Figure 4 reports the write counts onto different L2 cache
sets in Kilo Writes Per Set, KWPS, when running Dedup. From
this figure, it can be deduced that WWS of some sets is large
(43% of all sets receive more than 75KWs) while WWS of
some others sets is small (29% of all sets receive less than
50KWs). By remapping data within a set, we aim to increase
the write utilization of SRAM lines as well as uniform wear-out
of STT-RAM at line granularity. This non-uniform write
accesses between sets results in fast wear-out of some sets than
some others. To increase cache lifetime, it is worthwhile to
balance write accesses among the sets by reallocating some of
those STT-RAM lines in the write-stressed sets onto the SRAM
lines of the poorly-utilized ones. Simply stated, some
mechanism is required to increase set association and remap
the hot data of the former set to SRAM lines of the latter.