10:00
13:00

Andrès Rubio Proano PhD thesis: “Data placement strategies for heterogeneous and non-volatile memory in high performance computing”

Memory systems on high-performance computing platforms have undergone major changes in recent years. In addition to main memory, storage, and multiple levels of caches, servers are non-uniformly accessed (NUMA) and can have multiple types of memory. For example, high-speed memories (HBM) embedded in the processor as well as non-volatile memories (NVDIMM) have been introduced into the hierarchy. These changes are necessary to bring the data closer to the processing and thus have better performance. However, they force developers to adapt their applications to run properly on these different heterogeneous systems, which makes development much more complex. In practice, simply deciding to allocate a data buffer on the right type of memory in these heterogeneous systems is difficult and critical for the performance of the application. This thesis was carried out at Inria Bordeaux - Sud-Ouest and at LaBRI.
After presenting the state of the art of memory architectures, we characterized the different types of memory using simple attributes. We provided an interface that the hwloc library exposes to applications to understand the memory organization and allocate buffers. We then proposed a methodology for developers to tailor their applications for appropriate use of heterogeneous memory systems. Since access to different heterogeneous platforms is not always possible, we have identified numerous strategies for simulating heterogeneous memory performance, and emulating different memory topologies. Finally, we designed a strategy to facilitate sharing of non-volatile and heterogeneous memory platforms between HPC tasks co-executed on the same servers.

Inria (Salle Ada Lovelace)