HLT Ops 2013-03-22

Transcript

1. Intel Profiler Auditor v2 (not yet released) CPU profiling with Intel® VTune™ Amplifier XE 2013 Alexander Mazurov Ferrara University, CERN
2. 2 Measure frequency and duration of functions calls and/or code instructions. CPU Profiler
3. 3 Profiling Techniques - Hardware counters - Instrumenting the code
4. 4 Instrumenting the code - Statically: * Change code manually / automatically * Compiler assisted (gcc -pg) - Dynamically (at runtime): * Change code in runtime - Valgrind - Google Performance Tools - Intel VTune Amplifier
5. 5 VTune™ Amplifier XE Performance Profiling Tool - x86 (32 and 64-bit) - GUI and CLI
6. 6 VTune™ Features Runtime instrumenting profiler - User-mode sampling - Hardware-based sampling - Concurrency and locks and waits analysis - Threading timeline - Attach to a running process - Source view
7. 7 1) Interupts a process 2) Collect samples of all active instruction addresses 3) Restore a call sequence upon each sample. How user-mode sampling works?
8. 8 User-mode analysis types - Hotspots - Concurrency - Locks and Waits
9. 9 User-mode sampling Hotspots analysis:
10. 10 Group results
11. 11 Call Stack
12. 12 Filter by timeline
13. 13 CPU time by code line Debug mode (-g)
14. 14 User-mode sampling is a statistical method and does not provide a 100% accurate results. Accuracy depends on: - Duration of the collection - Speed of processor - Amount of software activity - Sampling interval * recommended value is 10 ms * profiling is only 5% slower Sampling Accuracy
15. 15 Gaudi Intel Profiling Auditor VTune™ User API + Gaudi Auditors API
16. 16 Algorithms profiling (I) CPU time per sequence branch
17. 17 Algorithms profiling (II)
18. 18 In Intel Profiler Auditor v2 you don't need to modify option files
19. 19 Run: $> intelprofiler2 job.py $> intelprofiler2 --start=1000 --stop=2000 job.py Analyze (GUI): $> amplxe-gui ./profile001hs Analyze (CLI): $> amplxe-cl -reports hotspots -r /collecter/data/r001hs
20. 20 Memory allocation functions operatornew from libstdc++ library: tc_new from tcmalloc library: tc_new uses twice less time then operatornew
21. 21 How to use a dev version (lxplus) $> export CMTCONFIG=$CMTDEB $> SetupProject Moore v14r11 –build-env # you can use any Moore version Moore v14r11 $> git clone git@github.com:mazurov/GaudiProfiling.git Moore v14r11 $> make Moore v14r11 $> SetupProject Moore v14r11 $> source /afs/cern.ch/sw/IntelSoftware/linux/all-setup.sh $> intelprofiler2 /path/to/yourjob $> amplxe-gui ./profile001hs
22. 22 Report example Job file: https://gist.github.com/mazurov/5220677
23. 23 • Don't reinvent the wheel – use the existing tools. • Need your feedback --- the new version will be released soon. Conclusions