Lustre short read

Short reads.

If you are trying to read a certain amount of data (i.e. 1 MB) the POSIX standard allows the read to return with less data than actually requested (i.e. 500 kB). This is called a short read. The read command will return the length of the record read.

It is the programmers responsiblity to check if the application actually read the amount data requested. If less data have been read this should be handled by the programm i.e. by re-reading the missing data.

The following example ensures that all data requested has been read:

   read:
   if(len > 0) {                                                      //len   - length of the array that is being read.
       count = fread(addr, (size_t)len, (size_t)1, (FILE*) data);     //addr  - buffer, where the data will be read to 
                                                                      //fread - returns the number of bytes read
       if((count<len && count>0 ) || (count==-1 && errno==EINTR)){     
           addr+=count;                                            
           len-=count;
           goto read;
       }
   }

Of course you should also check for reading errors.

Short reads and Other IO libraries

The following I/O libraries handle short reads in a proper way:

• MPIIO
• HDF5 parallel
• NETCDF 4

Issues occur with the following libraries

• C or C++ I/O The programmer has to handle the short read himself
• NETCDF 3 (version 4.3.2 or lower) leads to an unwanted behavior: If a short read occurs netcdf will fill up the missing data in the array with zeroes and returns successfully.

How to convert posix IO to MPIIO

Here is a short summary on how to covert from POSIX IO to MPIIO, which is short read save, so you do not have to handle it in your application.

Operation	C	Fortran	MPIIO	Comment
Open a file	fopen	Open	MPI_File_open(comm, filename, amode, info, fh)	fh corresponds to the C filepointer and the Fortran Unitnumber, info can be MPI_INFO_NULL
Write	fwrite	write	MPI_FILE_WRITE(fh, buf, count, datatype, status)
Read	fread	read	MPI_FILE_READ(fh, buf, count, datatype, status)
Close	fclose	close	MPI_FILE_CLOSE(fh)	Has to be called before MPI_Finalize