Accelerating memory decryption and authentication with frequent value prediction

Weidong Shi, Hsien-Hsin S. Lee
2007 Proceedings of the 4th international conference on Computing frontiers - CF '07  
This paper presents a novel architectural technique to hide fetch latency overhead of hardware encrypted and authenticated memory. A number of recent secure processor designs have used memory block encryption and authentication to protect un-trusted external memory. However, the latency overhead of certain encryption modes or authentication schemes can be intolerably high. This paper proposes novel techniques called frequent value ciphertext speculation and frequent value MAC speculation that
more » ... nergistically combine value prediction and the inherently pipelined cryptography hardware to address the issue of latency for memory decryption and authentication. Without sacrificing security, frequent value ciphertext speculation can speed up memory decryption or MAC integrity verification by speculatively encrypting predictable memory values and comparing the result ciphertext with the fetched ciphertext. In MAC speculation, a secure processor pre-computes MAC for speculated frequent values and compares the MAC result with the fetched MAC from memory. Using SPEC2000 benchmark suite and detailed architecture simulator, our results show that ciphertext speculation and MAC speculation can significantly improve performance for direct memory encryption modes based on only 8 most frequent 64-bit values. For eight benchmark programs, the speedup is over 10% and some benchmark programs achieve more than 20% speedup. For counter mode encrypted memory, MAC speculation can also significantly reduce the authentication overhead.
doi:10.1145/1242531.1242539 dblp:conf/cf/ShiL07 fatcat:z54jp3wr3nbs3ckzxyptwcdyme