Heterogeneous Memory Management for Graph Applications

Nome completo do aluno


Diego Braga Monteiro de Moura


Título do trabalho


Heterogeneous Memory Management for Graph Applications


Resumo do trabalho


We have seen an increased demand for memory in recent years from applications such as big data, streaming analytic, social network, machine learning and others. At the same time we have seen DRAM scalability issues. Not just density issues, but also problems with energy and costs. The main advantage of DRAM technology today is low latency. As an alternative to the problems faced by DRAM scalability, the company Intel launched in 2019 the Intel Optane, the first widely marketed Persistent Memory (PMEM), also known as Non-Volatile Memory (NVM). NVM has high density, low cost and low energy. But also high latency. It's that latency which makes NVM not replace DRAM for a while. Heterogeneous memories, like DRAM+NVM, are likely to become common, similarly to heterogeneous processors (e.g., big.LITTLE[6]), which are already a reality due to its benefits of balancing performance and energy efficiency. With two types of memory in the system, an important research question is how to perform an efficient data placement in order to get the best out of each one. This PhD dissertation focused on management data placement, in a totally real environment, using graph analytics applications. First we study the use of machine learning models with performance counters to predict the performance of an application in a context with multiple applications. In a second moment, we focused on a scenario with a single application, but with scheduling decisions at a lower granularity, at the level of objects/chunks. In this context, we worked on three approaches: (1) offline identification of which attributes most influence the performance of an object, (2) implementation of an offline scheduler at the chunk level, (3) implementation of an online scheduler at the chunk level.




Vinicius Petrucci




Daniel Mossé


Membro Titular Externo 1 (com afiliação)


Paul Carpenter (Barcelona Super Computer)


Link para o curriculum lattes




Membro Titular Externo 2 (com afiliação)


Luis Oliveira (University of Pittsburgh)


Link para o curriculum lattes




Membro Titular Interno 1 ou Titular Externo 3 (com afiliação)


Esbel Tomás Valero Orellana (UESC)


Link para o curriculum lattes




Membro Titular Interno 2 ou Titular Externo 4 (com afiliação)


George Marconi de Araújo Lima (UFBA)


Link para o curriculum lattes





Data da Defesa: 
09/03/2023 - 12:00
Tipo de Defesa: 
Defesa de Doutorado