First,
thanks to Andreas Gruenbacher,
which told me I should look up the blocksize using debugfs.
In fact is was 1024 bytes, which is perfectly consistent with my
testresults.
I think that some of us were interpreting the
stat.st_blksize in a wrong way, the stat() man page says:
---------
The value st_blksize gives the "preferred" blocksize
for efficient file system I/O. (Writing to a file in
smaller chunks may cause an inefficient read-modify-
rewrite.)
------
"preferred" is the keyword here, it's not the filesystem's blocksize
I think preferred means that you should use 4096bytes because it's
the dimension of a page on x86 CPUs.
That means for IO speed issues, we should look at the st_blksize,
but this has nothing to do with the filesystem's blocksize.
tschüss,
Benno.
