Re: [Gems-users] Token coherence protocol

[Mike Marty <mikem@xxxxxxxxxxx>]

MOESI_CMP_token can be configured as an SMP by setting PROCS_PER_CHIP to
1.  However because it was designed as a CMP protocol, tokens are used to
manage requests from L1 misses instead of L2 misses.  In the SMP-only
protocols, coherence is only done between SMP nodes such that the L1/L2 is
managed in the same controller.

I actually planned on releasing our SMP version of Token Coherence
(MOESI_SMP_token), however I pulled it for release 1.0 due to last-minute
technical problems, but will likely include it for release 1.1 sometime in
May.

Does this really matter?  Probably not to the first order.  There is more
going on in MOESI_CMP_token because of L1->L2 interactions that aren't
modeled in MOESI_SMP_token.  The controllers are specified seperately and
each has a persistent request table.  MOESI_SMP_token has exclusive L1/L2
caches (like all our SMP-only protocols) whereas MOESI_CMP_token has
non-inclusive as the L2 is essentially treated as a victim cache.  There
might be other subtle differences.

--Mike


> Hello.
>
> I am insterested on using the Token Coherence Protocol.  I have looked for any
> implementation of this protocol in the $GEMS/protocols directory of GEMS but
> I have only found the MOESI_CMP_token protocol. This protocol is used with
> Chip-Multiprocessors system (is it not?), but I want to test the SMP systems.
> I do not know if it is possible to use the MOESI_CMP_token protocol with SMP
> systems. If it is not possible, could anyone tell me if there is some
> implementation of the token coherence protocol for SMP systems?
>
> Thanks,
>
> Blas.
>
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>