I/O Strategies for the T3E

1
I/O Strategies for the T3E

• Jonathan Carter
• NERSC User Services

2
T3E Overview

• T3E is a set of Processing Elements (PE)
  connected by a fast 3D torus.
• PEs do not have local disk
• All PEs access all filesystems equivalently
• Path for I/O generally looks like
• user buffer space
• system buffer space
• I/O device buffer space

3
Filesystems

• /usr/tmp
• fast
• subject to 14 day purge, not backed up
• check quota with quota -s /usr/tmp (usually 75Gb
  and 6000 inodes)
• TMPDIR
• fast
• purged at end of job or session
• shares quota with /usr/tmp
• HOME
• slower
• permanent, backed up
• check quota with quota (usually 2Gb and 3500
  inodes)

4
Types of I/O

• Language I/O Fortran or C (ANSI or POSIX)
• Cray FFIO library (can be used from Fortran or C)
• MPI I/O
• Cray extensions to Fortran and C I/O (mostly for
  compatibility with PVP systems)

5
I/O Strategies - Exclusive access files

• Each PE reads and writes to a separate file
• Language I/O
• MPI I/O
• Increase language I/O performance with FFIO
  library (C must use POSIX style calls)

6
I/O Strategies - Communication and I/O PE

• One PE coordinates reading and writing and
  communicates data back and forth between other
  PEs via message passing
• Language I/O
• MPI I/O
• Increase language I/O performance with FFIO
  library

7
I/O Strategies - Shared files

• All PEs read and write the same file
  simultaneously
• Language I/O with FFIO library global layer
• MPI I/O
• Language I/O with FFIO library global layer and
  Cray extensions for additional flexibility

8
Cray FFIO library

• FFIO is a set of I/O layers tuned for different
  I/O characteristics
• Buffering of data (configurable size)
• Caching of data (configurable size)
• Available to regular Fortran I/O without
  reprogramming
• Available for C through POSIX-like calls, e.g.
  ffopen, ffwrite

9
The assign command

• the assign command controls
• controls which FFIO layer is active
• striping across multiple partitions
• lots more
• scope of assign
• File name
• Fortran unit number
• File type (e.g. all sequential unformatted files)

10
assign Examples

• read and write to file restart.file from all PEs
  by using the FFIO library global layer
• assign -F global1282 frestart.file
• use the FFIO library bufa layer to improve
  performance for file opened on Fortran unit 10
• assign -F bufa1282 u10
• use the FFIO library bufa layer to improve
  performance for all unformatted sequential
  Fortran files
• assign -F bufa1282 gsu

11
assign Examples

• To see all active assigns
• assign -V
• To remove all active assigns
• assign -R

12
bufa FFIO layer

• bufa is an asynchronous buffering layer
• performs read-ahead, write-behind
• specify buffer size with -F bufabsnbufs where
  bs is the buffer size in units of 4Kbyte blocks,
  and nbufs is the number of buffers
• buffer space increases your applications memory
  requirements

13
global FFIO layer

• global is a caching and buffering layer which
  enables multiple PEs to read and write to the
  same file
• if one PE has already read the data, an
  additional read request from another PE will
  result in a remote memory copy
• file open is a synchronizing event
• By default, all PEs must open a global file, this
  can be changed by calling GLIO_GROUP_MPI(comm)
• specify buffer size with -F globalbsnbufs where
  bs is the buffer size in units of 4Kbyte blocks,
  and nbufs is the number of buffers per PE

14
File positioning with the global FFIO layer

• Positioning of a read or write is your
  responsibility
• File pointers are private
• Fortran
• Use a direct access file, and read/write(recnum)
• Use Cray extensions setpos and getpos to
  position file pointer (not portable)
• C
• Use ffseek

15
FFIO considerations

• Examples above use an unblocked file structure,
  normal Fortran files are blocked. To read the
  file without the global or bufa layers you must
  use
• assign -s unblocked ffilename
• bufa and global do not allow backspace, or
  skipping over a partially read record. You can
  allow this behavior by using the cos layer in
  addition to bufa or global, but then setpos
  doesnt work.
• assign -s cos128,bufa1282 ffilename

16
More on FFIO

• There are many other FFIO layers, some pretty
  obscure
• cache and cachea layers, good for random access
  files
• man intro_ffio for a terse description
• Cray Publication - Application Programmers I/O
  Guide

17
More on assign

• Many text processing options
• Switch between Fortran 77 and Fortran 90 namelist
• File pre-allocation
• File striping

18
Further Information

• I/O on the T3E Tutorial by Richard Gerber at
  http//home.nersc.gov/training/tutorials
• Cray Publication - Application Programmers I/O
  Guide
• Cray Publication - Cray T3E Fortran Optimization
  Guide
• man assign

19
MPI I/O

• Part of MPI-2
• Interface for High Performance Parallel I/O
• data partitioning
• collective I/O
• asynchronous I/O
• portability and interoperability

20
MPI I/O Definitions

• An MPI file is an ordered collection of MPI
  types.
• A file may be opened individually or collectively
  by a group of processes
• The fileview defines a template for accessing the
  file and is used to partition the file amongst
  processes

21
Fileviews

• A fileview is composed of three pieces
• a displacement (in bytes) form the beginning of
  the file
• an elementary datatype (etype), which is the unit
  of data access and positioning within the file
• an filetype, which defines a template for
  accessing the file. A filetype can contain etypes
  or holes of the same extent as etypes.

