Open Source Summit + Embedded Linux Conference Europe 2020

Nowadays AI, machine learning are dominating the computing world, companies in this industry are willing to deploy their infrastructure in the machines with massive amount of memory and processors to train their algorithms. However, currently the cloud providers map the resources to virtual machines on demand by using the hardware provided mechanisms (such as page fault, etc.), this way is flexible indeed, unfortunately it will introduce tremendous performance drop for the guest with massive memory during migration or other scenarios. In this presentation, we propose our solution to this challenge, we developed a lockless mechanism to boost the concurrency while mapping guest memory. Relay on this solution, we could provide a faster and consistent environment to perform DL model training, inference and other workload which request substantial resources.