MPI-I/O

Performance:

• writing a large distributed array (500MB)
• distributed I/O with MPI2 versus I/O via master task:

  
  

  

  
  

• send&write faster (even across nodes)
• MPI-I/O depends less on block sizes
• 100 MB/s maximal write over fibre channel
• user I/O $\neq$ disk I/O (HP-UX buffer cache, FC60 cache memory)

MPI-I/O across hosts:

• Each host holds a file of the complete size
• contains valid data for the local tasks
• contains 0's for the remote tasks