Lecture 21: Parallel Filesystems Abhinav Bhatele, Department of - - PowerPoint PPT Presentation

▶
lecture 21 parallel filesystems
SMART_READER_LITE
LIVE PREVIEW

Lecture 21: Parallel Filesystems Abhinav Bhatele, Department of - - PowerPoint PPT Presentation

Apr 06, 2023 335 likes •457 views

Introduction to Parallel Computing (CMSC498X / CMSC818X) Lecture 21: Parallel Filesystems Abhinav Bhatele, Department of Computer Science Announcements Interim report was due yesterday Assignment 4 is due next week Abhinav Bhatele

slide-1
SLIDE 1

Lecture 21: Parallel Filesystems

Abhinav Bhatele, Department of Computer Science

Introduction to Parallel Computing (CMSC498X / CMSC818X)

slide-2
SLIDE 2

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Announcements

  • Interim report was due yesterday
  • Assignment 4 is due next week

2

slide-3
SLIDE 3

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Parallel filesystem

  • Home directories and scratch space typically on a parallel file system
  • Mounted on all login and compute nodes
  • Also referred to as I/O sub-system

3

http://wiki.lustre.org/Introduction_to_Lustre

slide-4
SLIDE 4

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Links between cluster and filesystem

4

Leaf Switch

Each SU (1 management node, 1 login node, 2 LNET router nodes, 2 gateway nodes) 9 9

Leaf Switch

36 total 9 9 9 9 9 9 LNET router node Object storage server (OSS) Compute node 9 total

slide-5
SLIDE 5

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Different parallel filesystems

  • Lustre: open-source (lustre.org)
  • GPFS: General Parallel File System from IBM, now called Spectrum Scale
  • PVFS: Parallel

Virtual File System

5

slide-6
SLIDE 6

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Tape drive

  • Store data on magnetic tapes
  • Used for archiving data
  • Use robotic arms to access the right tape: https://www.youtube.com/watch?v=d-

eWDuEo-3Q

6

slide-7
SLIDE 7

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Burst buffer

  • Fast, intermediate storage between compute nodes and the parallel filesystem
  • Two designs:
  • Node-local burst buffer
  • Remote (shared) burst buffer

7

slide-8
SLIDE 8

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

I/O libraries

  • High-level libraries: HDF5, NetCDF
  • Middleware: MPI-IO
  • Low-level: POSIX IO

8

slide-9
SLIDE 9

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

Different I/O patterns

  • One process reading/writing all the data
  • Multiple processes reading/writing data from/to shared file
  • Multiple processes reading/writing data from/to different files
  • Different performance depending upon number of readers/writers, file sizes,

filesystem etc.

9

slide-10
SLIDE 10

Abhinav Bhatele (CMSC498X/CMSC818X) LIVE RECORDING

I/O profiling tools

  • Darshan
  • Lightweight profiling tool from Argonne National Lab
  • Recorder
  • Research prototype from UIUC

10

slide-11
SLIDE 11

Abhinav Bhatele 5218 Brendan Iribe Center (IRB) / College Park, MD 20742 phone: 301.405.4507 / e-mail: bhatele@cs.umd.edu