Consider the following program:
program kl
integer i,j,k
integer, parameter :: m = 1000, n = 387
real x(m,n)
x = 1.e0
inquire(iolength=k) (x(i,1), i=1, m)
open(unit=1, file='foo.dat', access='direct', recl=k)
do j = 1, n
write(1,rec=j) (x(i,j), i=1, n)
end do
close(1)
end program kl
We find the following performance on multiple executions.
kargl[225] rm foo.dat
kargl[227] gfc40 -static -o z kl.f90
kargl[228] time ./z
0.11 real 0.05 user 0.05 sys
kargl[229] time ./z
29.44 real 0.36 user 28.13 sys
kargl[230] time ./z
28.77 real 0.25 user 28.07 sys
kargl[231] rm foo.dat
kargl[232] gfc41 -static -o z kl.f90
kargl[233] time ./z
0.23 real 0.06 user 0.05 sys
kargl[234] time ./z
0.11 real 0.05 user 0.06 sys
kargl[235] time ./z
0.11 real 0.03 user 0.08 sys
Here gfc40 is an unpatched gfortran from 4.0.1, and
gfc41 is a patched gfortran from HEAD. Now, image the
performance if m = 1080 and n = 25000. Oh, the inhumanity!
See attached patch for suggested fix.
--
Summary: Really, really, horrible IO performance
Product: gcc
Version: 4.0.0
Status: UNCONFIRMED
Severity: normal
Priority: P2
Component: libfortran
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: kargl at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21820
