Fig. 7From: Keep the PokerFace on! Thwarting cache side channel attacks by memory bus monitoring and cache obfuscationPokerFace performance overhead on Parsec benchmark. The graph shows the overhead incurred by Poker on the Parsec benchmark suite. The benchmark suite consists of different programs which perform a variety of tasks like cache-aware simulated annealing (canneal), frequent itemset mining (freqmine), online clustering (streamcluster), image processing (vips), video encoding (x264), etc. Since cloud instances are regularly used for machine learning and image/video processing applications, these set of benchmarks are a representative set of real world use cases. Memory intensive applications like streamcluster and canneal suffer from 5% overhead due to Poker and a maximum of 8% overall. Other applications which are predominantly CPU intensive like blackscholes (Black-Scholes partial differential equation), ferret (content similarity search) and fluidanimate (fluid dynamics for animation) incur overheads of up to 3%Back to article page