Skip to Main content Skip to Navigation
Theses

Data Placement Strategies for Heterogeneous and Non-Volatile Memories in High Performance Computing

Abstract : Memory systems in High-Performance Computing (HPC) systems have undergone major changes in recent years. Beside the main memory, the storage and multiple levels of caches, the servers come with non-uniform memory access (NUMA) and may contain different kinds of memory. For instance, high bandwidth memory (HBM) embedded on the processor package and non-volatile memory (NVDIMM) have been introduced into the hierarchy. These changes are necessary to bring the data closer and closer to processing and therefore have better performance. However, they require developers to adapt their applications to work properly on different heterogeneous memory systems, causing software development to become much more complex. In practice, the simple fact of deciding to allocate a data buffer on the appropriate memory in a heterogeneous system becomes difficult and critical to application performance.This thesis has been carried out at Inria Bordeaux - Sud-Ouest and LaBRI. After a presentation of the state of the art of memory architectures, we have characterised the memories through simple attributes. We have provided an interface that the hwloc library exposes to applications to understand the memory topology and allocate buffers. Then, we proposed a strategy to help developers adapt their applications for the proper use of heterogeneous memory systems. As accessing different heterogeneous platforms is not always possible, we identify several ways to simulate the performance of heterogeneous memory and to emulate different memory topologies. Finally, we built a strategy that eases the sharing of platforms with heterogeneous and non-volatile memory between HPC tasks co-scheduled on the same nodes.
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03431281
Contributor : ABES STAR :  Contact
Submitted on : Tuesday, November 16, 2021 - 3:48:10 PM
Last modification on : Tuesday, June 14, 2022 - 8:26:53 AM
Long-term archiving on: : Thursday, February 17, 2022 - 8:05:13 PM

File

RUBIO_PROANO_ANDRES_2021.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03431281, version 1

Collections

Citation

Andrès Rubio Proaño. Data Placement Strategies for Heterogeneous and Non-Volatile Memories in High Performance Computing. Distributed, Parallel, and Cluster Computing [cs.DC]. Université de Bordeaux, 2021. English. ⟨NNT : 2021BORD0224⟩. ⟨tel-03431281⟩

Share

Metrics

Record views

92

Files downloads

189