Software and hardware interfaces for approximate memories

Abstract

Memory components are sensitive to process, voltage, and temperature variability and, to ensure reliability in the stored data, vendors specify operating parameters considering the worst case in the process design with a guard-band margin. Approximate memories adjust parameters out of the guard-band range, which allows for energy savings at the cost of probabilistic errors in the stored data. While several applications tolerate some imprecision in their results, not all data are resilient to errors, and even resilient data have limits on their tolerance to errors. Uncontrolled errors may produce results with more imprecision than acceptable, which render them useless or even crash an application execution unexpectedly because of errors in critical data. Data access interfaces control what data are exposed to errors, through critical data protection, and how much error can be inserted to respect the imprecision limit of the results, through the configuration of the approximate memory. This control depends on the application and the error impact in the results of the computation. Interfaces typically rely on programmer annotations that change the application and domain-specific metrics to identify the data amenable to approximation or to configure the allowed error amount. Nevertheless, annotations hinder the portability and maintainability of the code, and domain-specific metrics demand some knowledge about the error impact on the application. Moreover, the error is a dynamic element dependent on a scenario that is composed of environment variables, such as temperature and fabrication process. Transparent interface mechanisms attempt to automatically protect critical data and control the error into the threshold that the application tolerates changing the error configuration according to the environment variables. This work proposes transparent interfaces for the control of memory approximation to improve execution resilience and increase energy efficiency. We analyze the execution of applications when data elements are exposed to probabilistic errors and find common critical data that cause execution crashes. Then, we propose hardware and software mechanisms to treat these data and avoid crashes aiming to generate useful results. We present protections for physical and virtual addressing and investigate the impact of error from different memory hierarchy levels, besides the exploration of alternatives for a supervisor system to treat incorrect executions for recovering application data. Lastly, this work relates the execution behavior of applications with their error tolerance to transparently configure the approximate memory. The behavior is abstracted from measurable execution statistics that are correlated with a configuration using a knowledge base of previous training executions. Our results show that these interfaces improve execution resilience by reducing a significant part of the crashes and configure approximate memories with energy savings and average quality close to the achieved by an exhaustive search.