Design Space Exploration for Value Prediction in Security Applications

• Main Author: Gunnarsson, Linnea
• Format: Others
• Language: English
• Published: Uppsala universitet, Institutionen för informationsteknologi 2019
• Subjects: Spectre; Meltdown; Value prediction; Other Computer and Information Science; Annan data- och informationsvetenskap
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-409971

With the introduction of Spectre and Meltdown, two new attacks thattarget the speculative instructions due to Out-of-Order execution intoday's processors, a new way to handle speculative loads has beenproposed. Instead of performing the speculative load, the approach isto predict them. This is a new way to use value predictors. In thiswork, the Last Value Predictor, which predicts based on the previouslyseen value, Value TAgged GEometric history length Predictor (VTAGE),which predicts based on the global branch history, and a stridepredictor, which predicts with help of strides, has been compared tosee which one has the best fit for this new use. They have been runwith the SPEC CPU 2017 benchmark suite in three different tests,different sizes, different threshold confidence and for VTAGE,different associativity. The VTAGE predictor performed best in terms ofvalues predicted and values correctly predicted. The thresholdconfidence level plays an important role in how many incorrectpredictions were made. The associativity in the VTAGE did not do muchdifference to the results.