The static profiling approach used to generate Fig. 5
requires running each application multiple times and does
not capture dynamically changing behavior. To predict
cache requirement and classify applications on the fly, we
explore the relation between application page access behavior
and cache utilization. The key insight is that the number
of hot pages (NHP) (active pages used in a particular time
interval) can reflect an application’s LLC demand in many
cases due to the DRAM row-buffer locality [29]. SysMon
can identify hot pages by examining the access bit [39], [40]
in the page table entry (PTE). Fig. 6 shows the correlation