Andrea Alimonda

Assessing Task Migration Impact on Embedded Soft Real-Time Streaming Multimedia Applications

Multiprocessor systems on chips (MPSoCs) are envisioned as the future of embedded platforms such as game-engines, smart-phones and palmtop computers. One of the main challenge preventing the widespread diffusion of these systems is the efficient mapping of multitask multimedia applications on processing elements. Dynamic solutions based on task migration has been recently explored to perform run-time reallocation of task to maximize performance and optimize energy consumption. Even if task migration can provide high flexibility, its overhead must be carefully evaluated when applied to soft real-time applications. In fact, these applications impose deadlines that may be missed during the migration process. In this paper we first present a middleware infrastructure supporting dynamic task allocation for NUMA architectures. Then we perform an extensive characterization of its impact on multimedia soft real-time applications using a software FM Radio benchmark.

A Feedback-Based Approach to DVFS in Data-Flow Applications

Runtime frequency and voltage adaptation has become very attractive for current and next generation embedded multicore platforms because it allows handling the workload variabilities arising in complex and dynamic utilization scenarios. The main challenge of dynamic frequency adaptation is to adjust the processing speed of each element to match the quality-of-service requirements in the presence of workload variations. In this paper, we present a control theoretic approach to dynamic voltage/frequency scaling for data-flow models of computations mapped to multiprocessor systems-on-chip architectures. We discuss, in particular, nonlinear control approaches to deal with general streaming applications containing both pipeline and parallel stages. Theoretical analysis and experiments, carried out by means of a cycle-accurate energy-aware multiprocessor simulation platform, are provided. We have applied the proposed control approach to realistic streaming applications such as Data Encryption Standard and software-based FM radio.

A control theoretic approach to energy-efficient pipelined computation in MPSoCs

In this work, we describe a control theoretic approach to dynamic voltage/frequency scaling (DVFS) in a pipelined MPSoC architecture with soft real-time constraints, aimed at minimizing energy consumption with throughput guarantees. Theoretical analysis and experiments carried out on a cycle-accurate, energy-aware, and multiprocessor simulation platform are provided. We give a dynamic model of the system behavior which allows to synthesize linear and nonlinear feedback control schemes for the run-time adjustment of the core frequencies. We study the characteristics of the proposed techniques in both transient and steady-state conditions. Finally, we compare the proposed feedback approaches and local DVFS policies from an energy consumption viewpoint.

Impact of Task Migration on Streaming Multimedia for Embedded Multiprocessors: A Quantitative Evaluation

Dynamic task mapping solutions based on task migration has been recently explored to perform run-time reallocation of task to maximize performance and optimize energy consumption in MPSoCs. Even if task migration can provide high flexibility, its overhead must be carefully evaluated when applied to soft real-time applications. In fact, these applications impose deadlines that may be missed during the migration process. In this paper we first present a middleware infrastructure supporting dynamic task allocation for NUMA architectures. Then we perform an extensive characterization of its impact on multimedia soft realtime applications using a software FM Radio benchmark.

A Control Theoretic Approach to Run-Time Energy Optimization of Pipelined Processing in MPSoCs

In this work we take a control-theoretic approach to feedback-based dynamic voltage scaling (DVS) in multi processor system on chip (MPSoC) pipelined architectures. We present and discuss a novel feedback approach based on both linear and non-linear techniques aimed at controlling interprocessor queue occupancy. Theoretical analysis and experiments, carried out on a cycle-accurate multiprocessor simulation platform, show that feedback-based control reduces energy consumption with respect to standard local DVS policies and highlight that non-linear strategies allows a more flexible and robust implementation in presence of variable workload conditions

Non-Linear Feedback Control for Energy Efficient On-Chip Streaming Computation

In this work we take a control theoretic approach to the dynamic voltage/frequency scaling (DVS) in a MPSoC architecture with mixed pipelined/parallel processing. The aim is that of minimizing energy consumption with throughput guarantees. Theoretical analysis and experiments, carried out on a cycle-accurate, energy-aware, multiprocessor simulation platform, are provided. We give a dynamic model of the system behavior on the basis of which we synthesize a non-linear feedback controller for the run-time adjustment the frequencies of the processing stages. We compare, from an energy consumption viewpoint, the proposed feedback approaches with local DVS policies.

Exploiting Memory-Boundedness in Energy-Efficient Hard Real-Time Scheduling

Dynamic voltage and frequency scaling (DVFS) has been extensively exploited in the context of hard real-time systems for the development of energy efficient task scheduling algorithms. However, when tasks are memory bounded, further energy improvement could be obtained. In this paper we analyze the effect of memory boundedness in a state-of-the-art energy efficient hard real-time scheduling algorithm, and we propose a new technique to take into account these effects to substantially improve energy efficiency of the scheduling algorithm while still preventing deadline misses. The proposed technique is compared to a state-of-the-art hard real-time scheduling algorithm from energy efficiency viewpoint. Results show an energy reduction from 15% to 90% depending on the amount of memory boundedness of the task.

Temperature and Leakage Aware Power Control for Embedded Streaming Applications

Leakage power has an increasing relevance for recent and future embedded technologies. Moreover, its contribution is heavily dependent on runtime temperature and process variations, that strongly impact the effectiveness of energy management strategies. Furthermore, in the context of streaming embedded systems, aggressive shutdown techniques that are typically applied to reduce leakage contribution are not suitable because of their negative effect on throughput. In this work we propose a leakage-aware approach for streaming applications that: i) Exploits feedback control policy to guarantee throughput by observing output data rate; ii) Levarages runtime temperature information to adjust the control law so that leakage energy is minimized independently from process variations and temperature conditions. Results show how temperature and leakage variations impact energy management strategies and how the proposed approach is effective in reducing total power consumption.

A modular digital VLSI architecture for stereo depth estimation in industrial applications

In this paper a modular architecture for real time stereo depth estimation is synthesized. The elaboration is described in terms of the action of a set of sub-processors working in parallel. Such modules have been implemented using proper mechanism to reduce their complexity. Simulations on real and synthetic images are presented and real-time performance evidenced.