I/O performance blues

[Salvatore Filippone]

Hello
In the last few days I have investigated a rather disappointing
phenomenon related to I/O.
The Harwell-Boeing storage format is a rather well-known format for
storing matrix data on file; sample codes and pointers to further
documentation are available at http://math.nist.gov/MatrixMarket/

The input is via read statements as follows
==================================================
    open(infile,file=filename, status='OLD', err=901, action='READ')

    read (infile,fmt=fmt10) mtitle_,key,totcrd,ptrcrd,indcrd,valcrd,rhscrd,&
         & type,nrow,ncol,nnzero,neltvl,ptrfmt,indfmt,valfmt,rhsfmt
    if (rhscrd > 0) read(infile,fmt=fmt11)rhstype,nrhs,nrhsix
    allocate(icp(ncol+1),ia(nnzero),val(nnzero),stat=ircode)
    if (ircode /= 0 ) then
      write(*,*) 'Memory allocation failed'
      goto 993
    end if

    read (infile,fmt=ptrfmt) (icp(i),i=1,ncol+1)
    read (infile,fmt=indfmt) (ia(i),i=1,nnzero)
    if (valcrd > 0) read (infile,fmt=valfmt) (val(i),i=1,nnzero)

    close(infile)
===================================================
Now, at the Matrix Market site there is an alternative implementation
of this I/O in C.
It turns out the C version consistently outperforms the Fortran version!!

[sfilippo@epsilon testhb]$ ./test_hbrd
 File name?
ultrasound80.hb
 Reading from matrix file: ultrasound80.hb
  nrows:      531441  ncols:      531441  nnzeros:    33076161
 C       Time read for ultrasound80.hb:   18.724172830581665
 Fortran Time read for ultrasound80.hb:   33.051491975784302
[sfilippo@epsilon testhb]$ ./test_hbrd
 File name?
thm_6814k.hb
 Reading from matrix file: thm_6814k.hb
  nrows:     6813804  ncols:     6813804  nnzeros:    33930324
 C       Time read for thm_6814k.hb:   19.886425018310547
 Fortran Time read for thm_6814k.hb:   38.642944097518921
[sfilippo@epsilon testhb]$ ./test_hbrd
 File name?
pde-50.hb
 Reading from matrix file: pde-50.hb
  nrows:      125000  ncols:      125000  nnzeros:      860000
 C       Time read for pde-50.hb:  0.32523894309997559
 Fortran Time read for pde-50.hb:  0.78256392478942871
====================================================
This is with current trunk; I have seen fluctuations in the exact
timings when trying 4.6.4/4.7.3/4.8.2 but the overall picture is
pretty much the same.

I have the full test program and many input cases available; is
anybody familiar with the IO in gfortran and willing to take a look?
And/or give me any suggestions where to look for clues and how? Should
I open a PR?


Thanks a lot
Salvatore