22
Fileviews (cont.)

• The filetype pattern is repeated, tiling the
  file
• Only the non-empty slots are available to read or
  write

23
Fileview (cont.)

• Each process can have a different filetype
• Process 0
• Process 1
• Process 2

24
MPI_File_set_view

• Called after MPI_File_open to set fileview
• MPI_File_set_view(fh, disp, etype, filetype,
  datarep, info)
• fh is a file handle
• disp, etype, and filetype define the fileview
• datarep is one of native, internal, or
  external32
• info is a set of hints to optimize performance

25
MPI Info object

• An info object bundles up a set of parameters
• integer finfo
• call MPI_Info_create(finfo, ierr)
• call MPI_Info_set(finfo, access_style,
  write_mostly, ierr)
• MPI I/O defines a set of parameters used to help
  optimize I/O performance
• MPI_Info_null can be used instead of an info
  object

26
Open and Close

• MPI_File_open(comm, filename, amode, info, fh)
• comm, open is collective over this communicator
• filename, string or character variable
• file access mode MPI_MODE_RDONLY, MPI_MODE_RDWR
  etc.
• info object, used to pass hints to open
• file handle
• MPI_File_close(fh)

27
Utility routines

• MPI_File_delete
• MPI_File_set_size
• MPI_File_preallocate
• MPI_File_set_info

28
Query routines

• MPI_File_get_size
• MPI_File_get_group
• MPI_File_get_amode
• MPI_File_get_info
• MPI_File_get_view

29
Data access routines

• Positioning
• Explicit, each call has an offset
• Individual, each PE maintains an individual file
  pointer
• Shared, the file pointer is maintained globally
• Synchronism
• Blocking, routine returns when complete
• Non-blocking, must call a termination routine to
  ensure completion
• Coordination
• Non-collective
• Collective

30
Summary of access routines
31
Summery of access routines (cont.)

• MPI_File_seek
• MPI_File_get_position
• MPI_File_get_byte_offset
• MPI_File_seek_shared (collective)
• MPI_File_get_position_shared

32
T3E Implementation

• No shared file pointers
• No non-blocking collective (split collective)
• SPR filed on non-blocking read
• Work in progress

33
Examples

• All the program fragments are available as
  working programs on the T3E
• Do module load training, then look in
  EXAMPLES/mpi_io
• All examples are of a distributed dot product
• initialize data with random numbers
• compute dot product of whole vector
• write out data into a shared file
• read back in and check dot product

PE 0
PE 1
PE 2
34
Naming convention

• First letter is positioning explicit,
  individual, or shared
• Second letter is synchronism blocking or
  non-blocking
• Third letter is coordination non-collective or
  collective
• ebn.f90 is the explicit, blocking non-collective
  example
• There are several ibn examples dealing with
  different fileviews

35
Filetype Example

• Process 0
• Process 1
• Process 2

36
Filetype Example
filemode MPI_MODE_RDWR MPI_MODE_CREATE call
MPI_INFO_CREATE(finfo, ierr) call
MPI_INFO_SET(finfo, 'access_style','write_mostly',
ierr) call MPI_FILE_OPEN(MPI_COMM_WORLD,
'vector', filemode, finfo, fhv, ierr) call
MPI_TYPE_CREATE_SUBARRAY(1, mnprocs, m, mme,
MPI_ORDER_FORTRAN, MPI_REAL, mpi_fileslice,
ierr) disp0 call MPI_FILE_SET_VIEW(fhv, disp,
MPI_REAL, mpi_fileslice, 'native',
MPI_INFO_NULL, ierr)
37
Individual, blocking, non-collective
call MPI_FILE_WRITE(fhv, b, m, MPI_REAL, status,
ierr) lresultsdot(m, b, 1, b, 1) call
MPI_REDUCE(lresult, result, 1, MPI_REAL, MPI_SUM,
0, MPI_COMM_WORLD, ierr) if (me.eq.0) then
write(6,) 'dot product ', result end if ! zero
vector and read it back in b0.0 disp0 call
MPI_FILE_SEEK(fhv, disp, MPI_SEEK_SET, ierr) call
MPI_FILE_READ(fhv, b, m, MPI_REAL, status, ierr)
38
Further Information on MPI I/O

• MPI-The Complete Reference
• Volume 1, The MPI Core
• Volume 2, The MPI Extensions