This is not performance troubleshooting post but in some sense related to that. I’ve been working on a paper where I suggest to apply reinforcement learning method of machine learning to design of spin-then-block strategy. The default behavior of latches and mutexes in recent versions of Oracle is, in fact, one implementation of the spin-then-block method. The primary idea of the paper is to have threads learn the cases where it should prefer either sleeping or spinning. The objective is to minimize cost associated with lock acquisition. The paper is out now. Unfortunately, I couldn’t make time to set up experiments to test this. However, the paper develops a self-tuning machine learning based framework that can wrap spin-then-block method. Obviously, some challenges are still there. The journal is indexed in such scientific databases as Scopus, Web of Science (ESCI) and etc. The paper is available here. Hope you will find it interesting.
Posted in General.