[Info-vax] booting vaxstation off alpha

[Johnny Billquist]

On 2013-02-13 23:05, Chris Scheers wrote:
> Johnny Billquist wrote:
>> On 2013-02-12 23:10, Chris Scheers wrote:
>>> Johnny Billquist wrote:
>>>> On 2013-02-11 13:44, Stephen Hoffman wrote:
>>>>> On 2013-02-11 07:54:11 +0000, Hans Vlems said:
>>>>>
>>>>>> DECnet over DSSI works fine, provided you get the incantations right.
>>>>>> All I tried was CTERM and FAL and both worked as expected (reliable
>>>>>> albeit slow).
>>>>>> Another example of an undocumented, unsupported feature that works
>>>>>> alright.
>>>>>
>>>>> Not that there is even a remote chance of seeing DSSI gear around, nor
>>>>> any likelihood of IP over FC nor connections, but...
>>>>>
>>>>> If you ran any tests[1] with that, how well does that "albeit slow"
>>>>> connection perform as compared with slow Ethernet?  CI wasn't known
>>>>> for
>>>>> its network performance, as compared with DECnet over even
>>>>> then-current
>>>>> 10 Mb Ethernet, and usual recommendations back then had CI at higher
>>>>> cost as a backup connection.  I can't see DSSI being much better in
>>>>> that
>>>>> regard.
>>>>
>>>> That sounds weird. Do you know why?
>>>> I mean, CI was after all two redundant full duplex 70 Mbit/s channels,
>>>> compared to the half-duplex 10Mbit/s ethernet. Not to mention the fact
>>>> that the MTU of CI is much larger.
>>>
>>> IIRC, CI is a form of token ring.  In particular, it has various "slots"
>>> that circulate that are allocated to applications.
>>
>> Nope. CI works similar to ethernet. It just checks if the line is free
>> before starting to transmit.
>> However, it is also different than ethernet in that each node has an
>> individual delay once the path goes silent, and you can only grab the
>> path if it's been silent "long enough", which is different for each
>> node. And collisions are detected by the fact that all messages on CI
>> are expected to be ACKed by the receiver. If no ACK comes back, you
>> retransmit.
>
> That's interesting.  I had been told it was a "token ring" type bus.
> Actually, now that I think of it, the description was more of a TDM type
> bus, but I don't remember if that term was common back then.

I can't remember for sure, but I think TDM did exist back then too. But 
CI don't match that any more than it matches "token ring".

> So CI is a full blown collision system?  Interesting.  Do you known of
> any technical documents on the net describing this?

It don't do collision detection the same way ethernet do.
Basically, if the cable is quite, you transmit. If the cable is busy, 
you wait until it is quiet, and then you wait a predefined time which is 
different for each CI node number, so that no two CI nodes will ever use 
the same wait time. This insures that after a transmission you will not 
get a collision if two nodes both want to send. If it is quiet you do 
have the possibility that you will get collision. However, that does not 
get detected as such.
The positive acknowledgement deals with that part. Any transmission 
needs to be acknowledged, and if no ACK is received, you do a 
retransmit. And if you have a collision at transmit, the received will 
get a scrambled message, and will not acknowledge. And since both nodes 
that tried to transmit failed, they will wait their individual 
predefined times, so at retransmit time, they will not collide again, 
and both will succeed. The last detail in this is that ACK for a 
transmission is done right after the transmission. All wait times for 
nodes are longer than the wait time before an ACK, so the ACK never have 
collision, and both nodes in an actual collision will know before they 
are allowed to send again if the previous send failed.

Hoff posted a link yesterday: 
http://lazowska.cs.washington.edu/p130-kronenberg.pdf, which have at 
least some half-good details.

CI is CSMA. Just not the /CD, which ethernet have.

>> As for allocation of traffic by type, this is totally up to the OS,
>> and CI itself don't give a damn. DEC defined SCA as the communications
>> layer on top of CI, and SCA have datagrams, messages and block data.
>
> It could very well be a VMS/driver construct.  I do know that when we
> did some trivial throughput tests, DECnet on the CI was slower than the
> Ethernet.  Of course, since we were doing disk to network to disk
> copies, the disk traffic also went over the CI.

That would of course affect things. The tests done in the paper linked 
above suggest that their tests using DECnet provided much better 
throughput than ethernet could even theoretically give, though.

	Johnny

-- 
Johnny Billquist                  || "I'm on a bus
                                   ||  on a psychedelic trip
email: bqt at softjar.se             ||  Reading murder books
pdp is alive!                     ||  tryin' to stay hip" - B. Idol