You've identified big CMP design issues that don't have clear answers.
I'm not sure what kind of help you want.
GEMS is an ideal platform for studying these issues. You need to first
come up with your conceptual target system. Then you will either need
to configure GEMS to match this, or add new functionality to GEMS.
GEMS currently supports 2 classes of systems. "SMPs" with private
caches. However make the interconnect latencies small, and you can
think of this as a CMP with private caches rather than an SMP. The
"CMP" protocols all have a shared L2 cache and support Multiple-CMP
systems. In included a new "SCMP" protocol in release 1.4 for
simulating a single-CMP, but allows flexible modifications for how this
is shared or private (shared by default). We also have a D-NUCA
implementation for a Single-CMP, but you shouldn't start with this.
The naming conventions are lousy, but I didn't know better several years
ago.
--mike
Niket Agarwal wrote:
Hi all,
I want to configure a CMP environment with many cores (32+) having
private L1 caches and a shared L2 cache which is distributed with the
cores.
As I have gathered, the distributed bank concept does not scale. Also,
having a unified shared L2 cache would not scale.
It would be great if someone could help me with some pointers.
Cheers,
Niket.
_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/"; to your search.
|
![[Computer Systems Lab]](https://www-auth.cs.wisc.edu/pics/csl_logo.gif){kind=logo}