Section: New Results

Profiling Feedback based Optimizations and Performance Debugging

Participants : Fabrice Rastello, Diogo Sampaio, Fabian Gruber, Christophe Guillon [STMicroelectronics] , Antoine Moynault [STMicroelectronics] , Changwan Hong [OSU, USA] , Aravind Sukumaran-Rajam [OSU, USA] , Jinsung Kim [OSU, USA] , Prashant Singh Rawat [OSU, USA] , Sriram Krishnamoorthy [PNNL, USA] , Louis-Noël Pouchet [CSU, USA] , P. Sadayappan [OSU, USA] .

Profiling feedback is an important technique used by developers for performance debugging, where it is usually used to pinpoint performance bottlenecks and also to find optimization opportunities. Our contributions in this area are twofold: (1) we developed a new technique that combines abstract simulation and sensitive analysis that allows to pinpoint performance bottleneck; (2) we developed a new technique to build a polyhedral representation out of an execution trace that allows to provide feedback on possible missed transformations.

Compiler Optimization for GPUs Using Bottleneck Analysis

Optimizing compilers generally use highly simplified performance models due to the significant challenges in developing accurate analytical performance models for complex computer systems. In this work, we develop an alternate approach to performance modeling using abstract execution of GPU kernel binaries. We use the performance model to predict the bottleneck resource for a given kernel's execution through differential analysis by performing multiple abstract executions with varying machine parameters. The bottleneck analysis is then used to develop an automated search through a configuration space of different grid reshaping, thread/block coarsening, and loop unrolling factors. Experimental results using a number of benchmarks from the Parboil/Rodinia/SHOC suites demonstrate the effectiveness of the approach. The bottleneck analysis is also shown to be useful in assisting high-level domain-specific code generators for GPUs.

This work is the fruit of the collaboration 9.4.1.1 with OSU. It has been presented at the ACM/SIGPLAN conference on Programming Language Design and Implementation, PLDI 2018.

Data-Flow/Dependence Profiling for Structured Transformations

Profiling feedback is an important technique used by developers for performance debugging, where it is usually used to pinpoint performance bottlenecks and also to find optimization opportunities. Assessing the validity and potential benefit of a program transformation requires accurate knowledge of the data flow and data dependencies, which can be uncovered by profiling a particular execution of the program.

In this work we develop Mickey, an end-to-end infrastructure for dynamic binary analysis, which produces feedback about the potential to apply structured transformations to uncover non-trivial parallelism and data locality via complex program rescheduling. Our tool can handle both inter- and intraprocedural aspects of the program in a unified way, thus providing structured interprocedural transformation feedback.

This work is the fruit of the collaboration 9.4.1.1 with CSU and the past collaboration Nano2017 with STMicroelectronics. It has been submitted for presentation at the ACM conference on Principles and Practice of Parallel Programming, PPoPP 2019.