lkj,
Here is a response from Elizabeth Rather on the question of the
PAUSE (multitasking, context switch):
>
> Hi Elizabeth,
>
> It appears that the PSC1000 did not do to well for a context switch. What
> kind of multitasking was used? What is the PSC1000 doing?
As you may know, the PSC1000 is a stack processor, rather than a
register-oriented processor. Its "operand stack" is Forth's data stack,
and its "local-variable" stack is the return stack. This makes most
operations extremely fast, as you can see from the posted benchmarks.
However, to do a context switch requires saving all the stack registers
(18 & 16, respectively) in the task's user area. There are single
instructions to do this, so it's not like having to move each cell
individually, but it's still more data transfers than most processors
require (typically 3). Our implementation on the PSC1000 improves
slightly on their published code to do this.
The multitasker model is the same one we've used for 20 years, a
cooperative round-robin in which a task relinquishes the CPU whenever it
does I/O, plus voluntary PAUSEs to allow "tuning" for CPU-intensive
operations. Since context switches are relatively infrequent, we think
the overall speed of operation of the PSC1000 more than compensates for
the slower context switch.
Cheers,
Elizabeth |
