Ruprecht-Karls-Universität Heidelberg

Multi-GPU Support for the Network-Attached Accelerators Approach

Master Thesis by Thomas Ott


Supercomputers can help to solve some of humankind’s biggest challenges. The development of new cancer drugs or the simulation of the future climate are two examples of high-performance computing. Solving problems faster, with more data and better energy-efficiency are key drivers behind the development of better supercomputers. Heterogeneous supercomputers play a key role in achieving these goals. They make use of accelerators like GPUs or FPGAs to support the CPUs. An innovative way of connecting accelerators with CPUs is provided by the “Network-Attached Accelerator”-concept.

A generic realization of this novel approach is provided by this thesis. This work is based on a previous proof-of-concept implementation and focuses on the support of multiple GPUs. A thorough conceptual analysis was conducted, and a Linux Kernel module was developed. It allows to attach PCI Express accelerator devices to a Linux host via the EXTOLL high-performance interconnection network. With some limitations, the functionality of the Kernel module could be demonstrated on a x86-based test system. CUDA programs using two EXTOLL-attached NVIDIA GPUs could be executed.


« back

back to top