Long tape records

[utzoo!decvax!ucbvax!unix-wizards]

>From ARPAVAX.ghg at Berkeley Mon Nov  2 12:18:03 1981
Purdue/EE School uses 100 block (51200 byte) tape records for disk
full dumps (epoch). The Unix "dump" program is not used for the
full dumps. "Dbuf", a locally written program similar to "dd"
in syntax, is used to make an image copy of disk filesystems
on tape. Dbuf forks into 4 processes, one reads, one writes,
one sends signals, and one is a watchdog. The read and write
processes swap roles on alternating records, thereby providing
the effect of double-buffered I/O.  Our tape drive  and disk transfer
rates are about the same, so the time required for a copy is
roughly equivalent to the hardware transfer rate (both are
transferring at same time). Tape drive is a Telex 6253,
800/1600/6250 BPI @125 in/sec on an AVIV Unibus controller.
Disks are CDC 9766's on System Industries 9400 SBI controllers
(looks like DEC MBA to software). A 65500 block (512-bytes/block)
filesystem can be dumped in 1-1/4 mins. Five 65500 block
filesystems will fit on one 2400' reel of tape. System is a
VAX 11/780 (dual cpu) running Berkeley 4.1BSD UNIX. Incremental
dumps are still made via dump 6 days/week, and the full dumps
are done on Sun morn. Both the disk and tape systems are pretty
much rock solid (tape sys costs approx $30K). We use Graham
Magnetics tape (Verituf or Epoch 480) which are suppose to
have 1 and 3 Million pass lives. Dbuf is also almost twice as
fast as dd for disk-disk xfers (on different controllers or
MBAs) and the same speed as dd for the same controller or UBA.
copies

There are no major problems with devices getting locked out
for the long transfers, the SI disk controllers have a 4 sector
buffer and the tape controller has a 4Kbyte FIFO for Unibus
latency. There are even 3 1MBaud DMC-11's on the same Unibus
which continue to function, although a little degraded. SI
disks don't data-late like DEC ones do, they just miss revs.
For this scheme to work there can be no badspots on the disk.
One must either have good packs (we use Dysan) or use the
Auto-badblock-forwarding in 4.1BSD to make them invisible.

Geo. Goble
--ghg (ucbvax!pur-ee!ghg or CSVAX.ghg at Berkeley)