ETD

• cost model
• gpgpu computing
• GPU
• performance model
• scheduler
• task
• virtual execution environment

In this thesis we developed a single task scheduler in a CPU-GPU heterogeneous environment.
We formulated a GPGPU performance model recognizing a ground model common to any GPGPU platform that must be refined to consider specific platforms. We proposed a model refinement for the Nvidia CUDA platform.
Moreover, we formulated a CPU performance model for the Common Language Infrastructure virtual execution environment.
Finally, we developed Libra.Net, a particular implementation of the scheduler for the Microsoft Common Language Runtime and evaluated its efficiency.