Re: [Gems-users] Cache Coherence Power in Opal/WATTCH

[Greg Byrd <gbyrd@xxxxxxxx>]

I don't know the answers to your questions, but you will at least need to say which coherence protocol you are using.

...Greg


On 9/27/2010 3:22 PM, Ragavendra wrote:
> Hello,
>
> I have a couple of questions regarding how power consumed due to
> cache coherence operations is accounted for in Opal. I searched the
> mailing list prior to posting and did not find any thread on this
> topic. I hope I'm not repeating any question.
>
> The power.C file in opal/power/ includes extra bits in the tag of a
> cache line for power calculations:
>
> tcolsb = a * (tagsize + 1 + 6) * ntspd/ntwl;
>
> I'm assuming 6 bits are for the state information and 1 is for the
> owner information stored in each cache line. Is this correct? The 7
> extra bits are added to both L1 and L2 cache tags, although the L1
> cache lines don't have as many states or owner information. Does this
> overestimate the power consumed?
>
> Also, where is the information about L1 sharers kept in L2 cache? I
> did not find any change in the tag or data fields of L2 cache to
> account for this. Where is the power consumed by these bits counted?