FairSDP: Fair and Secure Dynamic Cache Partitioning

Due to the widespread usage of cloud computing, and remote execution, where the resources of a computer are shared by multiple executions, hardware security gained more attention. It is possible to extract confidential information such as cryptographic keys through cache-based side-channel attacks as in Meltdown and Spectre attacks. Consequently, studies over secure cache architectures has gained a deeper focus. Secure cache architectures come with an inevitable performance penalty, since there is always an overhead for hiding information from the attackers. Subsequently, the performance degradation is traded off with the improvement in security. In this paper, we analyze the performance deterioration of the existing architectures and come up with a new solution which improves the fairness of the general framework. We propose a secure cache mechanism that respects fairness among the competing threads within a processor. We evaluate FairSDP architecture in 4-threaded and 8-threaded processors. As a result, we show that we can achieve up to 8.7% performance improvement over the baseline and 9.2% better performance compared to the static partitioning on the average, in an 8-threaded system. We also achieve almost identical results in terms of the fairness metric compared to a non-secure dynamic cache partitioning scheme.