[Bug libfortran/29568] implement unformatted files with subrecords (Intel style)

["jvdelisle at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #16 from jvdelisle at gcc dot gnu dot org  2006-11-21 08:25 -------
Here are some relative performance measures.  The first case is with the
program given in comment 14 with a very large record size as indicated.  This
gives before and after patch timing results.  The second case is before an
after with a smaller record size and setting recl=25 in the open statement.

There is a sigmificant performance penalty in this second case.  Lastly, I show
results with patch using -frecord-marker=8. Same penalty. 

Before patch:

$ rm testfile
$ time ./a.out
 reclength=   134217727
 arraysize=   536870908
 record size =           2684354556

real    3m26.163s
user    0m4.107s
sys     0m28.267s
$ ls -l testfile 
-rw-r--r-- 1 jerry jerry 5368709144 Nov 20 23:31 testfile

After patch:

$ rm testfile 
$ time ./a.out
 reclength=   134217727
 arraysize=   536870908
 record size =           2684354556

real    3m25.729s
user    0m4.058s
sys     0m28.657s
$ ls -l testfile 
-rw-r--r-- 1 jerry jerry 5368709144 Nov 20 23:53 testfile


With smaller file and setting RECL=2 in open statement:

Before patch:

$ rm testfile 
$ time ./a.out
 reclength=      500000
 arraysize=     2000000
 record size =             10000016

real    0m0.101s
user    0m0.018s
sys     0m0.083s


After patch:

$ rm testfile 
$ time ./a.out
 reclength=      500000
 arraysize=     2000000
 record size =             10000016

real    0m3.762s
user    0m0.877s
sys     0m2.885s


With -frecord-marker=8:

$ rm testfile 
$ time ./a.out
 reclength=      500000
 arraysize=     2000000
 record size =             10000016

real    0m3.827s
user    0m0.831s
sys     0m2.991s


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29568