Y.C. Tay National University of Singapore Dynamic System Optimization through Performance Modeling: Universality and Decomposition Y.C. Tay National University of Singapore

Dynamic optimization is hard. Examples: Prob(miss) = f(cache size) Given target Prob(miss), how to dynamically adjust cache size? Universality (2) Internet traffic equilibrium How to prevent performance collapse from congestion? Decomposition

(1) Universality reference pattern cache size management policy issue: Prob(miss) = f(cache size) reference pattern cache size management policy complex add: ● system customization ● hw/sw evolution ● autonomic configuration ● dynamic adjustment hardware variation data layout application mix software variation data instance intractable! system configuration

(1) Universality no change idea: Prob(miss) = f(cache size | parameters) change values ● system customization ● hw/sw evolution ● autonomic configuration ● dynamic adjustment

(1) Universality = (H + √(H 2 – 4))(n*+n0) – n0 idea: Prob(miss) = f(cache size | parameters) #miss = f (M | M*, M0, n*, n0) change values = (H + √(H 2 – 4))(n*+n0) – n0 1 2 where H = 1 + M* - M0 M - M0 ● system customization ● hw/sw evolution ● autonomic configuration ● dynamic adjustment

(1) Universality = (H + √(H 2 – 4))(n*+n0) – n0 #miss = f (M | M*, M0, n*, n0) = (H + √(H 2 – 4))(n*+n0) – n0 1 2 where H = 1 + M* - M0 M - M0 Example: cache=RAM