Due to the increasing demand for off-chip data transfers, the traditional Von-Neumann architecture faces challenges with modern data-intensive applications, leading to the memory-wall problem. Near-memory processing (NMP) provides a solution by placing computation units near the main memory, which reduces off-chip data transfers and improves system performance. Under this paradigm, some portions of the application are transferred and executed on the NMP side, known as computation offloading. This article introduces a novel computation offloading approach for NMP-enabled 3-D memory systems, considering several critical factors collectively, including data locality information at the last-level cache and execution time estimation of the offloadable portions, which have not been collectively explored in existing studies. Further, this article proposes two different load-balancing strategies to distribute workloads among the NMP cores in the 3-D memory, thereby improving overall performance further. Extensive experiments using a variety of applications from different application domains demonstrate the effectiveness of the proposed approach. Our approach achieves a maximum speedup of$2.34\times $and$1.91\times $compared to traditional and state-of-the-art approaches, respectively. The proposed approach also reduces off-chip data transfers by nearly$5.3\times $compared to traditional computing architectures. Furthermore, the best-proposed approach reduces energy consumption by 26% (maximum) and 21% (average) compared to various state-of-the-art approaches.

With the growing awareness of sustainability concerns in many application domains, energy-harvesting (EH) devices are increasingly being used with traditional non-energy-harvesting (non-EH) devices. This paper proposes a reinforcement learning (RL)-based task offloading and scheduling strategy called DTORA for a hybrid EH and non-EH enabled multi-access edge computing environment where EH devices harvest energy from solar radiation. The applications running on user devices can have varying levels of criticality (mixed-criticality). We formulate a mixed-integer energy and latency minimization programming problem based on the system and application model. To solve this, we use a recurrent neural network based long short-term memory (LSTM) model for solar energy prediction, and a Double Deep Q-learning is used for task-offloading decisions. The proposed strategy (DTORA) is benchmarked against several state-of-the-art (SOA) strategies and other baseline approaches, including SCOPE, OCO (Offloading Cost Optimization), a hybrid Particle Swarm Optimization and Genetic Algorithm (PSOGA), and Selective-Greedy (SG). The proposed strategy outperforms these strategies in terms of latency, energy consumption of user devices, task failure rate, and critical task failures by 39%, 77.51%, 60.98%, and 69.94%, respectively (on average). Compared to the existing best-performing strategy, DTORA achieves improvements of 17.45% for latency, 58.54% for energy consumption, 23.34% for task failure rate, and 27.77% for critical task failures. This improvement can be attributed to improved edge-cloud cooperation, an efficient energy prediction model, and efficient RL-based task offloading and scheduling in our proposed strategy.

Nowadays the cloud system receives requests from a wide horizon of users. In order to execute a large number of modern resource-intensive, latency-sensitive applications with deadline requests from the users, the cloud systems are equipped with powerful machines, and the machines run for a significant amount of time. This leads to an increase in the probability of failures of these machines. Hence, the reliability of the cloud system is to be duly considered while designing a scheduling strategy for executing resource-intensive, latency-sensitive applications on it. This paper proposes an efficient scheduling strategy for executing real-time applications (scientific applications) maintaining the reliability constraints of both the cloud system and applications and the deadline constraints of these applications. The proposed policy assigns recoveries for an optimal number of tasks of the application while scheduling them on the cloud considering the reliability constraints of both the cloud system and the application. The experimental evaluation proves that the proposed policy outperforms the state-of-the-art policy both for the synthetic task set and scientific workflows.

The data-intensive applications of today's big data era often produce a large memory footprint. As a result, a significant volume of data needs to travel from memory to the CPU under the traditional Von-Neumann computing paradigm. Near-memory processing (NMP) or processing-in-memory (PIM) is a potential alternate computation framework where a computation unit is placed near the memory (or inside the memory) and a portion of an application is executed on it (termed computation offloading) aiming to reduce the amount of data movement and its consequences. Although a few computation offloading strategies have been proposed in recent times, the existing approaches do not consider the data locality offered by the last level cache and the overall execution time of the application while designing their policies. In this paper, we propose a data locality-aware computation offloading strategy for a hybrid computing system comprising the host processor and NMP-enabled 3D memory. After the application code is instrumented using the LLVM compiler framework, the strategy offloads a portion of an application to NMP if its estimated overall execution time is less. An extensive simulation performed on a set of standard simulators for a bunch of large graph-based application benchmarks reports the effectiveness of the proposed strategy by achieving a maximum speedup of 40% and 11.8% as compared to the host-only configuration and the state-of-art policy, respectively. The proposed strategy also reduces the off-chip data transfer and energy consumption by a significant margin as compared to the host-only configuration (avg 27%) and the state-of-art policy (avg 28%). Further, the proposed policy reduces the LLC miss rate by 57% as compared to the state-of-art policy.

High processing capabilities of today's large systems are also used for real time applications, where executing tasks before their deadline is essential. On the other hand, with increase in the processing capability, energy consumption also increases for such systems. Thus energy efficient execution of real time tasks in such large systems has found to be promising research area in recent time. Scheduling tasks in such large systems using only low level power construct like DVFS is not efficient. In this paper, we have exploited the power consumption pattern of the recent commercial processors and derived a simple power model with a higher granularity for systems have large number of processor with each processor having multi-threading feature. We have then proposed an energy efficient scheduling technique namely, smart allocation policy for executing a set of aperiodic independent real time tasks on large system such that no task misses it deadline. We have analyzed the instantaneous power consumption and the overall energy consumption of the proposed policy along with other five baseline policies for a wide variety of synthetic data sets and real trace data. As execution time of tasks has a significant impact on scheduling and on the overall performance of the system, we have considered six different execution time models of task for our experiment. Experimental evaluation reveals that our proposed policy performs significantly better than baseline policies for all the variations of synthetic data and for real trace data.