Self-Adaptive Memory Approximation: A Formal Control Theory Approach

Abstract

Memory approximation enables trading off quality/accuracy for performance or energy gains. Traditionally, application programmers are burdened with the difficult task of setting memory approximation knobs to achieve the desired quality of service (QoS). Our self-adaptive approach for memory approximation eases the programmer's burden: simply specify the desired quality as a goal, with the system deploying a formal control-theoretic approach to tune the memory approximation knobs and deliver a guaranteed QoS. We model quality configuration tracking as a formal quality control problem, and outline a System Identification technique that captures memory approximation effects with variations in application input and system architecture. Preliminary results show that we can alleviate the programmer's burden of manual knob tuning for on-chip memory approximation. When compared to a manual calibration scheme we achieve 3× improvement in average settling time and upto 5× improvement in best case settling time.