Collective I/O Performance on the Santos Dumont Supercomputer

Abstract

The historical gap between processing and data access speeds causes many applications to spend a large portion of their execution on I/O operations. From the point of view of a large-scale, expensive, supercomputer, it is important to ensure applications achieve the best I/O performance to promote an efficient usage of the machine. In this paper, we evaluate the I/O infrastructure of the Santos Dumont supercomputer, the largest one from Latin America. More specifically, we investigate the performance of collective I/O operations. By conducting an analysis of a scientific application that uses the machine, we identify large performance differences between the available MPI implementations. We then further study the observed phenomenon using the BT-IO and IOR benchmarks, in addition to a custom microbenchmark. We conclude that the customized MPI implementation by Bull (used by more than 20% of the jobs) presents the worst performance for small collective write operations. Our results are being used to help the Santos Dumont users to achieve the best performance for their applications. Additionally, by investigating the observed phenomenon, we provide information to help improve future MPI-IO collective write implementations.

I/O performance of the Santos Dumont supercomputer

“…This article extends our previous one (Carneiro et al, 2018) by presenting a more comprehensive study of the I/O performance in the SDumont. In addition to the investigation on collective operations, we characterize the workload of the machine and propose a strategy for obtaining information about jobs I/O behaviors from traces.…”

“…In a previous work, we conducted an in-depth investigation of the performance differences between MPI implementations for small collective I/O operations on the SDumont. To the best of our knowledge, ours was the first work (Carneiro et al, 2018) to document this phenomenon.…”

Avaliação de Desempenho no Supercomputador SDumont de uma Estratégia de Decomposição de Domínio usando as Funcionalidades de Mapeamento Topológico do MPI para um Método Numérico de Escoamento de Fluidos.

“…Desta forma, como trabalhos futuros, pretendemos criar estrategias de localidade e acesso ao cache. Também pretendemos avaliar o desempenho para um maior número de nós computacionais com malhas maiores para avaliar o impacto do desempenho das operações coletivas do MPI conforme o trabalho [Carneiro et al 2018b]. Além disso, já iniciamos pesquisas em novas estratégias de E/S buscando uma melhor performance na escrita em disco, visualização e criação de checkpoint durante as simulações.…”

“…O trabalho de [Lima 2017] apresenta diversos métodos e exemplos de divisão de domínio para diversos tipos de malhas em arquiteturas de memoria distribuída. No artigo [Carneiro et al 2018a] foi investigado o desempenho das operações de E / S coletivas no Supercomputados Sdumont mostrando como as diversas implementações do MPI apresentam um gargalo em relaçãò as operações coletivas isso a medida em que aumenta o número de nós computacionais.…”

An Unsupervised Learning Approach for I/O Behavior Characterization

“…In the HPC context, a wide range of factors can negatively impact I/O performance. These factors were evaluated in several studies as in [11], [12], [13], [14], [15]. Moreover, understanding and characterizing a platform can provide extra insights on how the applications should perform I/O operations to obtain the best performance and thus, mitigate these negative factors.…”