Most modern multi-core edge devices work in outdoor situations with limited power supplies like energy harvester and batteries. Therefore, energy consumption is a fundamental issue in which the memory subsystem has a significant role. Scratchpad memories (SPM) can provide a broad potential for energy saving. Still, due to the insufficient SPM capacity in such edge devices, a rigorous SPM data allocation scheme is necessary to reduce the energy consumption of the memory subsystem. Emerging non-volatile memories (NVMs) are very useful to reduce the energy consumption of the memory subsystem. Compared with SRAM, NVMs have lower leakage power and higher density, but the read and write latencies of the NVMs are higher than the SRAM. Therefore, embedded and edge devices can take advantage of hybrid SPM composed of both NVM and SRAM to achieve further energy saving. This paper proposes MASTER, a task mapping, task scheduling, and dynamic SPM allocation scheme that efficiently utilizes the hybrid SPM space. To this end, we model the hybrid SPM allocation on a multi-core system with integer linear programming formulation to minimize the energy consumption of the memory subsystem. Experimental results show that MASTER improves the energy saving of the memory subsystem by up to 34 percent compared to EADA, which is a heuristic dynamic data allocation algorithm for multi-core systems with hybrid SPM.

Energy consumption and predictability are two important constraints in designing real-time embedded systems and one of the recently proposed solutions for the energy consumption problem is the use of non-volatile memories instead of conventional SRAM due to their lower leakage power consumption and smaller cell area. Furthermore, because of their non-volatile nature, the use of these memories helps normally-off computing and energy harvesting systems to resume their execution without a large startup delay. However, the write access latency of non-volatile memories is considerably more than that of SRAM which can decrease the performance and predictability of the system if not managed correctly. In this article, we present a predictable fully non-volatile data memory for real-time embedded systems which improves both worst-case execution time (WCET) and performance of the system using a hybrid hardware-software solution. As part of this solution, we add a special write buffer to the memory controller and adopt a multi-bank memory configuration which improves the overall latency of write operations. Since write buffers usually help with the performance problem but they make WCET estimation more complex, we also present a new low-overhead software-based optimization technique that makes the proposed system more predictable without imposing considerable overhead. Furthermore, we present the WCET analysis algorithm which can be used to estimate the WCET of applications during the design time. The results show that compared to a hybrid SRAM-NVM architecture, the proposed solution improves the WCET and performance by 33 and 47 percent, respectively.

Multi-level cell phase change memory (MLC-PCM), because of its very low leakage power and high density, is promising for embedded systems. Furthermore, for applications with inherent low sensitivity to errors, approximate write operations can be exploited in MLC-PCM to improve endurance and performance. However, data that reside in the approximate MLC-PCM for a rather long time without refreshing are prone to soft errors due to resistance drift phenomenon, while even for an application with inherent low sensitivity to errors, a high soft error rate can degrade its Quality of Result (QoR). The architecture-level approaches to decrease the drift effect incur considerable power overhead (about 100%), which is a prominent issue in embedded systems, and are dependent on the number of logic levels stored in the PCM cell (e.g., most of them are designed for 4LC-PCM). This article, taking a different approach, proposes a drift-aware frequency and voltage management to alleviate the drift-based soft-error rate. To this end, first we characterize the application data based on the degree of being exposed to the drift to identify the drift-prone application data. Then we assign the execution frequency and voltage to different regions of the application considering the drift. This frequency assignment speeds up the application regions wherein the drift-prone data are accessed to shorten the lifetime of the drift-prone data, thereby decreasing the soft error rate. An integer linear programming model implements our proposed Dynamic Voltage Frequency Scaling (DVFS). Also, the proposed approach is independent of the number of levels of PCM cells and can be applied to any MLC-PCM system. To evaluate the approach, the approximate MLC-PCM is simulated using empirical models and is integrated into a full-system simulator as data memory. The experimental results show that, by exploiting the approach, QoR is in the acceptable range, while its power overhead is about 84% (on average) less than that of the architecture-level approach.

Energy consumption is one of the prominent design constraints of multi-core embedded systems. Since the memory subsystem is responsible for a considerable portion of energy consumption of embedded systems, Non-Volatile Memories (NVMs) have been proposed as a candidate for replacing conventional memories such as SRAM and DRAM. The advantages of NVMs compared to conventional memories are that they consume less leakage power and provide higher density. However, these memories suffer from increased overhead of write operations and limited lifetime. In order to address these issues, researchers have proposed NVM-aware memory management techniques that consider the characteristics of the memories of the system when deciding on the placement of the application data. In systems equipped with memory management unit (MMU), the application data is partitioned into pages during the compile phase and the data is managed at page level during the runtime phase. In this paper we present an NVM-aware data partitioning and mapping technique for multi-core embedded systems equipped with MMU that specifies the placement of the application data based on access pattern of the data and characteristics of the memories. The experimental results show that the proposed technique improves the energy consumption of the system by 28.10 percent on average.