The extract that Nicolai provided pretty much nails down the point on fragment sizes. There is no exact match for fragment sizes, it is based on experience and the values listed are generally good for most environments.
On the other hand; A lot of the current tape technologies support fast locate-block which sort of removes the need for fragments, as NBU can tell the drive to fast forward and locate the blocks required for restore. However, using fragments is still a common approach, but not really required anymore. In the old days, you could end up with a restore request where the file/s required for restore was in the middle/last part of a image on tape, and the tape drive could only fast forward to the file, not the block, meaning NBU would have to read through x GB before actually being able to read the data... Hence the use of fragments as we could fast forward to the file on tape where the file was and minimize read of non-relevant data.
Fragment size is not equal to block size!
So, what about block sizes then? In Windows 2003SP1 MS broke the tape.sys driver and a lot of tuning in NBU failed. In SP2 it was again fixed. You can still encounter that drivers cannot support larger block sizes than 64KB. For LTO3, a block size of 256KB seems to perform best. for LTO4/5 it is a bit depending on vendor and driver, but most seem to be able to do 512KB, and some 1MB. The larger buffers the more throughput as we need less I/O interrupts. Fewer I/O interrupts lead to less CPU context switches, and so on...
To tune NBU, create the SIZE_DATA_BUFFERS in <install path>\Veritas\NetBackup\db\config and on the first line enter the size in bytes. 256KB is 262144 bytes. On Linux/UNIX it is /usr/openv/netbackup/db/config.
You can combine above with number of data buffers; NUMBER_DATA_BUFFERS. Each bptm tuple (2 for non-MPX, 1-n:1 for MPX backups). They share the buffer queue. So if you use 32 buffers á 256KB and run 100 concurrent jobs, you would at least need ~820MB of shared memory, and then add memory for the processes. With media servers with not enough memory you could "over tune", and end in problems and failed backups. So use care when tuning.
/A
