Scheduling for MIG-Capable GPUs: Accelerator-Aware Operating System Scheduling

The development of heterogeneous CPU-GPU systems for modern data centers in recent years increased the computational power massively. Multi-Instance GPUs (MIGs) allow to split a single physical GPU into multiple independent instances. Thus, more than one user can get exclusive acces to the same physical GPU. However, there are many different configurations for a single MIG-capable GPU that need to be controlled by a central instance. Today, only the system administrator can create and destroy MIG instances and assign them to the users. The goal of the thesis is to develop a scheduler that automates the creation and destruction of Multi-Instance GPU (MIG) instances. The thesis evaluate if such a scheduler can boost the performance and efficiency of the CPU-GPU system. The developed scheduler is based on an RSVT scheduler that follows an allocation-consumption scheme to decide which tasks to run next. The lack of a preemption mechanism for GPUs reduces the performance and usability of the scheduler