SeqAss: Using SeqUential Associative Caches to Mitigate Conflict-Based Cache Attacks with Reduced Cache Misses and Performance Overhead

Cache randomization has been proposed as an effective defense against conflict-based cache attacks. Mirage and Chameleon are two of the state-of-the-art randomized lastlevel caches achieving a strong defense. However, they rely on techniques intrusive to the traditional cache structure, such as cache skews, over-provided metadata space, and separated data storage, and prohibit the use of the LRU replacement policy. Mirage incurs 22% extra area and 21% extra power. When running memory heavy applications, Chameleon consumes significant dynamic power due to its high relocation rate. This paper proposes to mitigate conflict-based cache at tacks using sequential associativity. The proposed SeqAss cache retains the set-associative structure and supports LRU. It achieves a defense as strong as Mirage. Instead of raising cache miss rate, SeqAss actually reduces it by 11.4%. Its area and power overhead is 28.8% and 22.1%, respectively, lower than Mirage. When running memory heavy applications, it incurs $\sim 50 \%$ lower dynamic power overhead compared to Mirage and Chameleon.