Re: [Gems-users] Adapting SMP directory protocol to model private L2 CMP protocol

["Niket Agarwal (niketa@xxxxxxxxxxxxx)" <niketa@xxxxxxxxxxxxx>]

That works.

Thanks,

Niket

----- Original Message -----
From: Mike Marty <mike.marty@xxxxxxxxx>
Date: Friday, July 4, 2008 10:15 am
Subject: Re: [Gems-users] Adapting SMP directory protocol to model private L2 CMP protocol
To: Gems Users <gems-users@xxxxxxxxxxx>

> > I want to evaluate a private L2 CMP directory protocol in GEMS.
> Looking at
> > some of the previous discussion it seems that lowering the
> network latencies
> > and using an SMP protocol could mimic a private L2 CMP protocol.
> >
> > My concern is that this would not model an on-chip directory
> cache. If I
> > specify the directory link latency to be 200 cycles (to model
> off-chip
> > memory), every check of the sharing vectors would incur an off-
> chip 200
> > cycle access.
> >
>
> You can add a directory cache to MOESI_SMP_directory-dir.sm.  Just
> modifythe actions to check for a directory cache hit and use a
> different latency
> value if you hit.
>
>
> >
> > Am I correct? How have private L2 configurations been previously
> evaluated> inside Gems?
> >
>
> Yes they have.
>
> With private L2s, you can't really store directory state in the L2
> tags.
> --Mike
>