I have checked the etherlink model and etherint model which
has an etherint peer. Once the peer is set, two etherints
on both ends of the etherlink are binded together. So a direct
point-to-point link is set. As you said, I can just create a
switch model with some interfaces and make each etherint bind
with a switch interface instead of the etherints on other
systems. However, this network topology is a kind of star
network which is different from the fully-connected network
because the etherint on one system is shared to connect with
other systems, sometimes causing the transmission delay.
True, this is a star network, but as long as you implement the switch
properly and use link bandwidths that are appropriate, the performance of
the star will be the same as the performance of a fully connected network.
If you actually take into consideration the fact that the NICs end up
sharing some sort of bus at some point before getting into the processor,
all the star network is doing different is sharing a network link instead
of sharing a PCI bus.
I think each system should have N-1 etherints in order to
implement a fully-connected network of N systems. How do
you think of that? I know there is a limit on PCI devices.
As I said, I don't think it is necessary provided that you implement the
switch right and provide the right bandwidth to the link.
Do you think it possible to modify this limit? If yes, where
should I start with? Is this about the pci bus address limit?
If you really want to go this route, you're going to have to make changes
in several places. Potentially including the kernel. Though as I said,
you may be able to have a bunch of devices share PCI interrupts as long as
you separate the PCI functions. You'd just have to try.
Nate
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
