<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN">
<html><body>
<p>Hello,<br /><br />I would like to submit two papers that use LLVM to the Related Publications section.<br /><br />Both papers focus on code isolation applied to perform piecewise compiler optimizations.<br />The code isolation process is performed by CERE, an open source tool based on LLVM.<br /><br />The second paper is an extended version of the first one.<br /><br />1) Piecewise Holistic Autotuning of Compiler and Runtime Parameters</p>
<p><br />@inproceedings{popov2016piecewise,<br /> title={Piecewise Holistic Autotuning of Compiler and Runtime Parameters},<br /> author={Popov, Mihail and Akel, Chadi and Jalby, William and de Oliveira Castro, Pablo},<br /> booktitle={European Conference on Parallel Processing},<br /> pages={238--250},<br /> year={2016},<br /> organization={Springer}<br />}<br /><br />2) Piecewise holistic autotuning of parallel programs with CERE</p>
<p><br />@article{popov2017piecewise,<br /> title={Piecewise holistic autotuning of parallel programs with CERE},<br /> author={Popov, Mihail and Akel, Chadi and Chatelain, Yohan and Jalby, William and de Oliveira Castro, Pablo},<br /> journal={Concurrency and Computation: Practice and Experience},<br /> volume={29},<br /> number={15},<br /> year={2017},<br /> publisher={Wiley Online Library}<br />}<br /><br />Do not hesitate if you have any questions or if you need any additional documents.<br /><br />Thank you,<br />Mihail Popov<br /><br /><br />-----------------------------------------------------------------------------------<br /><br />PAPERS SUMMARY:<br /><br />Piecewise Holistic Autotuning of Compiler and Runtime Parameters<br /><br />Abstract. Current architecture complexity requires fine tuning of compiler <br />and runtime parameters to achieve full potential performance. Autotuning <br />substantially improves default parameters in many scenarios<br />but it is a costly process requiring a long iterative evaluation.<br />We propose an automatic piecewise autotuner based on CERE (Codelet<br />Extractor and REplayer). CERE decomposes applications into small<br />pieces called codelets: each codelet maps to a loop or to an OpenMP<br />parallel region and can be replayed as a standalone program.<br />Codelet autotuning achieves better speedups at a lower tuning cost. By<br />grouping codelet invocations with the same performance behavior, CERE<br />reduces the number of loops or OpenMP regions to be evaluated. Moreover <br />unlike whole-program tuning, CERE customizes the set of best <br />parameters for each specific OpenMP region or loop.<br />We demonstrate CERE tuning of compiler optimizations, number of<br />threads and thread affinity on a NUMA architecture. On average over the<br />NAS 3.0 benchmarks, we achieve a speedup of 1.08× after tuning. Tuning <br />a single codelet is 13× cheaper than whole-program evaluation and<br />estimates the tuning impact on the original region with a 94.7% accuracy. <br />On a Reverse Time Migration (RTM) proto-application we achieve<br />a 1.11× speedup with a 200× cheaper exploration.<br /><br /><br />Piecewise Holistic Autotuning of Parallel Programs with CERE<br /><br />Current architecture complexity requires fine tuning of compiler<br /> and runtime parameters to achieve best performance. Autotuning<br />substantially improves default parameters in many scenarios but it is a<br />costly process requiring long iterative evaluations.<br />We propose an automatic piecewise autotuner based on CERE (Codelet<br />Extractor and REplayer). CERE decomposes applications into small<br />pieces called codelets: each codelet maps to a loop or to an OpenMP<br />parallel region and can be replayed as a standalone program.<br />Codelet autotuning achieves better speedups at a lower tuning cost. By<br />grouping codelet invocations with the same performance behavior, CERE<br />reduces the number of loops or OpenMP regions to be evaluated. Moreover <br />unlike whole-program tuning, CERE customizes the set of best parameters<br /> for each specific OpenMP region or loop.<br />We demonstrate the CERE tuning of compiler optimizations, number<br />of threads, thread affinity, and scheduling policy on both NUMA and<br />heterogeneous architectures. Over the NAS benchmarks, we achieve an<br />average speedup of 1.08× after tuning. Tuning a codelet is 13× cheaper<br />than whole-program evaluation and predicts the tuning impact with a<br />94.7% accuracy. Similarly, exploring thread configurations and scheduling<br /> policies for a Black-Scholes solver on an heterogeneous big.LITTLE<br />architecture is over 40× faster using CERE.<br /><br /></p>
<p> </p>
<div> </div>
</body></html>