November 15, 2023
Alycia Lisito will be a PhD student in the team, starting from January 1st, 2024. She will work on a flexible and portable implementation of the HPL benchmark, using Chameleon and StarPU, in collaboration with Eviden.
There are currently various versions of the HPL code, such as the open HPL implementations Netlib and rocHPL, or the private implementations from manufacturers AMD (for CPUs), Intel and Nvidia. However, the performance of the reference version is below that of the manufacturer versions, and there are very few optimisation levers for the private versions. The number of computing architectures and associated programming models continues to grow. In particular, when it comes to GPUs, manufacturers such as AMD, Intel and Nvidia each offer a different programming model. From an academic and industrial point of view, it is therefore necessary to have a flexible and portable implementation of the HPL code, enabling it to be adapted as well as possible to current (and future) architectures, while allowing a maximum number of optimisation options. To guarantee a high level of flexibility and portability, it is possible to use a task-based implementation through an executive support (or runtime). This programming model has already proved its effectiveness in the implementation of various parallel algorithms, in particular for dense linear algebra (LU decomposition, Cholesky decomposition, QR, etc.). In this thesis, we will use Inria’s existing software stack, through the dense linear algebra library Chameleon and the executive support StarPU. These reference libraries for runtime linear algebra will be studied to enable the scaling up of more complex algorithms such as HPL.