|
|
Performance Analysis of Programs Based on PIN Framework
Močáry, Peter ; Fiedor, Jan (oponent) ; Pavela, Jiří (vedoucí práce)
The goal of this thesis is to extend the Performance Version System - Perun by implementing a new Tracer engine leveraging PIN instrumentation framework. This extension implements basic Tracer functionality and, in addition to that, a recording of function arguments' values as well as basic block run-times. The additional data, along with the visualizations introduced in this thesis, provide the necessary context that simplifies the detection of performance degradation. Besides the PIN framework, the new Tracer engine implements an analysis of debug information in DWARF format (using the python pyelftools library) to gather details about function arguments before the data collection process. The resulting engine was tested on multiple implementations of sorting algorithms and successfully detected the most time consuming functions along with the information about the effect of its parameter value on the functions complexity. Testing the PIN engine on a larger-scale project revealed that, in comparison to other Tracer engine implementations, the engine performs better or comparably, and produces the correct output.
|
|
|
Efficient Techniques for Program Performance Analysis
Pavela, Jiří ; Fiedor, Jan (oponent) ; Rogalewicz, Adam (vedoucí práce)
In this work, we propose optimization techniques focused on the data collection process of program performance analysis and profiling within the Perun framework. We enhance Perun (and especially its Tracer module) by extending their architecture and implementing novel optimization techniques that allow Perun to scale well even for large projects and test scenarios. In particular, we focus on improving the data collection precision, scaling down the amount of injected instrumentation, limiting the time overhead of the collection and profiling processes, reducing the volume of raw performance data and the size of the resulting profile. To achieve such optimization, we utilized statistical methods, several static and dynamic analysis approaches (as well as their combination) and exploited the advanced features and capabilities of SystemTap and eBPF frameworks. Based on the evaluation performed on two selected projects and numerous experiment cases, we were able to conclude that we successfully achieved significant levels of optimization for nearly all of the identified metrics and criteria.
|
|
|
Performance Analysis of Programs Based on PIN Framework
Močáry, Peter ; Fiedor, Jan (oponent) ; Pavela, Jiří (vedoucí práce)
The goal of this thesis is to extend the Performance Version System - Perun by implementing a new Tracer engine leveraging PIN instrumentation framework. This extension implements basic Tracer functionality and, in addition to that, a recording of function arguments' values as well as basic block run-times. The additional data, along with the visualizations introduced in this thesis, provide the necessary context that simplifies the detection of performance degradation. Besides the PIN framework, the new Tracer engine implements an analysis of debug information in DWARF format (using the python pyelftools library) to gather details about function arguments before the data collection process. The resulting engine was tested on multiple implementations of sorting algorithms and successfully detected the most time consuming functions along with the information about the effect of its parameter value on the functions complexity. Testing the PIN engine on a larger-scale project revealed that, in comparison to other Tracer engine implementations, the engine performs better or comparably, and produces the correct output.
|
|
|
Efficient Techniques for Program Performance Analysis
Pavela, Jiří ; Fiedor, Jan (oponent) ; Rogalewicz, Adam (vedoucí práce)
In this work, we propose optimization techniques focused on the data collection process of program performance analysis and profiling within the Perun framework. We enhance Perun (and especially its Tracer module) by extending their architecture and implementing novel optimization techniques that allow Perun to scale well even for large projects and test scenarios. In particular, we focus on improving the data collection precision, scaling down the amount of injected instrumentation, limiting the time overhead of the collection and profiling processes, reducing the volume of raw performance data and the size of the resulting profile. To achieve such optimization, we utilized statistical methods, several static and dynamic analysis approaches (as well as their combination) and exploited the advanced features and capabilities of SystemTap and eBPF frameworks. Based on the evaluation performed on two selected projects and numerous experiment cases, we were able to conclude that we successfully achieved significant levels of optimization for nearly all of the identified metrics and criteria.
|
host ::
RSS.