Unfortunately, I don't think there's an easy answer.  You could put a
finite write buffer in main memory, but the only current way to
provide backpressure when it fills up is through bus flow control,
which would stall reads as well as writes.  Another approach is to add
the ability to nack writebacks at the protocol level, but that's a
pretty significant change to a pretty complex and fragile part of the
system.

This fragility is a big part of the reason we're moving away from the
"classic" M5 memory system to Ruby... I don't know what the writeback
bandwidth modeling is like in Ruby, but I'm sure it's at least as good
:-), and even if it's not what you need, it might be easier to
enhance.  So overall your easiest path may well be to switch to Ruby.

Steve

On Mon, Jul 5, 2010 at 8:34 AM, sheng qiu <[email protected]> wrote:
> Hi Steve,
>
> Thanks for your answer. if i want to include the effect of writeback latency
> to memory into the running application(i.e. the bus bandwidth is not
> enough), what can i do for this?
>
> Thanks,
> Sheng
>
>
> _______________________________________________
> m5-users mailing list
> [email protected]
> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
>
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Reply via email to