Generalize Vortex’s IO-redistribution approach to multiple target GPUs

Determine how to extend the Vortex framework’s Exchange-based IO redistribution and scheduling to support data analytics using more than one target GPU, including the design of coordinated IO forwarding, compute orchestration, and resource sharing across multiple target GPUs in multi-GPU systems.

Background

Vortex currently targets a cold-start scenario where a single target GPU performs computation while leveraging the IO bandwidth of all GPUs in a multi-GPU node to transfer data from CPU DRAM. This design decouples IO scheduling from GPU kernel optimization and is demonstrated for one compute GPU with multiple forwarding GPUs.

The authors note that this approach may extend to cases where more than one GPU performs data analytics concurrently, but they have not developed or evaluated the necessary IO coordination and scheduling mechanisms for multi-target configurations.