Doctoral Research

"Data Placement Strategies for Heterogeneous and Non-Volatile Memories in HPC"

My research has consisted of exploring the opportunities that HPC applications have when facing hybrid memory systems including memories like HBM/MCDRAM, DDR and NVDIMMs.

During the thesis, I have first focused on studying different types of memory in different configurations that applications can access, especially I focused on the opportunities that applications have to access NVDIMMs(Intel Optane), the study also included DDR and HBM.

Then, knowing how to use these technologies, common metrics were used in the memories, and a comparative study was carried out, from which, different behaviours were obtained in the performance of benchmarks and applications already known associated with the metrics. On the other hand, that of the applications, those that might not be so well known in their behaviour, a study was carried out through profiling applications, by allocating them in different types of memories with the aim of being able to identify not only the general affinity of a application towards a type of memory or a metric, but also be able to identify the most important buffers that have this affinity.

Finally, this allows us to have some valid strategies that can be taken by developers as a path prior to development and in this way old and new applications can adapt and prepare to the increasing complexity of memory systems.