>Would you agree that this is not necessarily the case,
>if you can add a general-purpose CPU that the OS
>can allocate to that task, and doing so could have
>no marginal cost?
Maybe in some very unusual cases (development?
proof-of-concept?) but in general if you've got
a system to which you can "just add" a CPU you've
paid a premium for such a design. But for the sake
of argument let's say you dedicate one CPU (in a
dual CPU box) to, say, processing Enet traffic.
When that CPU is cranking in that system (if it's
anything like most current designs) that dedicated
CPU will eat interconnect bandwidth and cache behavior
as it touches memory and devices, and it will affect
latencies, throughput and determinacy (even while
processing packets that aren't of interest and are
ultimately discarded) to the point where I'd be
very surprised if you'd call it a win.
Contrast with a NIC (from a mfr who knows what
they're doing) installed in a uniprocessor box.
The box itself will have cost you less and the NIC
will be highly optimized to do what it does and
probably have economies-of-scale benefits associated
with the manufacturing process. It can touch the Enet
interface without getting in anybody else's way, it
can drop packets that aren't intended for its host,
and it can even (theoretically) handle part of the
protocol stack processing, all without ever perturbing
the host system, which need only respond to interrupts
and transfer data associated with relevant packets,
which may themselves already have been partially
(protocol-processed) by the NIC.
*****************************************************************
To unsubscribe from this list, send mail to [EMAIL PROTECTED]
with the text 'unsubscribe gnhlug' in the message body.
*****************************************************************
