Hello,
I have a few questions regarding prefetching implementation in Ruby. So far, I've added a component that predicts the data to be fetched, and modified SLICC files accordingly. In my simulations, even though the number of misses is less than the unmodified version, the number of Ruby cycles is higher. In addition, the number of data and instructions requests forwarded by Simics is higher. Here are my questions:
- The "total misses" reported by Ruby profiler includes only L2 misses in MOESI-SMP-hammer. Why is that? Why aren't L1 misses included in the total count?
- Do I need to track prefetch requests and modify SimicsDriver or SimicsProcessor accordingly to eliminate the extra data/instruction requests?
- What may cause the increase in ruby cycles? Any hints?
Thanks for your help!
Dave
__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
http://mail.yahoo.com
|