IBM gave a webinar on DISK ONLY backups including the advantages of DISK
vs. FILE device classes.  While your mileage may vary, in general it
seems that a FILE devclass will give better performance for large pools
(read TB not GB). Two quick examples:
1) With DISK TSM keeps track of each 4K block in the DISK volumes. This
means that TSM must maintain a map of all those blocks and search/update
that map every time a file is saved/expired.  Also your files will be
fragmented within those DISK volumes leading to further performance
problems.

2) When it's time to backup what's on disk to offsite tapes, TSM has a
speedy shortcut with SEQUENTIAL device classes.  TSM keeps a flag for
each sequential volume, when the sequential volume is backed the flag is
set, and when the sequential volume is written to the flag is cleared.
This means that when it's time to back-up those primary stgpool
sequential volumes to a copypool TSM only needs to examine those files
in the sequential volumes with the flag cleared.  With the way TSM
works, this greatly reduces the amount of time required by TSM to
determine what data needs to be backed up.  With a DISK device class,
TSM has no choice but to examine each file in the STGPOOL being backed
up to determine if it has been previously backed-up to the copypool.

Incidentally, IBM hinted that a future enhancement would be to allow a
list of mount points (directories) to be assigned as the destination to
FILE device classes.  This would allow utilization of dynamic allocation
across multiple file systems.  Of course one drawback with dynamic
allocation is that fragmentation can occur overtime.  Your particular OS
will greatly influence the severity of this problem, however defining
the stgpool volume explicitly will prevent that problem.

H. Milton Johnson