[Info-vax] Backup TK50 tapes

[glen herrmannsfeldt]

Bob Koehler <koehler at eisner.nospam.encompasserve.org> wrote:

(snip)
>   Even the DR780, which is a lot like an IBM channel, did semi-infinite
>   transfers as many smaller blocks.

(snip, then I wrote)
>> For S/360, you can write larger blocks with data chaining. That is,
>> more than one CCW for the block. (Among others, it allows for
>> scatter/gather I/O operations.)  Also, as you say, the timing is
>> critical. CCWs have to be fetched from memory (self modifying 
>> channel programs, or system modifying while they are running,
>> are both allowed). So, the channel has to fetch the new CCW and
>> continue the operation while data transfer continues.
 
>   Exactly what the DR780 did.  But that doesn't mean the I/O device at 
>   the other side of the CCW has an inifite number of bits in its length 
>   register.

I/O devices had very little buffering. When your whole computer had 8K
(the low-end S/360 models) how much are you going to put in an I/O
device?

So, the whole bus-and-tag and channel system was designed to synchronize
to the I/O controller. There is full  handshaking between the channel
and the I/O device. Normal bus-and-tag has an eight bit data bus, the
high-speed version uses an additional cable to run 16 bits. The channel
has to buffer at least the width of the data bus coming from memory,
which might be 64 bits on the high end machines. 

There are whole manuals describing the timing such that you can
figure out which devices, and under what load, can data chain.
Likely only slow devices on the smaller machines. Also, devices
nearer the channel (on daisy-chained cables) can run faster.

But yes, recent generations of I/O devices buffer whole blocks.
I believe I wrote 100K blocks on an Ultium-1 drive.

-- glen