high performance computing on graphics processing units: hgpu.org

General-purpose GPUs now account for substantial computing power on many platforms, but the management of GPU resources – cycles, memory, bandwidth – is frequently hidden in black-box libraries, drivers, and devices, outside the control of mainstream OS kernels. We believe that this situation is untenable, and that vendors will eventually expose sufficient information about cross-black-box interactions to enable whole-system resource management. In the meantime, we want to enable research into what that management should look like. We systematize, in this paper, a methodology to uncover the interactions within black-box GPU stacks. The product of this methodology is a state machine that captures interactions as transitions among semantically meaningful states. The uncovered semantics can be of significant help in understanding and tuning application performance. More importantly, they allow the OS kernel to intercept-and act upon-the initiation and completion of arbitrary GPU requests, affording it full control over scheduling and other resource management. While insufficiently robust for production use, our tools open whole new fields of exploration to researchers outside the GPU vendor labs.