Combine Vortex’s IO redistribution with intra-GPU slicing for higher system efficiency

Develop and evaluate techniques that integrate Vortex’s IO redistribution across GPUs with intra-GPU slicing (e.g., partitioning a single GPU into slices) to collocate complementary workloads and further improve overall system efficiency in multi-tenant environments.

Background

Prior work shows that GPU slicing can improve utilization by collocating workloads that stress different microarchitectural resources on the same GPU. Vortex’s approach redistributes IO across multiple GPUs to accelerate IO-bound analytics on a single target GPU.

The authors suggest combining these two ideas—cross-GPU IO pooling with sub-GPU slicing—to achieve higher system-wide efficiency, but leave the design and evaluation of such an integrated approach as future work.