Re: FYI: I2O Architecture (fwd)

Matthew Jacob (mjacob@feral.com)
Tue, 19 Nov 96 10:13:05 PST


>Looks horribly inefficient to me (as opposed to what they claim). I am
>unconvinced that a "message-passing interface" to device driver/OS
>interaction can provide "high-performance I/O systems." Why stick
>what essentially amounts to another protocol stack sandwiched between
>the hardware-specific points and the O/S layer? Eesh.

What might help here is hardware FIFOs between components, sorta
like the Auspex hardware (which is also a message passing distributed OS).

>
>The only slightly cool thing that could arise out of this would be a
>truly heterogeneous multiprocessor; with a little extension, one machine
>could call OS services on another machine via the protocol layer. This
>might be desirable for distributed-memory systems except for the fact that one
>bad kernel call could bring down all the machines in the cluster.
>

Again, see the Auspex model. That is heterogenous processors in a
distributed model (connected by a souped up VME backplane). Last I heard
they finally did solve the 'one bad processor dragging down the system'
problem.

A message passing distributed component system would be cool. I2O
*could* offer a software framework. So could a lot of other frameworks.