-Question about Figure 5: Shouldn't the black bar (no prefetching) have the equal height within simulations for a single benchmark? (“gromacs” and “sjeng” have different heights for the black bars. For “omnetpp”, black bars are not present.)

-Lack of benchmark-specific analysis. Show cross-stream correlations with the miss graphs for a particularly well performing and bad performing benchmark.

-Can there be other ways to exploit both spatial/temporal locality together?