Tracing and Visualization of GPU Offloading Using LTTng and Eclipse Trace Compass

Hrutvi Barad

Dr. Yonghong Yan

​

Motivations

• Examining the complexities and difficulties of GPU offloading, a difficult task.
• Utilizing GPU Computing: Making effective use of GPU processing capabilities.
• Increasing the performance of GPU applications by developing a thorough grasp of their behavior and analysis.
• Navigating Complexity: Understanding that GPU offloading is a difficult process and dealing with it.
• Insightful Tracing and Visualization: For intuitive and insightful analysis, use Trace Compass with LLTing.

​

Backgrounds

GPU (Graphics Processing Unit) Offloading:

• What is GPU Offloading
• The process of moving computational tasks from the CPU to the GPU for processing is referred to as GPU offloading.
• Why should we use it? How is it helpful?
• It accelerates the computation process, improves the performance and it is energy efficient.
• It is helpful for researchers to perform complex computations in less time and with greater precision.

​

Backgrounds

LTTng (Linux Trace Toolkit):

• What is LTTng?
• Open source software program that can be used for tracking and analyzing the activity of software systems.
• How is it helpful? Why should we use it?
• It is helpful to pinpoint kernel and application performance issues.
• It comprises kernel modules and analysis tools for data collection.

​

​

​

​

Backgrounds

Eclipse Trace Compass:

• What is Trace Compass?
• Eclipse trace compass is used to analyze and view traces and logs.
• Why should we used it ? How is it helpful?
• It makes the information user friendly by interpreting it by the visual representation.
• It also helps in accelerating debugging by providing detailed trace data for developer that helps developers identify and understand the issue.

Current Progress

• Installation and integration of the previous student's work into our research environment.
• Active learning and research in the field of GPU (Graphics Processing Unit) technology, which is essential for our project.
• Simultaneous work on the development of a research paper alongside our practical research activities.
• Additionally, we are actively involved in the creation of new plugins within our research environment to extend and enhance its capabilities.

Future Plans

• Addition of more features for visualization
• Showing GPU data movement and kernel computation
• Using XML-based data-drive analysis of Trace Compass
• https://archive.eclipse.org/tracecompass/doc/stable/org.eclipse.tracecompass.doc.user/Data-driven-analysis.html
• 3D-visualization
• Evaluation using benchmarks and applications
• From CUDA samples https://github.com/NVIDIA/cuda-samples/tree/master/Samples
• More applications such as GROMACS (https://developer.nvidia.com/blog/a-guide-to-cuda-graphs-in-gromacs-2023/)

References

• https://saturncloud.io/blog/why-offloading-to-a-gpu-is-a-good-idea/#:~:text=Offloading%20to%20a%20GPU%20means,Processing%20Unit)%20to%20a%20GPU.
• https://archive.fosdem.org/2018/schedule/event/the_lttng_approaches_to_solving_complex_problems/attachments/slides/2315/export/events/attachments/the_lttng_approaches_to_solving_complex_problems/slides/2315/lttng_fosdem_2018.pdf
• https://lttng.org/docs/v2.12/#:~:text=Welcome!,to%20trace%20the%20Linux%20kernel.
• https://archive.eclipse.org/tracecompass/doc/stable/org.eclipse.tracecompass.doc.user/Trace-Compass-Main-Features.html#Trace_Compass_Main_Features
• https://projects.eclipse.org/projects/tools.tracecompass/governance#:~:text=Eclipse%20Trace%20Compass%E2%84%A2%20is,is%20user%2Dfriendly%20and%20informative.

​

Thank You

Q & A