Steve Reinhardt:
> Hi Marius,
> 
> ReadEx is read exclusive, you're right.
> 
> The reason we get away with not having an entry for 
> transitionTable[Invalid][ReadEx].onRequest is that the protocol does not 
> handle the general case of multilevel coherent caches.  Effectively for 
> now you're limited to either a single level of caches or a set of 
> coherent L1 caches backed by a shared L2.  In the latter case the L2 
> should use 'uni' coherence and not the full coherence protocol.  Since 
> we don't support configurations where there's a coherent cache behind 
> another coherent cache, the protocol will never see ReadEx as a request.
> 
> We have plans to remedy this but haven't had the time to implement it. 
> Our current plan is to allow all the caches to snoop on a single shared 
> bus (much like in most modern systems, both the L1 and L2 snoop together 
> on external traffic).  The problem could also be solved by enforcing 
> inclusion but that's less common in industry these days (AFAIK... I'm 
> not exactly sure what Core 2 Duo does, for example).

I can probably get away with inclusion. Do you have support for this in 
M5? I couldn't find any references to this in BaseCache.py, or do I need
to hack the c++ code?
 
> In the longer run we would really like a more general network-based (not 
> just bus-based) coherence infrastructure; a big part of the v2 memory 
> system rewrite was to make that more feasible.  We haven't really 
> started on actually implementing that though.  There's been talk of 
> integrating Wisconsin's GEMS code but no action as of yet. 
> Contributions are welcome :-).

Yes, this would be very interesting. Also, I think that multi-level
cache coherence will become important to more people than just me in 
the future. I haven't had the time to really look at v2 yet, but I hope
to move my work to it when it gets out of beta.

-- 
Marius Grannæs
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Reply via email to