Towards On-Demand I/O Forwarding in HPC Platforms Jean Luca Bez, - - PowerPoint PPT Presentation

▶
towards on demand i o forwarding in hpc platforms
SMART_READER_LITE
LIVE PREVIEW

Towards On-Demand I/O Forwarding in HPC Platforms Jean Luca Bez, - - PowerPoint PPT Presentation

Jul 10, 2023 224 likes •506 views

Towards On-Demand I/O Forwarding in HPC Platforms Jean Luca Bez, Francieli Zanon Boito, Ramon Nou, Alberto Miranda, Toni Cortes, and Philippe O. A. Navaux jean.bez@inf.ufrgs.br PDSW 2020 International Parallel Data Systems Workshop

slide-1
SLIDE 1

Jean Luca Bez, Francieli Zanon Boito, Ramon Nou, Alberto Miranda, Toni Cortes, and Philippe O. A. Navaux

jean.bez@inf.ufrgs.br

PDSW 2020 — International Parallel Data Systems Workshop

Towards On-Demand I/O Forwarding in HPC Platforms

slide-2
SLIDE 2

INTRODUCTION

Agenda

  • The I/O Forwarding Layer
  • Motivation
  • FORGE The I/O Forwarding Explorer
  • Forwarding in MareNostrum 4
  • Forwarding in SDumont
  • Conclusion

2

slide-3
SLIDE 3

INTRODUCTION

The I/O Forwarding Layer

3

Application A Application B Application C Application D Application E Application X

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Metadata Servers Data Servers

Meta Server 1 Meta Server 2 Meta Server M

  • ● ●

Data Server 1 Data Server 2 Data Server N

  • ● ●

Compute Nodes Parallel File System

slide-4
SLIDE 4

INTRODUCTION

The I/O Forwarding Layer

4

Application A Application B Application C Application D Application E Application X

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Metadata Servers Data Servers

Meta Server 1 Meta Server 2 Meta Server M

  • ● ●

Data Server 1 Data Server 2 Data Server N

  • ● ●

Compute Nodes Parallel File System

slide-5
SLIDE 5

INTRODUCTION

The I/O Forwarding Layer

5

Application A Application B Application C Application D Application E Application X

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Metadata Servers Data Servers

ION 2 ION 3 ION K

  • ● ●

ION 1 Meta Server 1 Meta Server 2 Meta Server M

  • ● ●

Data Server 1 Data Server 2 Data Server N

  • ● ●

Compute Nodes Parallel File System Forwarding Layer

slide-6
SLIDE 6

INTRODUCTION

The I/O Forwarding Layer

6

Application A Application B Application C Application D Application E Application X

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Client Client Client

  • ● ●

Metadata Servers Data Servers

ION 2 ION 3 ION K

  • ● ●

ION 1 Meta Server 1 Meta Server 2 Meta Server M

  • ● ●

Data Server 1 Data Server 2 Data Server N

  • ● ●

Compute Nodes Parallel File System Forwarding Layer

slide-7
SLIDE 7

INTRODUCTION

Motivation

7

  • Investigate the impact of I/O forwarding on performance
  • Take into account the application’s access pattern
  • Most machines cannot be easily reconfigured
  • End-users are not allowed to change this layer
  • We need a research/exploration alternative!
  • When forwarding is the best choice?
  • How many I/O nodes should an application use?
slide-8
SLIDE 8

ARCHITECTURE

FORGE: The I/O FORwardinG Explorer

8

Compute Node FORGE CN Compute Node FORGE ION Compute Node FORGE ION Compute Node

Parallel File System

1 2 6 10 14 FORGE CN 3 7 11 15 FORGE CN 4 8 12 16 FORGE CN 5 9 13 16 Compute Node Compute Node MPI Rank

slide-9
SLIDE 9

FORGE

EXPERIMENTS

9

  • MareNostrum 4 (Spain) and Santos Dumont (Brazil) supercomputers
  • 189 distinct scenarios (access patterns and deployments):

○ Compute nodes: 8, 16, and 32 ○ Client processes per compute node: 12, 24, and 48 (96, 192, 384, 768, and 1536 processes in total) ○ File layout: file-per-process or shared file ○ Spatiality: contiguous or 1D-strided ○ Operation: WRITE ○ Request sizes: 32KB, 128KB, 512KB, 1MB, 4MB, 6MB, and 8MB ○ Stonewall: one second

slide-10
SLIDE 10

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

10

slide-11
SLIDE 11

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

11

slide-12
SLIDE 12

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

12

slide-13
SLIDE 13

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

13

slide-14
SLIDE 14

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

14

slide-15
SLIDE 15

I/O FORWARDING

MareNostrum 4

  • Bandwidth at client-side
  • 5 repetitions for each
  • Different days and periods

15

slide-16
SLIDE 16

I/O FORWARDING

MareNostrum 4

  • How many choices do we have to consider?
  • Dunn’s nonparametric test
  • 3 choices impact performance

46% patterns (88 out of 189)

  • What is the best number of I/O nodes?
  • No simple rule to fit all

16

slide-17
SLIDE 17

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

17

slide-18
SLIDE 18

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

18

slide-19
SLIDE 19

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

19

slide-20
SLIDE 20

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

20

slide-21
SLIDE 21

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

21

slide-22
SLIDE 22

I/O FORWARDING

Santos Dumont

  • Forwarding impact is different!
  • The more I/O nodes, the better
  • Not forwarding is an option

22

slide-23
SLIDE 23

RESULTS

Discussion

23

  • Increasing heterogeneous applications
  • Shift from must-use to on-demand I/O forwarding layer
  • Transparently reshape the flow of requests
  • Towards a dynamic allocation of I/O nodes
  • Idle or reserved set of compute nodes could act as I/O nodes
  • Interference on I/O could not be reduced or eliminated
slide-24
SLIDE 24

PRESENTATION

Conclusion

24

  • I/O forwarding is an established and widely-adopted technique
  • Not always possible to explore its advantages under different setups
  • Impact or disrupt production systems
  • FORGE: a lightweight forwarding layer in user-space
  • Understand the impact of forwarding different access patterns
  • Evaluation in MareNostrum 4 and Santos Dumont supercomputers
  • Shift from must-use to on-demand I/O forwarding layer
slide-25
SLIDE 25

25

ACKNOWLEDGMENTS

This study was financed by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001. It has also received support from the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), Brazil; It is also partially supported by the Spanish Ministry of Economy and Competitiveness (MINECO) under grants PID2019-107255GB; and the Generalitat de Catalunya under contract 2014—SGR—1051. The author thankfully acknowledges the computer resources, technical expertise and assistance provided by the Barcelona Supercomputing Center - Centro Nacional de Supercomputación. The authors acknowledge the National Laboratory for Scientific Computing (LNCC/MCTI, Brazil) for providing HPC resources of the SDumont supercomputer, which have contributed to the research results reported within this paper. URL: http://sdumont.lncc.br.

slide-26
SLIDE 26

Jean Luca Bez, Francieli Zanon Boito, Ramon Nou, Alberto Miranda, Toni Cortes, and Philippe O. A. Navaux

jean.bez@inf.ufrgs.br

PDSW 2020 — International Parallel Data Systems Workshop

Towards On-Demand I/O Forwarding in HPC Platforms