Re: [Gems-users] Cache Organization

[hb166307@xxxxxxxx]

Thanks!

Quoting Niket Agarwal <niketa@xxxxxxxxxxxxx>:

> The g_FLIT_SIZE is a GARNET specific parameter and GARNET does not
> use  
> the g_endpoint_bandwidth and bw_multiplier arguments.
> 
> g_FLIT_SIZE is used to specify the bandwidth (bytes/cycle) of on-chip
>  
> links. It was chosen assuming constant on-chip link latency. If there
>  
> is a desire to have different link bandwidths, the latencies in the 
> 
> topology file can be increased to incorporate that. For e.g. if a
> link  
> with half the bandwidth of g_FLIT_SIZE is desired, the latency for
> the  
> link can be doubled. So, if the g_FLIT_SIZE is 16 bytes/cycle and a 
> 
> link of 8 bytes/cycle is desired, the latency of the particular link 
> 
> can be made to 2. So, 16 bytes will take 2 cycles to cross the link =
>  
> 8 bytes/cycle. I think the pipelining in that particular link would 
> 
> have to be removed. Currently GARNET assumes that if a link has a  
> latency >  1, there can be multiple flits on the link at a time.
> 
> As regards the L2 cache organization, from my understanding, L2 on a 
> 
> "chip" is always shared. So, if you define multiple L2's on a chip
> the  
> L2 would be banked but would still be shared. The only way to get  
> private L2 is to have multiple "chips"
> 
> -Niket
> 
> On Oct 16, 2008, at 10:43 AM, hb166307@xxxxxxxx wrote:
> 
> > Dear Users,
> >
> >    I have been reading through the archives trying to understand
> how  
> > the cache
> > structure is organized and have come across information that leads 
> 
> > me to believe
> > that if I place one L2Bank on a "chip" or a CMP in a Multiple-CMP 
> 
> > system that it
> > is considered shared; and if I place more than one on a "chip" each
> is
> > considered private or is there more to it than that?
> >
> >    My next question is in regard to the g_endpoint_bandwith  
> > parameter versus
> > g_FLIT_SIZE:
> >
> >     1) g_FLIT_SIZE determines the amount of bytes/cycle across all 
> 
> > links?
> >     2) g_endpoint_bandwidth allows me to increase the bandwidth of 
> 
> > individual
> >        links when specified in a custom network design?
> >     3) So If I have g_FLIT_SIZE as 16, but my bw_multiplier = 64
> >        and g_endpoint_bandwidth = 1000 what is the bandwidth on the
>  
> > specified ink
> >        versus the others.
> >
> > I would appreciate any help trying to figure this out.
> >
> > Thanks
> > _______________________________________________
> > 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.
> >
> 
> _______________________________________________
> 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.
> 
>