Predicting Dynamic Properties of Heap Allocations using Neural Networks Trained on Static Code

Explore an innovative approach to predicting dynamic properties of heap allocations using neural networks trained on static code in this 18-minute video presentation from ISMM 2023. Delve into the challenges of traditional profile-guided optimization methods and discover how machine learning techniques can potentially overcome these limitations. Examine the trade-off space, promising directions, and experimental data supporting this novel approach. Gain insights into the potential for improving memory allocators and runtime systems without relying on performance profiles. Consider the implications for reducing profiling overheads, simplifying engineering complexity, and capturing a more comprehensive range of workload behaviors. Understand the challenges that future research in this area must address to advance this cutting-edge technique in memory management and profile-guided optimization.

[ISMM'23] Predicting Dynamic Properties of Heap Allocations using Neural Networks Trained on(…)