – time difference 2 – before the optimization (first version) 3 – after the optimization 13

14. Locks and waits analysis Other kind of analysis are provided in a similar way 14

15. Performing the analysis After the analysis you will be given an information according to the analysis type choosen 15

16. Analyzing the results Results are also could be viewed with the program call stack 1 – corresponding object, 2 – processor usage, 3 – wait cycles count 16

17. Analyze the code 1 – source code, 2 – processor usage, 3 – wait loop count, 4 - navigation 17

18. Comparing the results 1 – wit loop difference, 2 – wait loop count before, 3 – wait loop count after the optimizations, 4 – loop count difference, 5 и 6 – loop count 18

19. Useful events CPU_CLK_UNHALTED.CORE – processor clock ticks INST_RETIRED.ANY – number of instructions been executed BUS_TRANS_ANY.ALL_AGENTS – bus transaction count L2_LINES_IN.SELF.DEMAND –L2 cache misses. BR_INST_RETIRED.MISPRED – mispredicted branch count 19

20. Memory access time 20

21. Application performance analysis 21

22. Application performance analysis 22

23. Useful compiler options /Od (-O0 for Linux) – no optimizations (used for debug). /O2 (-O2 for linux) – only default optimizations. /O3 (-O3 for linux) – some additional optimization set. /xO (-xO for Linux) – non-intel arhitecture optimizations. /Qipo (-ipo)- interprocedural optimizations. /Qparallel (-parallel) –autoparallelization. /Qopt-report (-opt-report) /Qopt-report-file /Qopt-report-phase /Qopt-report-help /Qopt-report-routine /Qvec-report [1/2/3] 23

24. Cluster tools Intel Trace Collector & Analyzer Cluster tools for enterprise applications with MPI. 24

25. ITAC 25 Similar analysis, but different application and different proposes

26. 26

27. 27

28. 28

29. Thank you