PDP-11/74 MP

[Johnny Billquist]

On 2015-05-18 00:16, Noel Chiappa wrote:
>      > From: John Wilson
>
>      > Even if E11's mP feature is officially unsupported, it was a *crazy*
>      > amount of work
>
> What made it so much work? (Just curious about the technical aspects...)

John really needs to answer this one, as I think only he knows the 
innards of E11.
However, I know at least parts of the story, as related to how RSX 
works. Part of the problem is just memory coherency. The 11/74 
essentially guarantees memory consistency. If you write to a memory 
location from one CPU, it is immediately visible from all CPUs. At the 
same time, the 11/74 do not have any cache coherency. I believe in 
general, you can accept the memory coherency of modern systems as an 
acceptable alternative, since they essentially just gives you more 
coherency.
But it still needs examination and evaluation.

A second problem is how interprocess communication works. The 11/74 have 
a special device which takes care of this. It have something like a door 
bell for each CPU, and when one CPU wants to get the attention of 
another CPU, this device is used. It causes an interrupt on the second 
CPU, and it is then expected to figure out why it was knocked on, and 
acknowledge the request before the requesting CPU times out. This 
timeout is essentially done through a spin lock with a counter.
On modern machines, this counter runs way faster than on a real 11/74. 
In addition, the time spent on each thread, which implements the 
individual CPUs, might not match 1:1 to actual CPUs. So, in essence, the 
requesting CPU could be doing a whole lot of work before the requested 
CPU even gets to execute a single instruction. This will cause the 
interprocessor interrupt to time out in RSX. So, when the IIST 
interprocess interrupts are used, you need to somehow make sure that all 
CPUs will execute some code before the requesting CPU might time out. 
That is not trivial to accomplish.

Third, the actual IIST device documentation was not available when the 
E11 implementation was done, so it was a lot of reverse engineering 
based on how RSX use the device.

Fourth, some other bits and pieces of an 11/74 also exists and are 
different from other 11/70 systems (not to mention any other PDP-11). 
E11 don't do all of them correctly either, such as the MKA-11 memory 
boxes, which have their own CSRs.

Also, not all parts of how an 11/74 system works are really that clearly 
understood. For example, an 11/74 can start and stop individual CPUs 
while the system is running. Exactly how is this done? They obviously do 
not have normal M9312 bootstrap cards with standard boot roms, since 
that would not work. An 11/74 do have four separate consoles (if we have 
four CPUs), and four separate Unibuses. Also, four separate front 
panels. This also means that the same CSR addresses can be used by 
several devices, if they are on different buses. This also means that 
all device emulations also needs to now be aware of which bus they are 
sitting on.

I guess the list goes on... :-)

	Johnny