Self-adaptive Memory Approximation: A Formal Control Theory Approach

Abstract

Memory approximation enables trading off qual-ity/accuracy for performance or energy gains. Traditionally,application programmers are burdened with the difficult taskof setting memory approximation knobs to achieve the desiredquality of service (QoS). Our self-adaptive approach for memoryapproximation eases the programmer's burden: simply specifythe desired quality as a goal, with the system deploying a formalcontrol-theoretic approach to tune the memory approximationknobs and deliver a guaranteed QoS. We model quality con-figuration tracking as a formal quality control problem, andoutline a System Identification technique that captures memoryapproximation effects with variations in application input andsystem architecture. Preliminary results show that we can allevi-ate the programmer's burden of manual knob tuning for on-chipmemory approximation. When compared to a manual calibrationscheme we achieve 3×improvement in average settling time andupto 5×improvement in best case settling time.