results

Jonathan Lafiandra
Analysis of various C/B/S/T configurations for certain traces.

Perlbench: 
C  B S T
10 6 2 3  ~ AAT.best = 5.437, T being 3 seems best as predicted, as it has the benefit of both markov and n+1. C should be as big as possible within the constraints as the larger the cache, the better. B being around 5 to 6 seems to be sweet spot of large temporal locality without incredibly high pollution. S = 2 seems best for associativity as a little associativity goes a long way, and it gives associativity 4 way associativity while keeping HT relatively low.

Mcf:
C B S T 
10 7 2 3 ~ AAT.best = 4.302, These stats seem extremely similar to perlbench making me think that the two traces have extremely similar patterns. The only difference being that B = 7 is more optimal for MCF as perhaps it has even more temporal locality than perlbench.

Astar:
C B S T 
10 7 1 3 ~ AAT.best = 4.081, Again it seems that somewhere around this 10/(6,7)/(1,2)/3 setup would be the best setup for the standard trace. Astar seems to have less conflicts among its data, making it so that S=1 is actually the best associativity.


Bzip2:
C B S T 
10 4 1 2 ~ AAT.best = 2.742,  The most interesting of all of the traces. The prefetcher type actually seems to be best at 2, the n+1 prefetcher perhaps lots of temporal locality in Bzip. C as big as possible as the standard. B being around 4 seems to be, perhaps pollution happens quickly in this. s = 1 is a relatively low associativity, but it seems enough.