Ayudante: A Deep Reinforcement Learning Approach to Assist Persistent Memory...
summary
A deep RL model translates C/C++/Java volatile code into PM-aware code by inserting PMDK annotations, then a three-level code refining pipeline runs gcc, PM checkers and PMTest to report bugs. Evaluation on Intel Optane DC PM shows generated microbenchmarks and a key-value store pass PMDK checks, with a 78.7% checker pass rate on diverse repos and an average of 13 lines of code change needed; performance matches PMDK examples despite a 30% higher LLC store count on one structure.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:42Outline
- 5:28Deep RL-based Code Generator - the Model
- 7:09Deep RL-based Code Generator - Search Strategy
- 8:20Deep RL-based Code Generator - Transfer Learning
- 9:16An Example of RL-based Code Generation
- 9:54Ayudante Framework
- 11:06A refining suggestion report example
- 11:26Implementation
- 12:16Evaluation
- 14:28Conclusion
related talks
22mMachine Learning Inference in DPDK
10mChip floorplanning with deep reinforcement learning
37mIs the New 2016, but Make It Privacy: On Federated Memory, Contextual Privacy, and...
20mAccelerating Model Loading in LLM Inference by Programmable Page Cache
54mProgramming Language Design and Implementation for the Machine Learning Era: A Personal(…)
14mAIE-ML dev PMD for VRAN and UPF Datapath Acceleration
10mvLLM & Deepspeed Updates
13mScheduled Partial-Credit RL for Reliable Code Generation with Small Language Models (WIP)