Skip to Main content Skip to Navigation
Reports

Estimation of the impact of I/O forwarding on application performance

Francieli Zanon Boito 1
1 TADAAM - Topology-Aware System-Scale Data Management for High-Performance Computing
LaBRI - Laboratoire Bordelais de Recherche en Informatique, Inria Bordeaux - Sud-Ouest
Abstract : In high performance computing architectures, the I/O forwarding technique is often used to alleviate contention in the access to the shared parallel file system servers. Intermediate I/O nodes are placed between compute nodes and these servers, and are responsible for forwarding requests. In this scenario, it is important to properly distribute the number of available I/O nodes among the running jobs to promote an efficient usage of these resources and improve I/O performance. However, the impact different numbers of I/O nodes have on an application bandwidth depends on its characteristics. In this report, we explore the idea of predicting application performance by extracting information from a coarse-grained aggregated trace from a previous execution, and then using this information to match each of the application's I/O phases to an equivalent benchmark, for which we could have performance results. We test this idea by applying it to five different applications over three case studies, and find a mean error of approximately 20%. We extensively discuss the obtained results and limitations to the approach, pointing at future work opportunities.
Complete list of metadatas

Cited literature [12 references]  Display  Hide  Download

https://hal.inria.fr/hal-02969780
Contributor : Francieli Zanon Boito <>
Submitted on : Friday, October 16, 2020 - 7:03:00 PM
Last modification on : Tuesday, October 20, 2020 - 10:23:39 AM

File

RR-9366.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02969780, version 1

Collections

Citation

Francieli Zanon Boito. Estimation of the impact of I/O forwarding on application performance. [Research Report] RR-9366, Inria. 2020, pp.20. ⟨hal-02969780⟩

Share