------- Additional Comments From tkoenig at gcc dot gnu dot org 2005-07-18 14:35 ------- (In reply to comment #2)
> ultimately, things have to be written by a system call, and a system call is > expensive. (One system call per array element is out of the question.) With the current implementation (which may not be the best for performance) we use memory-mapped I/O. It then becomes a question of how efficient you transfer your data into the memory you mmapped. > Thus > we either dump the whole array to disk, byte for byte as it is in memory, or > we have to transform it into a temporary buffer and then write this buffer. Agreed. > C's fwrite uses a buffer (probably), fprintf does for sure. > > If you remove the padding, then you have to allocate and deallocate these > buffers If you use read() instead of mmap(), it is enough to allocate a single buffer, which can be reused. >and pack or unpack the array elements explicitly. Depending on your > I/O library, that may be a lot of work. I tried to show in my little benchmark above that packing things may be a little less work than actually copying the whole thing. This doesn't take disk I/O into account, which is also fairly slow. > And in the end you are still > incompatibly to C. Fortran unformatted I/O is incompatible to C anyway, because of the record markers, this is not really a big deal. I think a few benchmarks could help there. > (I'm afraid we may be going to repeat a discussion here that was already held > on the mailing list.) If there's a PR, I think it's better if people can read the discussion in its audit trail. The biggest problems in I/O performance at the moment are the overhead of individual function calls for unformatted array I/O. I think the best way to address that would be to let the library use array descriptors. For the general case, this would require pack / unpack operations, so a conversion to 10 bytes for numbers aligned to 16 bytes would only help. Hmm.... -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22519