70
edits
Changes
→Conclusion
For more information on System Overview click [https://www.intel.com/content/www/us/en/develop/documentation/vtune-help/top/analyze-performance/platform-analysis-group/system-overview-analysis.html here]
==='''Versions of the software:'''===
*Standalone VTune Profiler Graphical Interface
*Web Server Interface
====Performance====
As can be seen from the screenshot below, in the OpenMP solution, the work is spread unevenly between 8 threads. It can be described by the fact that the first node is responsible for initializing the arrays and single construct. Also, there is a lot of idle time due to the barrier construct. But the Prefix can itself seems to be spread almost evenlyfall into average optimal CPU utilization.
[[File:OMP_Scan.png]]
===TBB Prefix Scan===
Finally, the TBB solution that uses tbb:parallel scan and Body functor, as well as auto partitioning
====Code====
<source>
====Performance====
As can be seen from the screenshot below, there is a lot of overhead due to tbb::parallel_scan scheduling. Additionally, it seems that most work is done by thread 1, which can be explained by the fact that the array is still initialized serially. The solution can be optimized by choosing the proper grain size which is the first suggestion Vtune gave.
[[File:TBB_Scan.png]]