More IO hints 






Multidimensional arrays





As with all array-based computing, try to avoid strided memory access
through having the "wrong" inner loop index.  The linear layout of array
elements in memory and the banked nature of memory means that it is
easy to degrade IO and compute performance by having loops interchanged.




  
Fortran arrays:

       the leftmost array index varies fastest in memory
       (i.e. a 2D array has element order
       a(1,1), a(2,1), a(3,1), ...., a(1,2), a(2,2), ...
       in memory).
       So for any nested loop structure, the index of
       the inner DO loop should use the leftmost array index.
       For IO, use something like:
       
       do j=1,ny
          write(8) (c(i,j), i=3,nx-2)
       enddo

       
  
C arrays:

       the rightmost array index varies fastest in memory
       (i.e. a 2D array has element order
       a[0][0], a[0][1], a[0][2], ...., a[1][0], a[1][1], ... in
       memory).
       So for any nested loop structure, the index of
       the inner loop should use the righttmost array index.
       For IO, use something like:
       
       for(i=0;i<nx;i++)
          write(DATA, &c[i][0], 8*ny);
       

       





Fortran buffer size





When using sequential IO to a buffered filesystems from Fortran on the VPP,
it is a simple matter to artificially increase the size of your IO
records.  The default buffer size is reasonably small (4 or 8kB) which
leads to slow standard NFS IO being used from the secondary
PEs.  The -Wl,-g   runtime option
can be used to increase the buffersize - it
specifies the sequential IO buffer size (in kBs) for your program 
to use, eg.

a.out -Wl,-g128

 
causes writes to be buffered to 128kB doing a single write to disk.
Please use this option as much as possible.